US20190115923A1 - Processor for Implementing Non-Arithmetic Functions - Google Patents
Processor for Implementing Non-Arithmetic Functions Download PDFInfo
- Publication number
- US20190115923A1 US20190115923A1 US16/207,189 US201816207189A US2019115923A1 US 20190115923 A1 US20190115923 A1 US 20190115923A1 US 201816207189 A US201816207189 A US 201816207189A US 2019115923 A1 US2019115923 A1 US 2019115923A1
- Authority
- US
- United States
- Prior art keywords
- processor
- arithmetic
- lut
- alc
- function
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Abandoned
Links
Images
Classifications
-
- H—ELECTRICITY
- H03—ELECTRONIC CIRCUITRY
- H03K—PULSE TECHNIQUE
- H03K19/00—Logic circuits, i.e. having at least two inputs acting on one output; Inverting circuits
- H03K19/02—Logic circuits, i.e. having at least two inputs acting on one output; Inverting circuits using specified components
- H03K19/173—Logic circuits, i.e. having at least two inputs acting on one output; Inverting circuits using specified components using elementary logic circuits as components
- H03K19/177—Logic circuits, i.e. having at least two inputs acting on one output; Inverting circuits using specified components using elementary logic circuits as components arranged in matrix form
- H03K19/17724—Structural details of logic blocks
- H03K19/17728—Reconfigurable logic blocks, e.g. lookup tables
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F15/00—Digital computers in general; Data processing equipment in general
- G06F15/76—Architectures of general purpose stored program computers
- G06F15/78—Architectures of general purpose stored program computers comprising a single central processing unit
- G06F15/7839—Architectures of general purpose stored program computers comprising a single central processing unit with memory
- G06F15/7842—Architectures of general purpose stored program computers comprising a single central processing unit with memory on one IC chip (single chip microcontrollers)
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F15/00—Digital computers in general; Data processing equipment in general
- G06F15/76—Architectures of general purpose stored program computers
- G06F15/78—Architectures of general purpose stored program computers comprising a single central processing unit
- G06F15/7867—Architectures of general purpose stored program computers comprising a single central processing unit with reconfigurable architecture
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F7/00—Methods or arrangements for processing data by operating upon the order or content of the data handled
- G06F7/38—Methods or arrangements for performing computations using exclusively denominational number representation, e.g. using binary, ternary, decimal representation
- G06F7/48—Methods or arrangements for performing computations using exclusively denominational number representation, e.g. using binary, ternary, decimal representation using non-contact-making devices, e.g. tube, solid state device; using unspecified devices
- G06F7/483—Computations with numbers represented by a non-linear combination of denominational numbers, e.g. rational numbers, logarithmic number system or floating-point numbers
- G06F7/487—Multiplying; Dividing
- G06F7/4876—Multiplying
Definitions
- the present invention relates to the field of integrated circuit, and more particularly to processors.
- LBC logic-based computation
- Conventional logic circuits are suitable for arithmetic functions, whose operations include only basic arithmetic operations, i.e. the arithmetic operations performable by the conventional logic circuits per se.
- the basic arithmetic operations consist of addition, subtraction and multiplication.
- the conventional logic circuits are not suitable for non-arithmetic functions, whose operations include more than the basic arithmetic operations. Because they cannot be represented by a combination of the basic arithmetic operations, the non-arithmetic functions cannot be implemented by the conventional logic circuits alone.
- Exemplary non-arithmetic functions include transcendental functions and special functions.
- Non-arithmetic functions are computationally hard and their hardware implementation has been a major challenge.
- a conventional processor 300 generally comprises a logic circuit 380 and a memory circuit 370 .
- the logic circuit 380 comprises an arithmetic logic unit (ALU) for performing arithmetic operations, while the memory circuit 370 stores an LUT for the built-in function.
- ALU arithmetic logic unit
- the built-in function is approximated to a polynomial of a sufficiently high order.
- the LUT 370 stores the coefficients of the polynomial; and the ALU 380 calculates the polynomial. Because the ALU 380 and the LUT 370 are formed side-by-side on a semiconductor substrate 0 , this type of horizontal integration is referred to as two-dimensional (2-D) integration.
- the computational density is a figure of merit for parallel computation and it refers to the computational power (e.g. the number of floating-point operations per second) per die area.
- the computational complexity is a figure of merit for scientific computation and it refers to the total number of built-in functions supported by a processor.
- the 2-D integration severely limits computational density and computational complexity.
- FIG. 1B lists all built-in transcendental functions supported by an Intel Itanium (IA-64) processor (referring to Harrison et al. “The Computation of Transcendental Functions on the IA-64 Architecture”, Intel Technical Journal, Q4 1999, hereinafter Harrison).
- the IA-64 processor supports a total of 7 built-in transcendental functions, each using a relatively small LUT (from 0 to 24 kb) in conjunction with a relatively high-degree Taylor-series calculation (from 5 to 22).
- the prevailing framework of scientific computation comprises three layers: a foundation layer, a function layer and a modeling layer.
- the foundation layer includes built-in functions that can be implemented by hardware.
- the function layer includes mathematical functions that cannot be implemented by hardware (e.g. non-basic non-arithmetic functions).
- the modeling layer includes mathematical models, which are the mathematical descriptions of the input-output characteristics of a system component.
- the mathematical functions in the function layer and the mathematical models in the modeling layer are implemented by software.
- the function layer involves one software-decomposition step: mathematical functions are decomposed into combinations of built-in functions by software, before these built-in functions and the associated arithmetic operations are calculated by hardware.
- the modeling layer involves two software-decomposition steps: the mathematical models are first decomposed into combinations of mathematical functions; then the mathematical functions are further decomposed into combinations of built-in functions.
- the software-implemented functions e.g. mathematical functions, mathematical models
- extra software-decomposition steps e.g. for mathematical models
- FIGS. 2A-2B disclose a simple example—the simulation of an amplifier circuit 20 .
- the amplifier circuit 20 comprises a transistor 24 and a resistor 22 ( FIG. 2A ).
- All transistor models e.g. MOS3, BSIM3 V3.2, BSIM4 V3.0, PSP of FIG. 2B ) model the transistor behaviors based on the small set of built-in functions provided by the conventional processor 300 .
- the present invention discloses a three-dimensional processor (3D-processor).
- the present invention discloses a three-dimensional processor (3D-processor). It comprises an array of computing elements, with each computing element comprising an arithmetic logic circuit (ALC) formed on the semiconductor substrate and at least a three-dimensional memory (3D-M) array stacked above the ALC.
- the 3D-M array stores at least a portion of a look-up table (LUT, or 3DM-LUT) for a mathematical function, while the ALC performs arithmetic operations on selected 3DM-LUT data.
- the mathematical function implemented by the computing element is a non-arithmetic function, which includes more operations than arithmetic operations performable by the ALC.
- the 3D-M array and the ALC are communicatively coupled through a plurality of contact vias.
- the present invention further discloses a memory-based computation (MBC), which carries out computation primarily with the 3DM-LUT.
- MBC memory-based computation
- the 3DM-LUT used by the MBC has a much larger capacity than the conventional LUT.
- arithmetic operations are still performed for most MBCs, using a larger LUT as a starting point, the MBC only needs to calculate a polynomial to a smaller order.
- the fraction of computation done by the 3DM-LUT is significantly more than the ALC.
- 3D-M array is stacked above the ALC, this type of vertical integration is referred to as three-dimensional (3-D) integration.
- the 3-D integration has a profound effect on the computational density. Because the 3D-M array does not occupy any substrate area, the footprint of the computing element is roughly equal to that of the ALC. However, the footprint of a conventional processor is roughly equal to the sum of the footprints of the LUT and the ALU. By moving the LUT from aside to above, the computing element becomes smaller. The 3D-processor would contain more computing elements, become more computationally powerful and support massive parallelism.
- the 3-D integration also has a profound effect on the computational complexity.
- the total LUT capacity is less than 100 kb.
- the total 3DM-LUT capacity for a 3D-processor could reach 100 Gb (for example, a 3D-XPoint die has a storage capacity of 128 Gb). Consequently, a single 3D-processor die could support as many as 10,000 built-in functions, which are orders of magnitude more than the conventional processor.
- the present invention discloses a three-dimensional processor (3D-processor) for implementing a mathematical function, comprising: a semiconductor substrate with transistors thereon; first and second three-dimensional memory (3D-M) arrays above said semiconductor substrate, wherein said first 3D-M array stores at least a first portion of a first look-up table (LUT) for a first non-arithmetic function; and, said second 3D-M array stores at least a second portion of a second LUT for a second non-arithmetic function; an arithmetic logic circuit (ALC) on said semiconductor substrate for performing at least an arithmetic operation on selected data from said first or second LUT; and, a plurality of contact vias for communicatively coupling said first or second 3D-M array with said ALC; wherein said mathematical function is a combination of at least said first and second non-arithmetic functions.
- LUT look-up table
- ALC arithmetic logic circuit
- the present invention further discloses a processor for implementing a mathematical function, comprising: at least first and second memory arrays on a memory level, wherein said first memory array stores at least a first portion of a first look-up table (LUT) for a first non-arithmetic function; and, said second memory array stores at least a second portion of a second LUT for a second non-arithmetic function; at least an arithmetic logic circuit (ALC) on a logic level for performing at least an arithmetic operation on selected data from said first LUT or said second LUT, wherein said logic level is a different physical level than said memory level; and means for communicatively coupling said memory level and said logic level; wherein said mathematical function is a combination of at least said first and second non-arithmetic functions.
- LUT look-up table
- ALC arithmetic logic circuit
- FIG. 1A is a schematic view of a conventional processor (prior art);
- FIG. 1B lists all transcendental functions supported by an Intel Itanium (IA-64) processor (prior art);
- FIG. 2A is a circuit block diagram of an amplifier circuit
- FIG. 2B lists number of operations to calculate a current-voltage (I-V) point for various transistor models (prior art);
- FIG. 3A is a block diagram of a preferred 3D-processor
- FIG. 3B is a block diagram of a preferred computing element
- FIGS. 4A-4C are the block diagrams of three preferred ALC
- FIG. 5A is a cross-sectional view of a preferred computing element comprising at least a three-dimensional writable memory (3D-W) array
- FIG. 5B is a cross-sectional view of a preferred computing element comprising at least a three-dimensional printed memory (3D-P) array
- FIG. 5C is a perspective view of a preferred computing element
- FIG. 6A is a schematic view of a 3D-M cell comprising a diode or a diode-like device
- FIG. 6B is a schematic view of a 3D-M cell comprising a transistor or a transistor-like device
- FIGS. 7A-7C are the substrate layout views of three preferred 3D-processors
- FIG. 8A is a block diagram of a first preferred computing element
- FIG. 8B is its substrate layout view
- FIG. 8C is a detailed circuit diagram of the first preferred computing element
- FIG. 9A is a block diagram of a second preferred computing element;
- FIG. 9B is its substrate-circuit layout view;
- FIG. 10A is a block diagram of a third preferred computing element; FIG. 10B is its substrate-circuit layout view.
- the phrase “mathematical functions” refer to non-arithmetic functions only; the phrase “memory” is used in its broadest sense to mean any semiconductor-based holding place for information, either permanent or temporary; the phrase “permanent” is used in its broadest sense to mean any long-term storage; the phrase “communicatively coupled” is used in its broadest sense to mean any coupling whereby information may be passed from one element to another element; the phrase “on the substrate” means the active elements of a circuit (e.g. transistors) are formed on the surface of the substrate, although the interconnects between these active elements are formed above the substrate and do not touch the substrate; the phrase “above the substrate” means the active elements (e.g.
- 3DM-LUT refers to the look-up table (LUT) stored in the three-dimensional memory (3D-M) array(s), or the physical LUT circuit in the form of the 3D-M array(s); the symbol “/” means a relationship of “and” or “or”.
- a preferred three-dimensional processor (3D-processor) 100 comprises an array of computing elements 110 - 1 , 110 - 2 . . . 110 - i . . . 110 -N ( FIG. 3A ), which could realize a same function or different functions.
- Each computing element 110 - i has one or more input variables 150 , and one or more output variables 190 ( FIG. 3B ).
- It comprises at least a three-dimensional memory (3D-M) array 170 for storing at least a portion of the 3DM-LUT for a mathematical function and an arithmetic logic circuit (ALC) 180 for performing arithmetic operations on selected 3DM-LUT data.
- 3D-M three-dimensional memory
- ALC arithmetic logic circuit
- the mathematical function implemented by the computing element 110 - i is a non-arithmetic function, which includes more operations than arithmetic operations performable by the ALC 180 .
- the ALC 180 and the 3D-M array 170 are communicatively coupled by 3D-connections 160 . Formed on a different physical level than the ALC 180 , the 3D-M array 170 is represented by dotted line in all figures.
- the 3D-processor 100 uses memory-based computation (MBC), which carries out computation primarily with the 3DM-LUT 170 .
- MBC memory-based computation
- LBC logic-based computation
- the 3DM-LUT 170 used by the MBC has a much larger capacity than the conventional LUT 370 .
- arithmetic operations are still performed for most MBCs, using a larger LUT as a starting point, the MBC only needs to calculate a polynomial to a smaller order.
- the fraction of computation done by the 3DM-LUT 170 could be more than the ALC 180 .
- FIGS. 4A-4C are the block diagrams of three preferred ALC 180 .
- the first preferred ALC 180 comprises an adder 180 A
- the second preferred ALC 180 comprises a multiplier 180 M
- the third preferred ALC 180 comprising a multiply-accumulator (MAC), which includes an adder 180 A and a multiplier 180 M.
- the preferred ALC 180 could perform integer arithmetic operations, fixed-point arithmetic operations, or floating-point arithmetic operations.
- the computing element 110 - i comprising different types of 3D-M are disclosed.
- 3D-M was disclosed in U.S. Pat. No. 5,835,396 issued to Zhang on Nov. 10, 1998. It comprises a plurality of vertically stacked memory levels formed on a semiconductor substrate, with each memory level comprising a plurality of 3D-M arrays. Each 3D-M array is a collection of 3D-M cells in a memory level that share at least one address line.
- 3D-M can be categorized into 3D-RAM (random access memory) and 3D-ROM (read-only memory).
- RAM random access memory
- ROM read-only memory
- Most common 3D-M is 3D-ROM.
- the 3D-ROM is further categorized into 3-D writable memory (3D-W) and 3-D printed memory (3D-P).
- 3D-W data can be electrically written (or, programmable). Based on the number of programmings allowed, a 3D-W can be categorized into three-dimensional one-time-programmable memory (3D-OTP) and three-dimensional multiple-time-programmable memory (3D-MTP). The 3D-OTP can be written once, while the 3D-MTP is electrically re-programmable.
- An exemplary 3D-MTP is 3D-XPoint.
- Other types of 3D-MTP include memristor, resistive random-access memory (RRAM or ReRAM), phase-change memory, programmable metallization cell (PMC), conductive-bridging random-access memory (CBRAM), and the like.
- the 3DM-LUT 170 can be configured in the field. This becomes even better when the 3D-MTP is used, as the 3DM-LUT 170 would become re-configured.
- 3D-P data are recorded thereto using a printing method during manufacturing. These data are fixedly recorded and cannot be changed after manufacturing.
- the printing methods include photo-lithography, nano-imprint, e-beam lithography, DUV lithography, and laser-programming, etc.
- An exemplary 3D-P is three-dimensional mask-programmed read-only memory (3D-MPROM), whose data are recorded by photo-lithography. Because electrical programming is not required, a memory cell in the 3D-P can be biased at a larger voltage during read than the 3D-W and therefore, the 3D-P is faster than the 3D-W.
- FIG. 5A discloses a preferred computing element 110 - i comprising at least a 3D-W array. It comprises a substrate circuit 0 K formed on the substrate 0 .
- the ALC 180 is a portion of the substrate circuit 0 K.
- a first memory level 16 A is stacked above the substrate circuit 0 K, with a second memory level 16 B stacked above the first memory level 16 A.
- the substrate circuit 0 K includes the peripheral circuits of the memory levels 16 A, 16 B. It comprises transistors 0 t and the associated interconnect 0 M.
- Each of the memory levels (e.g. 16 A, 16 B) comprises a plurality of first address lines (i.e. y-lines, e.g.
- the first and second memory levels 16 A, 16 B are coupled to the ALC 180 through contact vias 1 av , 3 av , respectively.
- the LUTs stored in all 3D-M arrays coupled to the ALC 180 are collectively referred to as the 3DM-LUT 170 .
- Coupling the 3DM-LUT 170 with the ALC 180 the contact vias 1 av , 3 av are collectively referred to as 3D-connections 160 .
- the 3D-W cell 5 aa comprises a programmable layer 12 and a diode layer 14 .
- the programmable layer 12 could be an antifuse layer (which can be programmed once and is used for the 3D-OTP) or a re-programmable layer (which is used for the 3D-MTP).
- the diode layer 14 is broadly interpreted as any layer whose resistance at the read voltage is substantially lower than when the applied voltage has a magnitude smaller than or polarity opposite to that of the read voltage.
- the diode could be a semiconductor diode (e.g. p-i-n silicon diode), or a metal-oxide (e.g. TiO 2 ) diode.
- FIG. 5B discloses a preferred computing element 110 - i comprising at least a 3D-P array. It has a structure similar to that of FIG. 5A except for the memory cells.
- 3D-P has at least two types of memory cells: a high-resistance 3D-P cell 5 aa , and a low-resistance 3D-P cell 5 ac .
- the low-resistance 3D-P cell 5 ac comprises a diode layer 14
- the high-resistance 3D-P cell 5 aa comprises at least a high-resistance layer 13 .
- the diode layer 14 is similar to that in the 3D-W.
- the high-resistance layer 13 could simply be a layer of insulating dielectric (e.g. silicon oxide, or silicon nitride). It is physically removed at the location of the low-resistance 3D-P cell 5 ac during manufacturing.
- insulating dielectric e.g. silicon oxide, or silicon nitride
- FIG. 5C is a perspective view of the preferred computing element 110 - i .
- the ALC 180 is formed on the substrate 0 .
- the 3DM-LUT 170 is vertically stacked above and at least partially covers the ALC 180 .
- the 3-D integration moves the 3DM-LUT 170 physically close to the ALC 180 . Because the contact vias 1 av , 3 av coupling them are short (on the order of an um in length) and numerous (thousands at least), the 3D-connections 160 have a much larger bandwidth than the conventional processor 300 .
- the interconnects coupling them are much longer (hundreds of ums in length) and fewer (hundreds at most).
- FIGS. 6A-6B show two types of the preferred 3D-M cell 5 ab .
- the 3D-M cell 5 ab comprises a variable resistor 12 and a diode (or a diode-like device) 14 .
- the variable resistor 12 is realized by the programmable layer of FIG. 5A . It can be varied during manufacturing or after manufacturing.
- the diode (or diode-like device) 14 is realized by the diode layer of FIG. 5A . It is broadly interpreted as any two-terminal device whose resistance at the read voltage is substantially lower than when the applied voltage has a magnitude smaller than or polarity opposite to that of the read voltage.
- the 3D-M cell 5 ab comprises a transistor or a transistor-like device 16 .
- the transistor or transistor-like device 16 is broadly interpreted as any three- (or, more-) terminal device whose resistance between the first and second terminals can be modulated by an electrical signal on a third terminal.
- the device 16 further comprises a floating gate 18 for storing electrical charge which represents the digital information stored in the 3D-M cell 5 ab .
- the devices 16 can be organized into NOR-arrays or NAND-arrays.
- the 3D-M could be categorized into horizontal 3D-M (e.g. 3D-XPoint) and vertical 3D-M (e.g. 3D-NAND).
- FIGS. 7A-7C the substrate layout views of three preferred computing elements 110 - i are shown.
- the ALC 180 is only coupled with a single 3D-M array 170 o and processes the 3DM-LUT data therefrom.
- the 3DM-LUT 170 is stored in the 3D-M array 170 o .
- the ALC 180 is covered by the 3D-M array 170 .
- the 3D-M array 170 o has four peripheral circuits, including X-decoders 15 o , 15 o ′ and Y-decoders 17 i , 17 o ′.
- the ALC 180 is bound by these four peripheral circuits.
- As the 3D-M array is stacked above the substrate circuit 0 K and does not occupy any substrate area, its projection on the substrate 0 is shown by dotted lines in this and following figures.
- the ALC 180 is coupled with four 3D-M arrays 170 a - 170 d and processes the 3DM-LUT data therefrom.
- the 3DM-LUT 170 is stored in four 3D-M arrays 170 a - 170 d .
- each 3D-M array e.g. 170 a
- has two peripheral circuits e.g. X-decoder 15 a and Y-decoder 17 a ).
- the ALC 180 is bound by eight peripheral circuits (including X-decoders 15 a - 15 d and Y-decoders 17 a - 17 d ) and located below four 3D-M arrays 170 a - 170 d . Hence, the ALC 180 of FIG. 7B could be four times as large as that of FIG. 7A .
- the ALC 180 is coupled with eight 3D-M arrays 170 a - 170 d , 170 w - 170 z and processes the 3DM-LUT data therefrom.
- the 3DM-LUT 170 is stored in eight 3D-M arrays 170 a - 170 d , 170 w - 170 z .
- These 3D-M arrays are divided into two sets: a first set 150 a includes four 3D-M arrays 170 a - 170 d , and a second set 150 b includes four 3D-M arrays 170 w - 170 z .
- a first component 180 a of the ALC 180 is formed below the four 3D-M arrays 170 a - 170 d of the first set 150 a .
- a second component 180 b of the ALC 180 is formed below the four 3D-M array 170 w - 170 z of the second set 150 b .
- adjacent peripheral circuits e.g. adjacent x-decoders 15 a , 15 c , or, adjacent y-decoders 17 a , 17 b ) are separated by physical gaps G.
- routing channel 182 , 184 , 186 which provide coupling between different components 180 a , 180 b , or between different ALCs 180 a , 180 b .
- the ALC 180 of FIG. 7C could be eight times as large as that of FIG. 7A .
- the 3D-M array 170 is stacked above the ALC 180 , this type of vertical integration is referred to as 3-D integration.
- the 3-D integration has a profound effect on the computational density of the 3D-processor 100 .
- the 3D-M array 170 does not occupy any substrate area 0 , the footprint of the computing element 110 - i is roughly equal to that of the ALC 180 .
- This is much smaller than a conventional processor 300 , whose footprint is roughly equal to the sum of the footprints of the LUT 370 and the ALC 380 .
- the 3D-processor 100 would contain more computing elements 110 - 1 , become more computationally powerful and support massive parallelism.
- the 3-D integration also has a profound effect on the computational complexity of the 3D-processor 100 .
- the total LUT capacity is less than 100 kb.
- the total 3DM-LUT capacity for a 3D-processor 100 could reach 100 Gb (for example, a 3D-XPoint die has a storage capacity of 128 Gb). Consequently, a single 3D-processor die 100 could support as many as 10,000 built-in functions, which are orders of magnitude more than the conventional processor 300 .
- FIG. 8A is its circuit block diagram.
- the ALC 180 comprises a pre-processing circuit 180 R, a 3DM-LUT 170 P, and a post-processing circuit 180 T.
- the pre-processing circuit 180 R converts the input variable (X) 150 into an address (A) of the 3DM-LUT 170 P.
- the post-processing circuit 180 T converts it into the function value (Y) 190 .
- a residue (R) of the input variable (X) is fed into the post-processing circuit 180 T to improve the calculation precision.
- FIG. 8B is its substrate-circuit layout view.
- the 3D-M storing the 3DM-LUT 170 P comprises at least a 3D-M array 170 p , as well as its X-decoder 15 p and Y-decoder 17 p .
- the 3D-M array 170 p covers the pre-processing circuit 180 R and the post-processing circuit 180 T. Although a single 3D-M array 170 p is shown in this figure, the preferred embodiment could use multiple 3D-M arrays, as those shown in FIGS. 7B-7C .
- the 3-D integration between the 3D-M array 170 p and the ALC 180 leads to a smaller footprint for the computing element 110 - i.
- the input variable X 150 has 32 bits (x 31 . . . x 0 ).
- the pre-processing circuit 180 R extracts the higher 16 bits (x 31 . . . x 16 ) thereof and sends it as a 16-bit address A to the 3DM-LUT 170 P.
- the pre-processing circuit 180 R further extracts the lower 16 bits (x 15 . . . x 0 ) and sends it as a 16-bit residue R to the post-processing circuit 180 T.
- the 3DM-LUT 170 P comprises two 3DM-LUTs 170 Q, 170 R.
- the post-processing circuit 180 T comprises a multiplier 180 M and an adder 180 A.
- the output value (Y) 190 has 32 bits and is calculated from polynomial interpolation.
- higher-order polynomial interpolation e.g. higher-order Taylor series
- higher-order polynomial interpolation can be used to improve the calculation precision.
- a single-precision function can be realized using a total of 4 Mb LUT (2 Mb for function values, and 2 Mb for first-derivative values) in conjunction with a first-order Taylor series calculation. This is significantly less than the LUT-only approach (4 Mb vs. 128 Gb).
- FIG. 9A is its schematic circuit block diagram.
- the preferred computing element 110 - i comprises two 3DM-LUTs 170 S, 170 T and a multiplier 180 M.
- the 3DM-LUT 170 S stores the Log( ) values, while the 3DM-LUT 170 T stores the Exp( ) values.
- the input variable X is used as an address 150 for the 3DM-LUT 170 S.
- the output Log(X) 160 a from the 3DM-LUT 170 S is multiplied by an exponent parameter K at the multiplier 180 M.
- FIG. 9B is its substrate-circuit layout view.
- the substrate circuit 0 K comprises the X-decoders 15 s , 15 t and the Y-decoders 17 s , 17 t for the 3D-M arrays 170 s , 170 t , as well as a multiplier 180 M. Placed side-by-side, both 3D-M arrays 170 s , 170 t partially cover the multiplier 180 M.
- both embodiments in FIG. 8C and FIG. 9A comprise two 3DM-LUTs. These 3DM-LUTs could be stored in a single 3D-M array 170 p (as in FIG.
- the 3DM-LUT can be stored in more 3D-M arrays.
- FIGS. 10A-10B a third preferred computing element 110 - i to simulate the amplifier circuit 20 of FIG. 2A is disclosed. It uses the model-by-LUT method.
- FIG. 10A is its schematic circuit block diagram.
- the preferred computing element 110 - i comprises a 3DM-LUT 170 U, an adder 180 A and a multiplier 180 M.
- the 3DM-LUT 170 U stores the data associated with the behaviors (e.g. input-output characteristics) of the transistor 24 .
- V IN input voltage value
- the readout 160 of the 3DM-LUT 170 U is the drain-current value (I D ).
- the multiplication result ( ⁇ R*I D ) is added to the V DD value by the adder 180 A to generate the output voltage value (V OUT ) 190 .
- the 3DM-LUT 170 U stores different forms of mathematical models.
- the mathematical model data stored in the 3DM-LUT 170 U is raw measurement data, i.e. the measured input-output characteristics of the transistor 24 .
- One example is the measured drain current vs. the applied gate-source voltage (I D ⁇ V GS ) characteristics.
- the mathematical model data stored in the 3DM-LUT 170 U is the smoothed measurement data.
- the raw measurement data could be smoothed using a purely mathematical method (e.g. a best-fit model). Or, this smoothing process can be aided by a physical transistor model (e.g. a BSIM4 V3.0 transistor model).
- the mathematical data stored in the 3DM-LUT include not only the measured data, but also its derivative values.
- the 3DM-LUT data include not only the drain-current values of the transistor 24 (e.g. the I D ⁇ V GS characteristics), but also its transconductance values (e.g. the G m ⁇ V GS characteristics).
- polynomial interpolation can be used to improve the modeling precision using a reasonable-size 3DM-LUT, as in the case of FIG. 8C .
- FIG. 10B is its substrate-circuit layout view.
- the substrate circuit 0 K comprises the X-decoder 15 u and the Y-decoder 17 u for the 3D-M array 170 u , as well as the multiplier 180 M and the adder 180 A.
- the 3D-M array 170 u covers the multiplier 180 M and the adder 180 A.
- a single 3D-M array 170 u is shown in this figure, the preferred embodiment could use multiple 3D-M arrays 170 u , as those shown in FIGS. 7B-7C .
- Model-by-LUT offers many advantages. By skipping two software-decomposition steps (from mathematical models to mathematical functions, and from mathematical functions to built-in functions), it saves substantial modeling time and energy. Model-by-LUT may need less LUT than function-by-LUT. Because a transistor model (e.g. BSIM4 V3.0) has hundreds of model parameters, calculating the intermediate functions of the transistor model requires extremely large LUTs. However, if we skip function-by-LUT (namely, skipping the transistor models and the associated intermediate functions), the transistor behaviors can be described using only three parameters (including the gate-source voltage V GS , the drain-source voltage V DS , and the body-source voltage V BS ). Describing the mathematical models of the transistor 24 requires relatively small LUTs.
- a transistor model e.g. BSIM4 V3.0
- calculating the intermediate functions of the transistor model requires extremely large LUTs.
- the transistor behaviors can be described using only three parameters (including the gate-source voltage V GS , the
- the processor could be a micro-controller, a controller, a central processing unit (CPU), a digital signal processor (DSP), a graphic processing unit (GPU), a network-security processor, an encryption/decryption processor, an encoding/decoding processor, a neural-network processor, or an artificial intelligence (AI) processor.
- CPU central processing unit
- DSP digital signal processor
- GPU graphic processing unit
- AI artificial intelligence
- processors can be found in consumer electronic devices (e.g. personal computers, video game machines, smart phones) as well as engineering and scientific workstations and server machines. The invention, therefore, is not to be limited except in the spirit of the appended claims.
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- Computer Hardware Design (AREA)
- General Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- Computing Systems (AREA)
- Mathematical Physics (AREA)
- Computational Mathematics (AREA)
- Mathematical Analysis (AREA)
- Mathematical Optimization (AREA)
- Pure & Applied Mathematics (AREA)
- Nonlinear Science (AREA)
- Semiconductor Integrated Circuits (AREA)
- Semiconductor Memories (AREA)
- Complex Calculations (AREA)
Abstract
A processor comprises an array of computing elements, with each computing element comprising an arithmetic logic circuit (ALC) and at least one three-dimensional memory (3D-M) array. The 3D-M array stores at least a portion of a look-up table (LUT) for a non-arithmetic function, while the ALC performs arithmetic operations on the LUT data. Because they include more operations than the basic arithmetic operations (i.e. addition, subtraction and multiplication), the non-arithmetic functions cannot be implemented by the conventional logic circuits alone.
Description
- This application is a continuation of U.S. patent application Ser. No. 15/487,366, filed Apr. 13, 2017, which claims priority from Chinese Patent Application 201610083747.7, filed on Feb. 13, 2016; Chinese Patent Application 201610260845.3, filed on Apr. 22, 2016; Chinese Patent Application 201610289592.2, filed on May 2, 2016; Chinese Patent Application 201710237780.5, filed on Apr. 12, 2017, in the State Intellectual Property Office of the People's Republic of China (CN), the disclosure of which are incorporated herein by references in their entireties.
- The present invention relates to the field of integrated circuit, and more particularly to processors.
- Conventional processors use logic-based computation (LBC), which carries out computation primarily with logic circuits (e.g. XOR circuit). Conventional logic circuits are suitable for arithmetic functions, whose operations include only basic arithmetic operations, i.e. the arithmetic operations performable by the conventional logic circuits per se. As is well known in the art, the basic arithmetic operations consist of addition, subtraction and multiplication. However, the conventional logic circuits are not suitable for non-arithmetic functions, whose operations include more than the basic arithmetic operations. Because they cannot be represented by a combination of the basic arithmetic operations, the non-arithmetic functions cannot be implemented by the conventional logic circuits alone. Exemplary non-arithmetic functions include transcendental functions and special functions. Non-arithmetic functions are computationally hard and their hardware implementation has been a major challenge.
- For the conventional processors, only few basic non-arithmetic functions (e.g. basic algebraic functions and basic transcendental functions) are implemented by hardware and they are referred to as built-in functions. These built-in functions are realized by a combination of logic circuits and look-up tables (LUT). For example, U.S. Pat. No. 5,954,787 issued to Eun on Sep. 21, 1999 taught a method for generating sine/cosine functions using LUTs; U.S. Pat. No. 9,207,910 issued to Azadet et al. on Dec. 8, 2015 taught a method for calculating a power function using LUTs.
- Realization of built-in functions is further illustrated in
FIG. 1A . Aconventional processor 300 generally comprises alogic circuit 380 and amemory circuit 370. Thelogic circuit 380 comprises an arithmetic logic unit (ALU) for performing arithmetic operations, while thememory circuit 370 stores an LUT for the built-in function. To obtain a desired precision, the built-in function is approximated to a polynomial of a sufficiently high order. TheLUT 370 stores the coefficients of the polynomial; and the ALU 380 calculates the polynomial. Because the ALU 380 and theLUT 370 are formed side-by-side on asemiconductor substrate 0, this type of horizontal integration is referred to as two-dimensional (2-D) integration. - Computation has been developed along the directions of computational density and computational complexity. The computational density is a figure of merit for parallel computation and it refers to the computational power (e.g. the number of floating-point operations per second) per die area. The computational complexity is a figure of merit for scientific computation and it refers to the total number of built-in functions supported by a processor. The 2-D integration severely limits computational density and computational complexity.
- For the 2-D integration, inclusion of the
LUT 370 increases the die size of theconventional processor 300 and lowers its computational density. This has an adverse effect on parallel computation. Moreover, because the ALU 380 is the primary component of theconventional processor 300 and occupies a large die area, theLUT 370 is left with a small die area and only supports few built-in functions.FIG. 1B lists all built-in transcendental functions supported by an Intel Itanium (IA-64) processor (referring to Harrison et al. “The Computation of Transcendental Functions on the IA-64 Architecture”, Intel Technical Journal,Q4 1999, hereinafter Harrison). The IA-64 processor supports a total of 7 built-in transcendental functions, each using a relatively small LUT (from 0 to 24 kb) in conjunction with a relatively high-degree Taylor-series calculation (from 5 to 22). - This small set of built-in functions (˜10 types, including arithmetic operations) is the foundation of scientific computation. Scientific computation uses advanced computing capabilities to advance human understandings and solve engineering problems. It has wide applications in computational mathematics, computational physics, computational chemistry, computational biology, computational engineering, computational economics, computational finance and other computational fields. The prevailing framework of scientific computation comprises three layers: a foundation layer, a function layer and a modeling layer. The foundation layer includes built-in functions that can be implemented by hardware. The function layer includes mathematical functions that cannot be implemented by hardware (e.g. non-basic non-arithmetic functions). The modeling layer includes mathematical models, which are the mathematical descriptions of the input-output characteristics of a system component.
- The mathematical functions in the function layer and the mathematical models in the modeling layer are implemented by software. The function layer involves one software-decomposition step: mathematical functions are decomposed into combinations of built-in functions by software, before these built-in functions and the associated arithmetic operations are calculated by hardware. The modeling layer involves two software-decomposition steps: the mathematical models are first decomposed into combinations of mathematical functions; then the mathematical functions are further decomposed into combinations of built-in functions. Apparently, the software-implemented functions (e.g. mathematical functions, mathematical models) run much slower and less efficient than the hardware-implemented functions (i.e. built-in functions), and extra software-decomposition steps (e.g. for mathematical models) would make these performance gaps even more pronounced.
- Because the arithmetic operations performable by the ALC consist of addition, subtraction and multiplication, the mathematical models that can be represented by the ALC alone are linear models only. Typical mathematical models are nonlinear and cannot be represented by a combination of these arithmetic operations. To illustrate how computationally intensive a mathematical model could be,
FIGS. 2A-2B disclose a simple example—the simulation of anamplifier circuit 20. Theamplifier circuit 20 comprises atransistor 24 and a resistor 22 (FIG. 2A ). All transistor models (e.g. MOS3, BSIM3 V3.2, BSIM4 V3.0, PSP ofFIG. 2B ) model the transistor behaviors based on the small set of built-in functions provided by theconventional processor 300. Due to the limited choice of the built-in functions, calculating even a single current-voltage (I-V) point for thetransistor 24 requires a large amount of computation (FIG. 2B ). As an example, the BSIM4 V3.0 transistor model needs 222 additions, 286 multiplications, 85 divisions, 16 square-root operations, 24 exponential operations, and 19 logarithmic operations. This large amount of computation makes simulation extremely slow and inefficient. - It is a principle object of the present invention to provide a paradigm shift for scientific computation.
- It is a further object of the present invention to provide a processor with improved computational complexity.
- It is a further object of the present invention to provide a processor with a large set of built-in functions.
- It is a further object of the present invention to realize non-arithmetic functions rapidly and efficiently.
- It is a further object of the present invention to realize rapid and efficient modeling and simulation.
- It is a further object of the present invention to provide a processor with improved computational density.
- In accordance with these and other objects of the present invention, the present invention discloses a three-dimensional processor (3D-processor).
- The present invention discloses a three-dimensional processor (3D-processor). It comprises an array of computing elements, with each computing element comprising an arithmetic logic circuit (ALC) formed on the semiconductor substrate and at least a three-dimensional memory (3D-M) array stacked above the ALC. The 3D-M array stores at least a portion of a look-up table (LUT, or 3DM-LUT) for a mathematical function, while the ALC performs arithmetic operations on selected 3DM-LUT data. The mathematical function implemented by the computing element is a non-arithmetic function, which includes more operations than arithmetic operations performable by the ALC. The 3D-M array and the ALC are communicatively coupled through a plurality of contact vias.
- The present invention further discloses a memory-based computation (MBC), which carries out computation primarily with the 3DM-LUT. Compared with the conventional logic-based computation (LBC), the 3DM-LUT used by the MBC has a much larger capacity than the conventional LUT. Although arithmetic operations are still performed for most MBCs, using a larger LUT as a starting point, the MBC only needs to calculate a polynomial to a smaller order. For the MBC, the fraction of computation done by the 3DM-LUT is significantly more than the ALC.
- Because the 3D-M array is stacked above the ALC, this type of vertical integration is referred to as three-dimensional (3-D) integration. The 3-D integration has a profound effect on the computational density. Because the 3D-M array does not occupy any substrate area, the footprint of the computing element is roughly equal to that of the ALC. However, the footprint of a conventional processor is roughly equal to the sum of the footprints of the LUT and the ALU. By moving the LUT from aside to above, the computing element becomes smaller. The 3D-processor would contain more computing elements, become more computationally powerful and support massive parallelism.
- The 3-D integration also has a profound effect on the computational complexity. For a conventional processor, the total LUT capacity is less than 100 kb. In contrast, the total 3DM-LUT capacity for a 3D-processor could reach 100 Gb (for example, a 3D-XPoint die has a storage capacity of 128 Gb). Consequently, a single 3D-processor die could support as many as 10,000 built-in functions, which are orders of magnitude more than the conventional processor.
- Significantly more built-in functions shall flatten the prevailing framework of scientific computation (including the foundation, function and modeling layers). The hardware-implemented functions, which were only available to the foundation layer, now become available to the function and modeling layers. Not only mathematical functions in the function layer can be directly realized by hardware, but also mathematical models in the modeling layer can be directly described by hardware. In the function layer, mathematical functions can be realized by a function-by-LUT method, i.e. the function values are calculated by reading the 3DM-LUT plus polynomial interpolation. In the modeling layer, mathematical models can be described by a model-by-LUT method, i.e. the input-output characteristics of a system component are modeled by reading the 3DM-LUT plus polynomial interpolation. Rapid and efficient computation would lead to a paradigm shift for scientific computation.
- Accordingly, the present invention discloses a three-dimensional processor (3D-processor) for implementing a mathematical function, comprising: a semiconductor substrate with transistors thereon; first and second three-dimensional memory (3D-M) arrays above said semiconductor substrate, wherein said first 3D-M array stores at least a first portion of a first look-up table (LUT) for a first non-arithmetic function; and, said second 3D-M array stores at least a second portion of a second LUT for a second non-arithmetic function; an arithmetic logic circuit (ALC) on said semiconductor substrate for performing at least an arithmetic operation on selected data from said first or second LUT; and, a plurality of contact vias for communicatively coupling said first or second 3D-M array with said ALC; wherein said mathematical function is a combination of at least said first and second non-arithmetic functions.
- The present invention further discloses a processor for implementing a mathematical function, comprising: at least first and second memory arrays on a memory level, wherein said first memory array stores at least a first portion of a first look-up table (LUT) for a first non-arithmetic function; and, said second memory array stores at least a second portion of a second LUT for a second non-arithmetic function; at least an arithmetic logic circuit (ALC) on a logic level for performing at least an arithmetic operation on selected data from said first LUT or said second LUT, wherein said logic level is a different physical level than said memory level; and means for communicatively coupling said memory level and said logic level; wherein said mathematical function is a combination of at least said first and second non-arithmetic functions.
-
FIG. 1A is a schematic view of a conventional processor (prior art);FIG. 1B lists all transcendental functions supported by an Intel Itanium (IA-64) processor (prior art); -
FIG. 2A is a circuit block diagram of an amplifier circuit;FIG. 2B lists number of operations to calculate a current-voltage (I-V) point for various transistor models (prior art); -
FIG. 3A is a block diagram of a preferred 3D-processor;FIG. 3B is a block diagram of a preferred computing element; -
FIGS. 4A-4C are the block diagrams of three preferred ALC; -
FIG. 5A is a cross-sectional view of a preferred computing element comprising at least a three-dimensional writable memory (3D-W) array;FIG. 5B is a cross-sectional view of a preferred computing element comprising at least a three-dimensional printed memory (3D-P) array;FIG. 5C is a perspective view of a preferred computing element; -
FIG. 6A is a schematic view of a 3D-M cell comprising a diode or a diode-like device;FIG. 6B is a schematic view of a 3D-M cell comprising a transistor or a transistor-like device; -
FIGS. 7A-7C are the substrate layout views of three preferred 3D-processors; -
FIG. 8A is a block diagram of a first preferred computing element;FIG. 8B is its substrate layout view;FIG. 8C is a detailed circuit diagram of the first preferred computing element; -
FIG. 9A is a block diagram of a second preferred computing element;FIG. 9B is its substrate-circuit layout view; -
FIG. 10A is a block diagram of a third preferred computing element;FIG. 10B is its substrate-circuit layout view. - It should be noted that all the drawings are schematic and not drawn to scale. Relative dimensions and proportions of parts of the device structures in the figures have been shown exaggerated or reduced in size for the sake of clarity and convenience in the drawings. The same reference symbols are generally used to refer to corresponding or similar features in the different embodiments.
- Throughout this specification, the phrase “mathematical functions” refer to non-arithmetic functions only; the phrase “memory” is used in its broadest sense to mean any semiconductor-based holding place for information, either permanent or temporary; the phrase “permanent” is used in its broadest sense to mean any long-term storage; the phrase “communicatively coupled” is used in its broadest sense to mean any coupling whereby information may be passed from one element to another element; the phrase “on the substrate” means the active elements of a circuit (e.g. transistors) are formed on the surface of the substrate, although the interconnects between these active elements are formed above the substrate and do not touch the substrate; the phrase “above the substrate” means the active elements (e.g. memory cells) are formed above the substrate and do not touch the substrate; the term “3DM-LUT” refers to the look-up table (LUT) stored in the three-dimensional memory (3D-M) array(s), or the physical LUT circuit in the form of the 3D-M array(s); the symbol “/” means a relationship of “and” or “or”.
- Those of ordinary skills in the art will realize that the following description of the present invention is illustrative only and is not intended to be in any way limiting. Other embodiments of the invention will readily suggest themselves to such skilled persons from an examination of the within disclosure.
- Referring now to
FIG. 3A-3B , a preferred three-dimensional processor (3D-processor) 100 is disclosed. It comprises an array of computing elements 110-1, 110-2 . . . 110-i . . . 110-N (FIG. 3A ), which could realize a same function or different functions. Each computing element 110-i has one ormore input variables 150, and one or more output variables 190 (FIG. 3B ). It comprises at least a three-dimensional memory (3D-M)array 170 for storing at least a portion of the 3DM-LUT for a mathematical function and an arithmetic logic circuit (ALC) 180 for performing arithmetic operations on selected 3DM-LUT data. The mathematical function implemented by the computing element 110-i is a non-arithmetic function, which includes more operations than arithmetic operations performable by theALC 180. TheALC 180 and the 3D-M array 170 are communicatively coupled by 3D-connections 160. Formed on a different physical level than theALC 180, the 3D-M array 170 is represented by dotted line in all figures. - The 3D-
processor 100 uses memory-based computation (MBC), which carries out computation primarily with the 3DM-LUT 170. Compared with the conventional logic-based computation (LBC), the 3DM-LUT 170 used by the MBC has a much larger capacity than theconventional LUT 370. Although arithmetic operations are still performed for most MBCs, using a larger LUT as a starting point, the MBC only needs to calculate a polynomial to a smaller order. For the MBC, the fraction of computation done by the 3DM-LUT 170 could be more than theALC 180. -
FIGS. 4A-4C are the block diagrams of threepreferred ALC 180. The firstpreferred ALC 180 comprises anadder 180A, the secondpreferred ALC 180 comprises amultiplier 180M, with the thirdpreferred ALC 180 comprising a multiply-accumulator (MAC), which includes anadder 180A and amultiplier 180M. Thepreferred ALC 180 could perform integer arithmetic operations, fixed-point arithmetic operations, or floating-point arithmetic operations. - Referring now to
FIGS. 5A-5C , the computing element 110-i comprising different types of 3D-M are disclosed. 3D-M was disclosed in U.S. Pat. No. 5,835,396 issued to Zhang on Nov. 10, 1998. It comprises a plurality of vertically stacked memory levels formed on a semiconductor substrate, with each memory level comprising a plurality of 3D-M arrays. Each 3D-M array is a collection of 3D-M cells in a memory level that share at least one address line. - 3D-M can be categorized into 3D-RAM (random access memory) and 3D-ROM (read-only memory). As used herein, the phrase “RAM” is used in its broadest sense to mean any memory for temporarily holding information, including but not limited to registers, SRAM, and DRAM; the phrase “ROM” is used in its broadest sense to mean any memory for permanently holding information, wherein the information being held could be either electrically alterable or un-alterable. Most common 3D-M is 3D-ROM. The 3D-ROM is further categorized into 3-D writable memory (3D-W) and 3-D printed memory (3D-P).
- For the 3D-W, data can be electrically written (or, programmable). Based on the number of programmings allowed, a 3D-W can be categorized into three-dimensional one-time-programmable memory (3D-OTP) and three-dimensional multiple-time-programmable memory (3D-MTP). The 3D-OTP can be written once, while the 3D-MTP is electrically re-programmable. An exemplary 3D-MTP is 3D-XPoint. Other types of 3D-MTP include memristor, resistive random-access memory (RRAM or ReRAM), phase-change memory, programmable metallization cell (PMC), conductive-bridging random-access memory (CBRAM), and the like. For the 3D-W, the 3DM-
LUT 170 can be configured in the field. This becomes even better when the 3D-MTP is used, as the 3DM-LUT 170 would become re-configured. - For the 3D-P, data are recorded thereto using a printing method during manufacturing. These data are fixedly recorded and cannot be changed after manufacturing. The printing methods include photo-lithography, nano-imprint, e-beam lithography, DUV lithography, and laser-programming, etc. An exemplary 3D-P is three-dimensional mask-programmed read-only memory (3D-MPROM), whose data are recorded by photo-lithography. Because electrical programming is not required, a memory cell in the 3D-P can be biased at a larger voltage during read than the 3D-W and therefore, the 3D-P is faster than the 3D-W.
-
FIG. 5A discloses a preferred computing element 110-i comprising at least a 3D-W array. It comprises a substrate circuit 0K formed on thesubstrate 0. TheALC 180 is a portion of the substrate circuit 0K. Afirst memory level 16A is stacked above the substrate circuit 0K, with asecond memory level 16B stacked above thefirst memory level 16A. The substrate circuit 0K includes the peripheral circuits of thememory levels second memory levels ALC 180 throughcontact vias 1 av, 3 av, respectively. The LUTs stored in all 3D-M arrays coupled to theALC 180 are collectively referred to as the 3DM-LUT 170. Coupling the 3DM-LUT 170 with theALC 180, thecontact vias 1 av, 3 av are collectively referred to as 3D-connections 160. - The 3D-
W cell 5 aa comprises aprogrammable layer 12 and adiode layer 14. Theprogrammable layer 12 could be an antifuse layer (which can be programmed once and is used for the 3D-OTP) or a re-programmable layer (which is used for the 3D-MTP). Thediode layer 14 is broadly interpreted as any layer whose resistance at the read voltage is substantially lower than when the applied voltage has a magnitude smaller than or polarity opposite to that of the read voltage. The diode could be a semiconductor diode (e.g. p-i-n silicon diode), or a metal-oxide (e.g. TiO2) diode. -
FIG. 5B discloses a preferred computing element 110-i comprising at least a 3D-P array. It has a structure similar to that ofFIG. 5A except for the memory cells. 3D-P has at least two types of memory cells: a high-resistance 3D-P cell 5 aa, and a low-resistance 3D-P cell 5 ac. The low-resistance 3D-P cell 5 ac comprises adiode layer 14, while the high-resistance 3D-P cell 5 aa comprises at least a high-resistance layer 13. Thediode layer 14 is similar to that in the 3D-W. The high-resistance layer 13, on the other hand, could simply be a layer of insulating dielectric (e.g. silicon oxide, or silicon nitride). It is physically removed at the location of the low-resistance 3D-P cell 5 ac during manufacturing. -
FIG. 5C is a perspective view of the preferred computing element 110-i. TheALC 180 is formed on thesubstrate 0. The 3DM-LUT 170 is vertically stacked above and at least partially covers theALC 180. The 3-D integration moves the 3DM-LUT 170 physically close to theALC 180. Because thecontact vias 1 av, 3 av coupling them are short (on the order of an um in length) and numerous (thousands at least), the 3D-connections 160 have a much larger bandwidth than theconventional processor 300. As the 2-D integration places theALU 380 and theLUT 370 side-by-side on thesubstrate 0, the interconnects coupling them are much longer (hundreds of ums in length) and fewer (hundreds at most). -
FIGS. 6A-6B show two types of the preferred 3D-M cell 5 ab. In the preferred embodiment ofFIG. 6A , the 3D-M cell 5 ab comprises avariable resistor 12 and a diode (or a diode-like device) 14. Thevariable resistor 12 is realized by the programmable layer ofFIG. 5A . It can be varied during manufacturing or after manufacturing. The diode (or diode-like device) 14 is realized by the diode layer ofFIG. 5A . It is broadly interpreted as any two-terminal device whose resistance at the read voltage is substantially lower than when the applied voltage has a magnitude smaller than or polarity opposite to that of the read voltage. - In the preferred embodiment of
FIG. 6B , the 3D-M cell 5 ab comprises a transistor or a transistor-like device 16. The transistor or transistor-like device 16 is broadly interpreted as any three- (or, more-) terminal device whose resistance between the first and second terminals can be modulated by an electrical signal on a third terminal. In this preferred embodiment, thedevice 16 further comprises a floatinggate 18 for storing electrical charge which represents the digital information stored in the 3D-M cell 5 ab. To those skilled in the art, thedevices 16 can be organized into NOR-arrays or NAND-arrays. Depending on the direction of the current flow between the first and second terminals in thedevices 16, the 3D-M could be categorized into horizontal 3D-M (e.g. 3D-XPoint) and vertical 3D-M (e.g. 3D-NAND). - Referring now to
FIGS. 7A-7C , the substrate layout views of three preferred computing elements 110-i are shown. In the embodiment ofFIG. 7A , theALC 180 is only coupled with a single 3D-M array 170 o and processes the 3DM-LUT data therefrom. The 3DM-LUT 170 is stored in the 3D-M array 170 o. TheALC 180 is covered by the 3D-M array 170. The 3D-M array 170 o has four peripheral circuits, including X-decoders 15 o, 15 o′ and Y-decoders 17 i, 17 o′. TheALC 180 is bound by these four peripheral circuits. As the 3D-M array is stacked above the substrate circuit 0K and does not occupy any substrate area, its projection on thesubstrate 0 is shown by dotted lines in this and following figures. - In the embodiment of
FIG. 7B , theALC 180 is coupled with four 3D-M arrays 170 a-170 d and processes the 3DM-LUT data therefrom. The 3DM-LUT 170 is stored in four 3D-M arrays 170 a-170 d. Different fromFIG. 7A , each 3D-M array (e.g. 170 a) has two peripheral circuits (e.g. X-decoder 15 a and Y-decoder 17 a). TheALC 180 is bound by eight peripheral circuits (including X-decoders 15 a-15 d and Y-decoders 17 a-17 d) and located below four 3D-M arrays 170 a-170 d. Apparently, theALC 180 ofFIG. 7B could be four times as large as that ofFIG. 7A . - In the embodiment of
FIG. 7C , theALC 180 is coupled with eight 3D-M arrays 170 a-170 d, 170 w-170 z and processes the 3DM-LUT data therefrom. The 3DM-LUT 170 is stored in eight 3D-M arrays 170 a-170 d, 170 w-170 z. These 3D-M arrays are divided into two sets: afirst set 150 a includes four 3D-M arrays 170 a-170 d, and asecond set 150 b includes four 3D-M arrays 170 w-170 z. Below the four 3D-M arrays 170 a-170 d of thefirst set 150 a, afirst component 180 a of theALC 180 is formed. Similarly, below the four 3D-M array 170 w-170 z of thesecond set 150 b, asecond component 180 b of theALC 180 is formed. In this embodiment, adjacent peripheral circuits (e.g.adjacent x-decoders decoders 17 a, 17 b) are separated by physical gaps G. These physical gaps allow the formation of therouting channel different components different ALCs ALC 180 ofFIG. 7C could be eight times as large as that ofFIG. 7A . - Because the 3D-
M array 170 is stacked above theALC 180, this type of vertical integration is referred to as 3-D integration. The 3-D integration has a profound effect on the computational density of the 3D-processor 100. Because the 3D-M array 170 does not occupy anysubstrate area 0, the footprint of the computing element 110-i is roughly equal to that of theALC 180. This is much smaller than aconventional processor 300, whose footprint is roughly equal to the sum of the footprints of theLUT 370 and theALC 380. By moving the LUT from aside to above, the computing element becomes smaller. The 3D-processor 100 would contain more computing elements 110-1, become more computationally powerful and support massive parallelism. - The 3-D integration also has a profound effect on the computational complexity of the 3D-
processor 100. For aconventional processor 300, the total LUT capacity is less than 100 kb. In contrast, the total 3DM-LUT capacity for a 3D-processor 100 could reach 100 Gb (for example, a 3D-XPoint die has a storage capacity of 128 Gb). Consequently, a single 3D-processor die 100 could support as many as 10,000 built-in functions, which are orders of magnitude more than theconventional processor 300. - Significantly more built-in functions shall flatten the prevailing framework of scientific computation (including the foundation, function and modeling layers). The hardware-implemented built-in functions, which were only available to the foundation layer, now become available to the function and modeling layers. Not only mathematical functions in the function layer can be directly realized by hardware (
FIGS. 8A-9B ), but also mathematical models in the modeling layer can be directly described by hardware (FIGS. 10A-10B ). In the function layer, mathematical functions can be realized by a function-by-LUT method, i.e. the function values are calculated by reading the 3DM-LUT data plus polynomial interpolation. In the modeling layer, mathematical models can be described by a model-by-LUT method, i.e. the input-output characteristics of a system component are modeled by reading the 3DM-LUT data plus polynomial interpolation. Rapid and efficient computation would lead to a paradigm shift for scientific computation. - Referring now to
FIGS. 8A-8C , a first preferred computing element 110-i implementing a built-in function Y=f(X) is disclosed. It uses the function-by-LUT method.FIG. 8A is its circuit block diagram. TheALC 180 comprises apre-processing circuit 180R, a 3DM-LUT 170P, and apost-processing circuit 180T. Thepre-processing circuit 180R converts the input variable (X) 150 into an address (A) of the 3DM-LUT 170P. After the data (D) at the address (A) is read out from the 3DM-LUT 170P, thepost-processing circuit 180T converts it into the function value (Y) 190. A residue (R) of the input variable (X) is fed into thepost-processing circuit 180T to improve the calculation precision. -
FIG. 8B is its substrate-circuit layout view. The 3D-M storing the 3DM-LUT 170P comprises at least a 3D-M array 170 p, as well as its X-decoder 15 p and Y-decoder 17 p. The 3D-M array 170 p covers thepre-processing circuit 180R and thepost-processing circuit 180T. Although a single 3D-M array 170 p is shown in this figure, the preferred embodiment could use multiple 3D-M arrays, as those shown inFIGS. 7B-7C . Because the 3D-M array 170 p does not occupy any substrate area, the 3-D integration between the 3D-M array 170 p and the ALC 180 (including thepre-processing circuit 180R and thepost-processing circuit 180T) leads to a smaller footprint for the computing element 110-i. -
FIG. 8C discloses the first preferred computing element 110-i which realizes a single-precision built-in function Y=f(X). The inputvariable X 150 has 32 bits (x31 . . . x0). Thepre-processing circuit 180R extracts the higher 16 bits (x31 . . . x16) thereof and sends it as a 16-bit address A to the 3DM-LUT 170P. Thepre-processing circuit 180R further extracts the lower 16 bits (x15 . . . x0) and sends it as a 16-bit residue R to thepost-processing circuit 180T. The 3DM-LUT 170P comprises two 3DM-LUTs LUTs LUT 170Q stores the functional value D1=f(A), while the 3DM-LUT 170R stores the first-order derivative value D2=f′(A). Thepost-processing circuit 180T comprises amultiplier 180M and anadder 180A. The output value (Y) 190 has 32 bits and is calculated from polynomial interpolation. In this case, the polynomial interpolation is a first-order Taylor series: Y(X)=D1+D2*R==f(A)+f′(A)*R. To those skilled in the art, higher-order polynomial interpolation (e.g. higher-order Taylor series) can be used to improve the calculation precision. - When calculating a built-in function, combining the LUT with polynomial interpolation can achieve a high precision without using an excessively large LUT. For example, if only LUT (without any polynomial interpolation) is used to realize a single-precision function (32-bit input and 32-bit output), it would have a capacity of 232*32=128 Gb, which is impractical. By including polynomial interpolation, significantly smaller LUTs can be used. In the above embodiment, a single-precision function can be realized using a total of 4 Mb LUT (2 Mb for function values, and 2 Mb for first-derivative values) in conjunction with a first-order Taylor series calculation. This is significantly less than the LUT-only approach (4 Mb vs. 128 Gb).
- Referring now to
FIGS. 9A-9B , a second preferred computing element 110-i implementing a composite function Y=exp[K*log(X)]=XK is disclosed. It uses the function-by-LUT method.FIG. 9A is its schematic circuit block diagram. The preferred computing element 110-i comprises two 3DM-LUTs multiplier 180M. The 3DM-LUT 170S stores the Log( ) values, while the 3DM-LUT 170T stores the Exp( ) values. The input variable X is used as anaddress 150 for the 3DM-LUT 170S. The output Log(X) 160 a from the 3DM-LUT 170S is multiplied by an exponent parameter K at themultiplier 180M. The multiplication result K*Log(X) is used as anaddress 160 b for the 3DM-LUT 170T, whoseoutput 190 is Y=XK. -
FIG. 9B is its substrate-circuit layout view. The substrate circuit 0K comprises the X-decoders 15 s, 15 t and the Y-decoders M arrays 170 s, 170 t, as well as amultiplier 180M. Placed side-by-side, both 3D-M arrays 170 s, 170 t partially cover themultiplier 180M. Note that both embodiments inFIG. 8C andFIG. 9A comprise two 3DM-LUTs. These 3DM-LUTs could be stored in a single 3D-M array 170 p (as inFIG. 8B ), in two 3D-M arrays 170 s, 170 t placed side-by-side (as inFIG. 9B ), or in two vertically stacked 3D-M arrays (i.e. ondifferent memory levels FIGS. 5A-5C ). Apparently, the 3DM-LUT can be stored in more 3D-M arrays. - Referring now to
FIGS. 10A-10B , a third preferred computing element 110-i to simulate theamplifier circuit 20 ofFIG. 2A is disclosed. It uses the model-by-LUT method.FIG. 10A is its schematic circuit block diagram. The preferred computing element 110-i comprises a 3DM-LUT 170U, anadder 180A and amultiplier 180M. The 3DM-LUT 170U stores the data associated with the behaviors (e.g. input-output characteristics) of thetransistor 24. By using the input voltage value (VIN) as anaddress 150 for the 3DM-LUT 170U, thereadout 160 of the 3DM-LUT 170U is the drain-current value (ID). After the ID value is multiplied with the minus resistance value (−R) of theresistor 22 by themultiplier 180M, the multiplication result (−R*ID) is added to the VDD value by theadder 180A to generate the output voltage value (VOUT) 190. - The 3DM-
LUT 170U stores different forms of mathematical models. In one case, the mathematical model data stored in the 3DM-LUT 170U is raw measurement data, i.e. the measured input-output characteristics of thetransistor 24. One example is the measured drain current vs. the applied gate-source voltage (ID−VGS) characteristics. In another case, the mathematical model data stored in the 3DM-LUT 170U is the smoothed measurement data. The raw measurement data could be smoothed using a purely mathematical method (e.g. a best-fit model). Or, this smoothing process can be aided by a physical transistor model (e.g. a BSIM4 V3.0 transistor model). In a third case, the mathematical data stored in the 3DM-LUT include not only the measured data, but also its derivative values. For example, the 3DM-LUT data include not only the drain-current values of the transistor 24 (e.g. the ID−VGS characteristics), but also its transconductance values (e.g. the Gm−VGS characteristics). With derivative values, polynomial interpolation can be used to improve the modeling precision using a reasonable-size 3DM-LUT, as in the case ofFIG. 8C . -
FIG. 10B is its substrate-circuit layout view. The substrate circuit 0K comprises the X-decoder 15 u and the Y-decoder 17 u for the 3D-M array 170 u, as well as themultiplier 180M and theadder 180A. The 3D-M array 170 u covers themultiplier 180M and theadder 180A. Although a single 3D-M array 170 u is shown in this figure, the preferred embodiment could use multiple 3D-M arrays 170 u, as those shown inFIGS. 7B-7C . - Model-by-LUT offers many advantages. By skipping two software-decomposition steps (from mathematical models to mathematical functions, and from mathematical functions to built-in functions), it saves substantial modeling time and energy. Model-by-LUT may need less LUT than function-by-LUT. Because a transistor model (e.g. BSIM4 V3.0) has hundreds of model parameters, calculating the intermediate functions of the transistor model requires extremely large LUTs. However, if we skip function-by-LUT (namely, skipping the transistor models and the associated intermediate functions), the transistor behaviors can be described using only three parameters (including the gate-source voltage VGS, the drain-source voltage VDS, and the body-source voltage VBS). Describing the mathematical models of the
transistor 24 requires relatively small LUTs. - While illustrative embodiments have been shown and described, it would be apparent to those skilled in the art that many more modifications than that have been mentioned above are possible without departing from the inventive concepts set forth therein. For example, the processor could be a micro-controller, a controller, a central processing unit (CPU), a digital signal processor (DSP), a graphic processing unit (GPU), a network-security processor, an encryption/decryption processor, an encoding/decoding processor, a neural-network processor, or an artificial intelligence (AI) processor. These processors can be found in consumer electronic devices (e.g. personal computers, video game machines, smart phones) as well as engineering and scientific workstations and server machines. The invention, therefore, is not to be limited except in the spirit of the appended claims.
Claims (20)
1. A three-dimensional processor (3D-processor) for implementing a mathematical function, comprising:
a semiconductor substrate with transistors thereon;
first and second three-dimensional memory (3D-M) arrays above said semiconductor substrate, wherein said first 3D-M array stores at least a first portion of a first look-up table (LUT) for a first non-arithmetic function; and, said second 3D-M array stores at least a second portion of a second LUT for a second non-arithmetic function;
an arithmetic logic circuit (ALC) on said semiconductor substrate for performing at least an arithmetic operation on selected data from said first or second LUT; and,
a plurality of contact vias for communicatively coupling said first or second 3D-M array with said ALC;
wherein said mathematical function is a combination of at least said first and second non-arithmetic functions.
2. The processor according to claim 1 , wherein said first or second non-arithmetic function includes more operations than arithmetic operations performable by said ALC.
3. The processor according to claim 2 , wherein said arithmetic operations performable by said ALC consist of addition, subtraction and multiplication.
4. The processor according to claim 1 , wherein said first or second non-arithmetic function cannot be represented by a combination of arithmetic operations performable by said ALC.
5. The processor according to claim 4 , wherein said arithmetic operations performable by said ALC consist of addition, subtraction and multiplication.
6. The processor according to claim 1 , wherein said first or second 3D-M array at least partially covers said ALC.
7. The processor according to claim 1 , wherein said first LUT includes functional values of said mathematical function.
8. The processor according to claim 1 , wherein said second LUT includes derivative values of said mathematical function.
9. The processor according to claim 1 , wherein said mathematical function is a composite function of said first and second non-arithmetic functions.
10. The processor according to claim 1 , wherein said processor is a micro-controller, a controller, a central processing unit (CPU), a digital signal processor (DSP), a graphic processing unit (GPU), a network-security processor, an encryption/decryption processor, an encoding/decoding processor, a neural-network processor, or an artificial intelligence (AI) processor.
11. A processor for implementing a mathematical function, comprising:
at least first and second memory arrays on a memory level, wherein said first memory array stores at least a first portion of a first look-up table (LUT) for a first non-arithmetic function; and, said second memory array stores at least a second portion of a second LUT for a second non-arithmetic function;
at least an arithmetic logic circuit (ALC) on a logic level for performing at least an arithmetic operation on selected data from said first LUT or said second LUT, wherein said logic level is a different physical level than said memory level; and
means for communicatively coupling said memory level and said logic level;
wherein said mathematical function is a combination of at least said first and second non-arithmetic functions.
12. The processor according to claim 11 , wherein said first or second non-arithmetic function includes more operations than arithmetic operations performable by said ALC.
13. The processor according to claim 12 , wherein said arithmetic operations performable by said ALC consist of addition, subtraction and multiplication.
14. The processor according to claim 11 , wherein said first or second non-arithmetic function cannot be represented by a combination of arithmetic operations performable by said ALC.
15. The processor according to claim 14 , wherein said arithmetic operations performable by said ALC consist of addition, subtraction and multiplication.
16. The processor according to claim 11 , further comprising a semiconductor substrate, wherein said logic level is disposed on said semiconductor substrate; said memory level is disposed above said substrate; and, said logic level and said memory level are communicatively coupled by a plurality of contact vias.
17. The processor according to claim 11 , wherein said first LUT includes the functional values of said mathematical function.
18. The processor according to claim 11 , wherein said second LUT includes the derivative values of said mathematical function.
19. The processor according to claim 11 , wherein said mathematical function is a composite function of said first and second non-arithmetic functions.
20. The processor according to claim 11 , wherein said processor is a micro-controller, a controller, a central processing unit (CPU), a digital signal processor (DSP), a graphic processing unit (GPU), a network-security processor, an encryption/decryption processor, an encoding/decoding processor, a neural-network processor, or an artificial intelligence (AI) processor.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
US16/207,189 US20190115923A1 (en) | 2016-02-13 | 2018-12-03 | Processor for Implementing Non-Arithmetic Functions |
Applications Claiming Priority (9)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201610083747.7A CN107085452B (en) | 2016-02-13 | 2016-02-13 | Three-dimensional printed memory (3D-P) based processor |
CN201610260845 | 2016-04-22 | ||
CN201610260845.3 | 2016-04-22 | ||
CN201610289592 | 2016-05-02 | ||
CN201610289592.2 | 2016-05-02 | ||
CN201710237780.5 | 2017-04-12 | ||
CN201710237780 | 2017-04-12 | ||
US15/487,366 US10763861B2 (en) | 2016-02-13 | 2017-04-13 | Processor comprising three-dimensional memory (3D-M) array |
US16/207,189 US20190115923A1 (en) | 2016-02-13 | 2018-12-03 | Processor for Implementing Non-Arithmetic Functions |
Related Parent Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
US15/487,366 Continuation US10763861B2 (en) | 2016-02-13 | 2017-04-13 | Processor comprising three-dimensional memory (3D-M) array |
Publications (1)
Publication Number | Publication Date |
---|---|
US20190115923A1 true US20190115923A1 (en) | 2019-04-18 |
Family
ID=59561801
Family Applications (3)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
US15/487,366 Active 2037-07-15 US10763861B2 (en) | 2016-02-13 | 2017-04-13 | Processor comprising three-dimensional memory (3D-M) array |
US16/200,630 Abandoned US20190115922A1 (en) | 2016-02-13 | 2018-11-26 | Processor For Implementing Mathematical Functions or Models |
US16/207,189 Abandoned US20190115923A1 (en) | 2016-02-13 | 2018-12-03 | Processor for Implementing Non-Arithmetic Functions |
Family Applications Before (2)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
US15/487,366 Active 2037-07-15 US10763861B2 (en) | 2016-02-13 | 2017-04-13 | Processor comprising three-dimensional memory (3D-M) array |
US16/200,630 Abandoned US20190115922A1 (en) | 2016-02-13 | 2018-11-26 | Processor For Implementing Mathematical Functions or Models |
Country Status (2)
Country | Link |
---|---|
US (3) | US10763861B2 (en) |
WO (1) | WO2017137015A2 (en) |
Cited By (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US11527523B2 (en) * | 2018-12-10 | 2022-12-13 | HangZhou HaiCun Information Technology Co., Ltd. | Discrete three-dimensional processor |
Families Citing this family (10)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US10445067B2 (en) * | 2016-05-06 | 2019-10-15 | HangZhou HaiCun Information Technology Co., Ltd. | Configurable processor with in-package look-up table |
KR102408858B1 (en) | 2017-12-19 | 2022-06-14 | 삼성전자주식회사 | A nonvolatile memory device, a memory system including the same and a method of operating a nonvolatile memory device |
US10628295B2 (en) | 2017-12-26 | 2020-04-21 | Samsung Electronics Co., Ltd. | Computing mechanisms using lookup tables stored on memory |
US11398453B2 (en) * | 2018-01-09 | 2022-07-26 | Samsung Electronics Co., Ltd. | HBM silicon photonic TSV architecture for lookup computing AI accelerator |
CN108599849B (en) * | 2018-04-14 | 2021-01-01 | 上海交通大学 | Photon processing system and processing method for intelligent decision |
US11296068B2 (en) * | 2018-12-10 | 2022-04-05 | HangZhou HaiCun Information Technology Co., Ltd. | Discrete three-dimensional processor |
CN113918506A (en) * | 2018-12-10 | 2022-01-11 | 杭州海存信息技术有限公司 | Discrete three-dimensional processor |
US10901694B2 (en) * | 2018-12-31 | 2021-01-26 | Micron Technology, Inc. | Binary parallel adder and multiplier |
CN112287632A (en) * | 2020-10-26 | 2021-01-29 | 成都华微电子科技有限公司 | Method for prejudging wiring of integrated circuit |
US11709790B2 (en) * | 2021-02-24 | 2023-07-25 | Xilinx, Inc. | Spatial distribution in a 3D data processing unit |
Family Cites Families (36)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US4870302A (en) | 1984-03-12 | 1989-09-26 | Xilinx, Inc. | Configurable electrical circuit having configurable logic elements and configurable interconnects |
US5046038A (en) | 1989-07-07 | 1991-09-03 | Cyrix Corporation | Method and apparatus for performing division using a rectangular aspect ratio multiplier |
US5060182A (en) | 1989-09-05 | 1991-10-22 | Cyrix Corporation | Method and apparatus for performing the square root function using a rectangular aspect ratio multiplier |
US5604499A (en) | 1993-12-28 | 1997-02-18 | Matsushita Electric Industrial Co., Ltd. | Variable-length decoding apparatus |
US5901274A (en) | 1994-04-30 | 1999-05-04 | Samsung Electronics Co. Ltd. | Method for enlargement/reduction of image data in digital image processing system and circuit adopting the same |
US5835396A (en) | 1996-10-17 | 1998-11-10 | Zhang; Guobiao | Three-dimensional read-only memory |
KR100202956B1 (en) | 1996-12-26 | 1999-06-15 | 전주범 | Triangle function lookup table access method and apparatus of digital signal process system |
US6263470B1 (en) | 1998-02-03 | 2001-07-17 | Texas Instruments Incorporated | Efficient look-up table methods for Reed-Solomon decoding |
WO2000004436A1 (en) | 1998-07-17 | 2000-01-27 | Intergraph Corporation | Graphics processing with transcendental function generator |
US7366748B1 (en) | 2000-06-30 | 2008-04-29 | Intel Corporation | Methods and apparatus for fast argument reduction in a computing system |
US7206410B2 (en) | 2001-10-10 | 2007-04-17 | Stmicroelectronics S.R.L. | Circuit for the inner or scalar product computation in Galois fields |
US20040044710A1 (en) | 2002-08-28 | 2004-03-04 | Harrison John R. | Converting mathematical functions to power series |
US7028247B2 (en) | 2002-12-25 | 2006-04-11 | Faraday Technology Corp. | Error correction code circuit with reduced hardware complexity |
JP4199100B2 (en) | 2003-12-12 | 2008-12-17 | 富士通株式会社 | Function calculation method and function calculation circuit |
JP3845636B2 (en) | 2004-01-21 | 2006-11-15 | 株式会社東芝 | Function approximation calculator |
US20060106905A1 (en) | 2004-11-17 | 2006-05-18 | Chren William A Jr | Method for reducing memory size in logarithmic number system arithmetic units |
US7512647B2 (en) | 2004-11-22 | 2009-03-31 | Analog Devices, Inc. | Condensed Galois field computing system |
US7574468B1 (en) | 2005-03-18 | 2009-08-11 | Verisilicon Holdings (Cayman Islands) Co. Ltd. | Digital signal processor having inverse discrete cosine transform engine for video decoding and partitioned distributed arithmetic multiply/accumulate unit therefor |
US7539927B2 (en) | 2005-04-14 | 2009-05-26 | Industrial Technology Research Institute | High speed hardware implementation of modified Reed-Solomon decoder |
US8203564B2 (en) | 2007-02-16 | 2012-06-19 | Qualcomm Incorporated | Efficient 2-D and 3-D graphics processing |
US8106918B2 (en) | 2007-05-01 | 2012-01-31 | Vivante Corporation | Apparatus and method for texture level of detail computation |
US7962543B2 (en) | 2007-06-01 | 2011-06-14 | Advanced Micro Devices, Inc. | Division with rectangular multiplier supporting multiple precisions and operand types |
US20100140750A1 (en) * | 2008-12-10 | 2010-06-10 | Qualcomm Incorporated | Parallel Plane Memory and Processor Coupling in a 3-D Micro-Architectural System |
US9207910B2 (en) | 2009-01-30 | 2015-12-08 | Intel Corporation | Digital signal processor having instruction set with an xK function using reduced look-up table |
US9015452B2 (en) | 2009-02-18 | 2015-04-21 | Texas Instruments Incorporated | Vector math instruction execution by DSP processor approximating division and complex number magnitude |
KR101728068B1 (en) * | 2010-06-01 | 2017-04-19 | 삼성전자 주식회사 | Stacked semiconductor memory device, memory system including the same, and method of repairing defects of through silicon vias |
US9363068B2 (en) | 2010-08-03 | 2016-06-07 | Intel Corporation | Vector processor having instruction set with sliding window non-linear convolutional function |
US8273610B2 (en) * | 2010-11-18 | 2012-09-25 | Monolithic 3D Inc. | Method of constructing a semiconductor device and structure |
CN103959192B (en) | 2011-12-21 | 2017-11-21 | 英特尔公司 | For estimating the mathematical circuit surmounted function |
CN104205234B (en) * | 2012-03-30 | 2017-07-11 | 英特尔公司 | For the conventional data scrambler of memory test circuit engine |
CN103633091B (en) * | 2012-08-22 | 2016-03-30 | 成都海存艾匹科技有限公司 | Three-dimensional storage containing integrated intermediate circuit chip |
US9753695B2 (en) | 2012-09-04 | 2017-09-05 | Analog Devices Global | Datapath circuit for digital signal processors |
US8737108B2 (en) * | 2012-09-25 | 2014-05-27 | Intel Corporation | 3D memory configurable for performance and power |
US9287196B2 (en) * | 2012-12-28 | 2016-03-15 | Intel Corporation | Resonant clocking for three-dimensional stacked devices |
US9385058B1 (en) * | 2012-12-29 | 2016-07-05 | Monolithic 3D Inc. | Semiconductor device and structure |
US9606796B2 (en) | 2013-10-30 | 2017-03-28 | Texas Instruments Incorporated | Computer and methods for solving math functions |
-
2017
- 2017-04-13 US US15/487,366 patent/US10763861B2/en active Active
- 2017-04-13 WO PCT/CN2017/080462 patent/WO2017137015A2/en active Application Filing
-
2018
- 2018-11-26 US US16/200,630 patent/US20190115922A1/en not_active Abandoned
- 2018-12-03 US US16/207,189 patent/US20190115923A1/en not_active Abandoned
Cited By (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US11527523B2 (en) * | 2018-12-10 | 2022-12-13 | HangZhou HaiCun Information Technology Co., Ltd. | Discrete three-dimensional processor |
Also Published As
Publication number | Publication date |
---|---|
US20170237440A1 (en) | 2017-08-17 |
US10763861B2 (en) | 2020-09-01 |
WO2017137015A3 (en) | 2017-10-05 |
WO2017137015A2 (en) | 2017-08-17 |
US20190115922A1 (en) | 2019-04-18 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
US20190115923A1 (en) | Processor for Implementing Non-Arithmetic Functions | |
Li et al. | Analogue signal and image processing with large memristor crossbars | |
Zidan et al. | A general memristor-based partial differential equation solver | |
Chen et al. | Efficient in-memory computing architecture based on crossbar arrays | |
US11068771B2 (en) | Integrated neuro-processor comprising three-dimensional memory array | |
Chen et al. | Multiply accumulate operations in memristor crossbar arrays for analog computing | |
Haj-Ali et al. | Imaging: In-memory algorithms for image processing | |
US20170323042A1 (en) | Simulation Processor with Backside Look-Up Table | |
US20190114170A1 (en) | Processor Using Memory-Based Computation | |
CN107305594B (en) | Processor containing three-dimensional memory array | |
US11776944B2 (en) | Discrete three-dimensional processor | |
US20170323041A1 (en) | Simulation Processor with In-Package Look-Up Table | |
US20190114139A1 (en) | Configurable Processor with Backside Look-Up Table | |
US10372359B2 (en) | Processor for realizing at least two categories of functions | |
US11128303B2 (en) | Three-dimensional memory (3D-M)-based configurable processor singlet | |
Eshraghian et al. | A behavioral model of digital resistive switching for systems level DNN acceleration | |
Liu et al. | Cellular automata imbedded memristor-based recirculated logic in-memory computing | |
US11080229B2 (en) | Processor for calculating mathematical functions in parallel | |
US11966715B2 (en) | Three-dimensional processor for parallel computing | |
CN112328535B (en) | Processor with three-dimensional memory array | |
Amin et al. | IMAC-Sim: A Circuit-level Simulator For In-Memory Analog Computing Architectures | |
Zhao et al. | A Multi-Bit Non-Volatile Compute-in-Memory Architecture with Quantum-Dot Transistor Based Unit | |
Bengel et al. | Devices and Architectures for Efficient Computing In-Memory (CIM) Design | |
Cubero-Cascante et al. | A Calibratable Model for Fast Energy Estimation of MVM Operations on RRAM Crossbars | |
Marinella et al. | Neuromorphic Algorithm Acceleration with Resistive Memory NanoCrossbars. |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
STPP | Information on status: patent application and granting procedure in general |
Free format text: NON FINAL ACTION MAILED |
|
STPP | Information on status: patent application and granting procedure in general |
Free format text: RESPONSE TO NON-FINAL OFFICE ACTION ENTERED AND FORWARDED TO EXAMINER |
|
STPP | Information on status: patent application and granting procedure in general |
Free format text: FINAL REJECTION MAILED |
|
STCB | Information on status: application discontinuation |
Free format text: ABANDONED -- FAILURE TO RESPOND TO AN OFFICE ACTION |