Classifications

• • G—PHYSICS
  • G06—COMPUTING OR CALCULATING; COUNTING
  • G06F—ELECTRIC DIGITAL DATA PROCESSING
  • G06F13/00—Interconnection of, or transfer of information or other signals between, memories, input/output devices or central processing units
  • G06F13/14—Handling requests for interconnection or transfer
  • G06F13/16—Handling requests for interconnection or transfer for access to memory bus
• • G—PHYSICS
  • G06—COMPUTING OR CALCULATING; COUNTING
  • G06F—ELECTRIC DIGITAL DATA PROCESSING
  • G06F13/00—Interconnection of, or transfer of information or other signals between, memories, input/output devices or central processing units
  • G06F13/14—Handling requests for interconnection or transfer
  • G06F13/16—Handling requests for interconnection or transfer for access to memory bus
  • G06F13/1668—Details of memory controller
  • G06F13/1694—Configuration of memory controller to different memory types
• • G—PHYSICS
  • G06—COMPUTING OR CALCULATING; COUNTING
  • G06F—ELECTRIC DIGITAL DATA PROCESSING
  • G06F13/00—Interconnection of, or transfer of information or other signals between, memories, input/output devices or central processing units
  • G06F13/14—Handling requests for interconnection or transfer
  • G06F13/20—Handling requests for interconnection or transfer for access to input/output bus
• • G—PHYSICS
  • G06—COMPUTING OR CALCULATING; 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
• • G—PHYSICS
  • G06—COMPUTING OR CALCULATING; 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 OR CALCULATING; COUNTING
  • G06F—ELECTRIC DIGITAL DATA PROCESSING
  • G06F3/00—Input arrangements for transferring data to be processed into a form capable of being handled by the computer; Output arrangements for transferring data from processing unit to output unit, e.g. interface arrangements
  • G06F3/06—Digital input from, or digital output to, record carriers, e.g. RAID, emulated record carriers or networked record carriers
  • G06F3/0601—Interfaces specially adapted for storage systems
  • G06F3/0628—Interfaces specially adapted for storage systems making use of a particular technique
  • G06F3/0629—Configuration or reconfiguration of storage systems
  • G06F3/0631—Configuration or reconfiguration of storage systems by allocating resources to storage systems
• • G—PHYSICS
  • G06—COMPUTING OR CALCULATING; COUNTING
  • G06F—ELECTRIC DIGITAL DATA PROCESSING
  • G06F3/00—Input arrangements for transferring data to be processed into a form capable of being handled by the computer; Output arrangements for transferring data from processing unit to output unit, e.g. interface arrangements
  • G06F3/06—Digital input from, or digital output to, record carriers, e.g. RAID, emulated record carriers or networked record carriers
  • G06F3/0601—Interfaces specially adapted for storage systems
  • G06F3/0628—Interfaces specially adapted for storage systems making use of a particular technique
  • G06F3/0646—Horizontal data movement in storage systems, i.e. moving data in between storage devices or systems
  • G06F3/065—Replication mechanisms
• • G—PHYSICS
  • G06—COMPUTING OR CALCULATING; COUNTING
  • G06F—ELECTRIC DIGITAL DATA PROCESSING
  • G06F3/00—Input arrangements for transferring data to be processed into a form capable of being handled by the computer; Output arrangements for transferring data from processing unit to output unit, e.g. interface arrangements
  • G06F3/06—Digital input from, or digital output to, record carriers, e.g. RAID, emulated record carriers or networked record carriers
  • G06F3/0601—Interfaces specially adapted for storage systems
  • G06F3/0668—Interfaces specially adapted for storage systems adopting a particular infrastructure
  • G06F3/0671—In-line storage system
  • G06F3/0673—Single storage device
  • G06F3/068—Hybrid storage device

Definitions

• Configurable platforms are being deployed in data centers and are promising candidate architectures for accelerating certain classes of workloads.
• these configurable platforms typically only outperform graphics processing units (GPUs) and other types of compute-dense processing units for certain types of calculations (those with irregular data flows, computations on non standard bit-width data types, etc.).
• field- programmable gate array (FPGA) vendors have started incorporating hardened logic, hardcoded logic blocks, or hardened compute blocks (collectively“hardened logic blocks”) for a number of compute element types including central processing units (CPUs), floating point (FP) units and the like in the FPGAs.
• a conventional technique incorporates multiple discrete devices of each kind in the system at a board level (e.g., a CPU and a GPU along with the FPGA). This technique increases system-level cost and complexity as multiple discrete devices need to be incorporated and coordinated.
• Another conventional technique incorporates hardened logic blocks on the FPGA device. This technique requires the manufacture of many different devices with varying mixes of hardened logic block types and still runs the risk of not being optimal for any particular workload.
• Figure 1 is an example high level block diagram of a configurable computing substrate with near-memory hardened logic in accordance with some embodiments
• Figure 2 is an example detailed block diagram of a base die integrated with a memory stack in accordance with some embodiments
• Figure 3 is another example block diagram of a configurable computing substrate with near-memory hardened logic in accordance with some embodiments
• Figure 4 is yet another example block diagram of a configurable computing substrate with near-memory hardened logic in accordance with some embodiments
• Figure 5 is an example flowchart for using a configurable computing substrate with near-memory hardened logic in accordance with some embodiments.
• Figure 6 is a block diagram of an example device in which one or more disclosed embodiments may be implemented.
• Configurable computing substrates have gained considerable attention recently for datacenter deployments and machine learning acceleration.
• Configurable computing substrates can refer to a field-programmable gate array (FPGA), complex programmable logic devices (CPLD), coarse-grain reconfigurable arrays (CGRA), gate arrays and other similar platforms or devices.
• FPGA field-programmable gate array
• CPLD complex programmable logic devices
• CGRA coarse-grain reconfigurable arrays
• gate arrays and other similar platforms or devices.
• such configurable computing substrates can be augmented with hardened logic blocks for efficient support of computations with dense arithmetic or complex control flows. That is, configurable computing substrates typically incorporate hardened logic blocks for functionality that cannot be efficiently implemented using configurable logic blocks.
• Described herein is a configurable computing system which uses near memory and in-memory hardened logic blocks to enhance the capabilities of configurable computing substrates without requiring additional discrete components to be added to the configurable computing substrates, such as an FPGA device.
• These hardened logic blocks can include but are not limited to, central processing units (CPUs), CPU cores, graphics processing units (GPUs), GPU cores, hard-coded accelerators, application specific integrated circuit (ASIC) blocks, execution engines, data-parallel floating-point-capable execution engines and other similar devices or logic and combinations thereof.
• the hardened logic blocks are incorporated into memory modules.
• the hardened logic blocks are on a base or logic die or a memory die of the memory module.
• a separate die can be added and included as part of the memory module for implementing the hardened logic blocks.
• the memory modules include an interface or communication logic to communicate between the configurable computing substrate and the memory module.
• the interface between the configurable computing substrate and the memory modules can be proprietary on a condition that the interfaces are compatible to enable mix-and-match use.
• the interface may be a standard such as Joint Electron Device Engineering Council (JEDEC) high bandwidth memory (HBM) to enable the use of commodity memory modules. Standardization of the interface would enhance flexibility by allowing use of third party memory modules.
• the memory modules can include on-die memory such as static random access memory (SRAM) or other forms of non-configurable logic to enable more efficient processing for a variety of operations.
• SRAM static random access memory
• the memory modules can include a portion of configurable computing substrate logic fabric to enable more efficient processing for a variety of operations.
• the memory modules can include SRAM and a portion of configurable computing substrate logic fabric to enable more efficient processing for a variety of operations.
• FIG. 1 shows an example high level block diagram of a configurable platform 100 in accordance with certain implementations.
• the configurable platform 100 includes a configurable computing substrate 110 connected to or in communication with (collectively “connected to”) memory module(s) 120.
• the memory module(s) 120 can be 3D-stacked memory modules which have a base or logic die 130 stacked with memory dies 140.
• the configurable computing substrate 110 writes data to the memory dies 140 via the configurable computing substrate interface 200 and the memory interface 210.
• the hardened logic block 220 reads the data from the memory dies 140 via the memory interface 210.
• the results from the hardened logic block 220 are written into the memory dies 140 via the memory interface 210.
• the configurable computing substrate 110 then reads the results from the memory dies
• the configurable computing substrate 110 and the hardened logic block 220 can use memory store and load operations to write/read the data in/out of memory and operate on it or vice versa.
• the configurable computing substrate interface 200 between the configurable computing substrate 110 and the memory module 120 may be an industry-standard memory interface (as only loads/stores are needed from outside the memory module 120) or a proprietary one.
• FIG. 3 shows an example high level block diagram of a configurable platform 300 in accordance with certain implementations.
• the configurable platform 300 includes a configurable computing substrate 310 connected to or in communication with (collectively“connected to”) a memory module 320.
• the memory module 320 can be a 3D-stacked memory module which has a base die 330 connected to a stacked memory 340.
• the base die 330 includes a configurable computing substrate interface 350 (shown as CCS Interface 350 in Figure 3) for communicating with the configurable computing substrate 310.
• the configurable computing substrate interface 350 is connected to a memory interface 360, which in turn is connected to the stacked memory 340.
• the memory interface 360 is further connected to a hardened logic block 370.
• the hardened logic block 370 can include one or more logic blocks; for ease of description, only one logic block is referred to herein.
• the configurable computing substrate interface 350 is further connected to an on-die memory 380 within the base die 330 which is included in the address space or range of the memory module 320.
• the on-die memory 380 can be SRAM.
• the on-die memory 380 provides efficient processing of data when the data is sized compatibly with the capacity of the on-die memory 380 since reads and writes to the stacked memory 340 can be avoided.
• the on-die memory 380 can be provisioned on the configurable computing substrate 310.
• the configurable computing substrate interface 350 enables addressing of the on-die memory 380 from the hardened logic block 370 in the memory module 320.
• the configurable computing substrate 310 writes data via the configurable computing substrate interface 350.
• the amount of data to be communicated with the hardened logic block determines where the data is stored. If the amount of data is compatible with the capacity of the on-die memory 380, then the configurable computing substrate 310 writes the data to the on-die memory 380 (using the address space associated with the on-die memory 380) via the configurable computing substrate interface 350. If the amount of data is not compatible with the capacity of the on-die memory 380, then the configurable computing substrate 310 writes the data to the stacked memory 340 (using the address space associated with the stacked memory 340) via the configurable computing substrate interface 350 and the memory interface 360.
• the hardened logic block 370 reads the data from either the stacked memory 340 via the memory interface 360 or from the on-die memory 380.
• the results from the hardened logic block 220 are written into the stacked memory 340 via the memory interface 360 or in the on-die memory 380, depending on the amount of data associated with the results and using the appropriate address space.
• the configurable computing substrate 310 then reads the results from the stacked memory 340 via the configurable computing substrate interface 350 and the memory interface 360 or from the on-die memory 380 via the configurable computing substrate interface 350, using the appropriate address space.
• FIG. 4 shows an example high level block diagram of a configurable platform 400 in accordance with certain implementations.
• the configurable platform 400 includes a configurable computing substrate 410 connected to or in communication with (collectively“connected to”) a memory module 420.
• the memory module 420 can be a 3D-stacked memory module which has a base die 430 connected to a stacked memory 440.
• the base die 430 includes a configurable computing substrate interface 450 (shown as CCS Interface 450 in Figure 4) for communicating with the configurable computing substrate 410.
• the configurable computing substrate interface 450 is connected to a memory interface 460, which in turn is connected to the stacked memory 440.
• the memory interface 460 is further connected to hardened logic block 470.
• the configurable computing substrate 410 communicates with the hardened logic block 470 via the configurable computing substrate interface 450.
• the nature or type of the computation determines how data is communicated from the configurable computing substrate 410 to the hardened logic block 470.
• the nature of the computation can refer to streaming computations where the configurable computing substrate fabric 480 perform some step(s) of a computation and the hardened logic block 470 performs the next step(s) (or vice versa). In this instance, there is not a lot of data that is collected between the steps that happen on each device and data simply gets passed from one device to the other device as the data becomes available. Consequently, having configurable computing substrate fabric 480 right next to the hardened logic block 470 helps with the handoff in these types of computations.
• the computation is such that the configurable computing substrate fabric 480 on the memory module 420 can be used to interface with the hardened logic block 470
• data is passed from the configurable computing substrate 410 to the hardened logic block 470 via the configurable computing substrate interface 450 and mediated by the configurable computing substrate fabric 480 on the memory module 420.
• the nature of the computation is such that the configurable computing substrate fabric 480 on the memory module 420 cannot be used to interface with the hardened logic block 470, then the data is stored in the stacked memory 440 via the configurable computing substrate interface 450 and the memory interface 460.
• the hardened logic block 470 receives the data through the configurable computing substrate fabric 480 on the memory module 420 or accesses the data from the stacked memory 440 via the memory interface 460.
• the results from the hardened logic block 470 are written into the stacked memory 440 via the memory interface 460 or communicated to the configurable computing substrate 410 via the configurable computing substrate interface 450 and mediated by the configurable computing substrate fabric 480 on the memory module 420, depending on the nature of the computation.
• the configurable computing substrate 410 then reads the results from the stacked memory 440 via the configurable computing substrate interface 450 and the memory interface 460 or operates on them directly via the configurable computing substrate interface 450 and mediated by the configurable computing substrate fabric 480 on the memory module 420, as appropriate.
• a configurable platform can include a configurable computing substrate connected to a memory module, where the memory module includes both an on-die memory as described in Figure 3 and a configurable computing substrate fabric as described in Figure 4.
• the memory module includes both an on-die memory as described in Figure 3 and a configurable computing substrate fabric as described in Figure 4.
• Such an implementation can consider both data size and the nature of the computation with additional logic to determine priority considerations between the data size and the nature of the computation determinations.
• the hardened logic block(s) can be implemented on a memory die or on a separate die connected to the memory module.
• Figure 5 is an example high level flowchart 500 for data processing using
• a configurable platform which has a configurable computing substrate in communication with hardened logic block(s) implemented on memory module(s).
• the configurable computing substrate writes data to the memory module (step 510).
• the data is stored in a memory die or memory stack.
• the data is stored in an on-die memory associated with the hardened logic bock(s) when the data is compatible with the capacity of the on-die memory.
• the data is stored in logic or storage elements within an extension of the configurable logic fabric in the memory module.
• the hardened logic block(s) accesses the data from the memory module (step 520).
• the data is read from a memory die or memory stack.
• the results are read from a memory die or memory stack. In another implementation, the results are read from an on-die memory associated with the hardened logic block(s) when the size of the data associated with the results is compatible with the capacity of the on-die memory. In another implementation, the results are accessed from logic or storage elements within an extension of the configurable logic fabric in the memory module.
• FIG. 6 is a block diagram of an example device 600 in which one or more features of the disclosure can be implemented.
• the device 600 can include, for example, a computer, a gaming device, a handheld device, a set-top box, a television, a mobile phone, or a tablet computer.
• the device 600 includes a processor 602, a memory 604, a storage 606, one or more input devices 608, and one or more output devices 610.
• the device 600 can also optionally include an input driver 612 and an output driver 614. It is understood that the device 600 can include additional components not shown in Figure 6.
• the processor 602 includes a central processing unit (CPU), a graphics processing unit (GPU), a CPU and GPU located on the same die, or one or more processor cores, wherein each processor core can be a CPU or a GPU.
• the memory 604 can be located on the same die as the processor 602, or is located separately from the processor 602.
• the memory 604 includes a volatile or non-volatile memory, for example, random access memory (RAM), dynamic RAM, or a cache.
• the storage 606 includes a fixed or removable storage, for example, a hard disk drive, a solid state drive, an optical disk, or a flash drive.
• the input devices 608 include, without limitation, a keyboard, a keypad, a touch screen, a touch pad, a detector, a microphone, an accelerometer, a gyroscope, a biometric scanner, or a network connection (e.g., a wireless local area network card for transmission and/or reception of wireless IEEE 802 signals).
• the output devices 610 include, without hmitation, a configurable computing substrate, a display, a speaker, a printer, a haptic feedback device, one or more lights, an antenna, or a network connection (e.g., a wireless local area network card for transmission and/or reception of wireless IEEE 802 signals).
• a network connection e.g., a wireless local area network card for transmission and/or reception of wireless IEEE 802 signals.
• the input driver 612 communicates with the processor 602 and the input devices 608, and permits the processor 602 to receive input from the input devices 608.
• the output driver 614 communicates with the processor 602 and the output devices 610, and permits the processor 602 to send output to the output devices 610. It is noted that the input driver 612 and the output driver 614 are optional components, and that the device 600 will operate in the same manner if the input driver 612 and the output driver 614 are not present.
• a configurable computing platform includes a configurable computing substrate and a memory module connected to the configurable computing substrate.
• the memory module including a memory die and another die.
• the another die including a hardened logic block configured to operate on data sent by the configurable computing substrate, a configurable computing substrate interface configured to communicate with the configurable computing substrate and a memory interface configured to communicate with the memory die, the configurable computing substrate interface, and the hardened logic block.
• the hardened logic block is at least one of: a central processing unit (CPU), a CPU core, a graphics processing unit (GPU), a GPU core, a hard-coded accelerator, an application specific integrated circuit block, an execution engine, and a data-parallel floating-point-capable execution engine.
• the another die further includes an on-die memory configured to communicate with the configurable computing substrate interface and the hardened logic block.
• the on-die memory is further configured to communicate with the memory interface and functions as a cache for the memory die.
• the on-die memory is included in the address range of the memory module.
• the data from the configurable computing substrate is written to the on-die memory when the size of the data is substantially matched with the capacity of the on-die memory.
• results from the hardened logic block are written to the on-die memory when a data size of the results is substantially matched with the capacity of the on-die memory.
• the die further includes a configurable computing substrate fabric configured to communicate with the configurable computing substrate interface and the hardened logic block.
• the configurable computing substrate sends the data to the hardened logic block via the configurable computing substrate fabric based on a type of computation.
• the die further includes an on-die memory configured to communicate with the configurable computing substrate interface and the hardened logic block and a configurable computing substrate fabric configured to communicate with the configurable computing substrate interface and the hardened logic block.
• the die is a logic die.
• the die is a memory die.
• a method for computing using a configurable computing platform includes connecting a configurable computing substrate with a memory module.
• the configurable computing substrate writes data to a memory die included in the memory module via a configurable computing substrate interface and a memory interface, the configurable computing substrate interface and the memory interface being provided on a die included in the memory module.
• Data from the memory die is accessed by a hardened logic block via the memory interface, the hardened logic block being provided on the die.
• results are written to the memory die by the hardened logic block via the memory interface.
• data is written to an on-die memory via the configurable computing substrate interface when the size of the data is substantially matched with a capacity of the on-die memory, the on-die memory being provided on the die.
• results are written to the on-die memory by the hardened logic block when a data size of the results is substantially matched with the capacity of the on-die memory.
• a portion of the data is written to the on-die memory when the data is being written to the memory die and subsequent reads are serviced to the portion of the data.
• the on-die memory is included in the address range of the memory module.
• the data between the configurable computing substrate and the hardened logic block is communicated via a configurable computing substrate fabric and the configurable computing substrate interface based on a type of computation, the configurable computing substrate fabric being provided on the die.
• the data is written to an on-die memory via the configurable computing substrate interface when the size of the data is substantially matched with the capacity of the on-die memory, the on-die memory being provided on the die.
• results are written to the memory die by the hardened logic block when a data size of the results is substantially matched with the capacity of the on-die memory.
• processors include, by way of example, a general purpose processor, a special purpose processor, a conventional processor, a digital signal processor (DSP), a plurality of microprocessors, one or more microprocessors in association with a DSP core, a controller, a microcontroller, Application Specific Integrated Circuits (ASICs), Field Programmable Gate Arrays (FPGAs) circuits, any other type of integrated circuit (IC), and/or a state machine.
• DSP digital signal processor
• ASICs Application Specific Integrated Circuits
• FPGAs Field Programmable Gate Arrays
• Such processors may be manufactured by configuring a manufacturing process using the results of processed hardware description language (HDL) instructions and other intermediary data including netlists (such instructions capable of being stored on a computer readable media).
• HDL hardware description language
• netlists such instructions capable of being stored on a computer readable media.
• the results of such processing may be maskworks that are then used in a semiconductor manufacturing process to manufacture a processor which implements aspects of the embodiments.

Landscapes

• Engineering & Computer Science (AREA)
• Theoretical Computer Science (AREA)
• Physics & Mathematics (AREA)
• General Engineering & Computer Science (AREA)
• General Physics & Mathematics (AREA)
• Computer Hardware Design (AREA)
• Human Computer Interaction (AREA)
• Logic Circuits (AREA)
• Memory System (AREA)
• Stored Programmes (AREA)

Priority Applications (4)

Applications Claiming Priority (2)

Publications (1)

Family

ID=67213994

Family Applications (1)

Country Status (6)

Families Citing this family (4)

Citations (7)

Family Cites Families (19)

• 2018
  • 2018-01-16 US US15/872,943 patent/US10579557B2/en active Active
  • 2018-12-21 CN CN201880086468.3A patent/CN111602124B/zh active Active
  • 2018-12-21 JP JP2020536026A patent/JP7403457B2/ja active Active
  • 2018-12-21 EP EP18900755.2A patent/EP3740876A4/en active Pending
  • 2018-12-21 KR KR1020207021744A patent/KR102789084B1/ko active Active
  • 2018-12-21 WO PCT/US2018/067067 patent/WO2019143442A1/en not_active Ceased

Patent Citations (7)

Non-Patent Citations (1)

Also Published As

Similar Documents

Legal Events