JP5575947B2 - Multiprocessor device - Google Patents

Description

  The present invention relates to an optimal bus configuration and layout configuration of a multiprocessor device in which a plurality of multiprocessor groups are implemented in the same LSI.

  The bus configuration of a multiprocessor device that implements a plurality of multiprocessors of the same architecture and different architectures such as CPU and DSP on the same semiconductor chip is as one bus as described in Non-Patent Document 1 below. A configuration in which all multiprocessors are connected, and a multi-processor having the same protocol as described in Non-Patent Document 2 below has a local bus for each CPU to connect the bus, and multiple local processors The buses were bridged together.

  When all the multiprocessors are connected to one bus, the external bus I / F is connected to the same bus regardless of whether there is one LSI or a plurality of external bus I / Fs.

  When the buses are divided into multiple buses and bridged together, only one processor is connected to the local bus, and each local bus is connected to one bus master and connected to the external bus I / F. One bus is connected.

Toshiba, EmotionEngine, SCE / IBM / Toshiba, Cell, February 9, 2005, [Search January 9, 2007], Internet <http://ascii24.com/news/i/tech/article/2005/02 /09/654178-000.html> Renesas, G1, February 2006, ISSCC2006 Fig.29.5.1 `` A Power Manegement Scheme Controlling 20 Power Domains for a Single-Chip Mobile Processor ''

  However, when multiple multiprocessors are connected to a single bus and include different architectures, there is usually a difference in processing performance due to the processor, so the operation of the high-speed processor is hindered by the low-speed processor, and the performance of the high-speed processor There was a problem that did not go up. If a processor that mainly processes data such as DSP and SIMD type massively parallel processors is included, the DSP and SIMD type massively parallel processors handle a large amount of data. There was a problem that the benefits of improved performance by the processor could not be received.

  In addition, the problem of cache coherency is guaranteed in the case of multiprocessors of the same architecture, but is not guaranteed in the case of different architectures, and there is a problem that consistency is not achieved. .

  Also, when running a multiprocessor-compatible OS, developers of different architectures have different developers, so there is almost no need to create an OS that supports multiple processors, and multiprocessor-compatible OSs only support processors with the same architecture. Mostly not. Therefore, a processor of a different architecture will have a different OS, but if different OSs are on the same bus, it will be the same as a bus master IP connection that the OS is not familiar with on the same bus, There was a problem that performance improvement such as scheduling by multi-processor compatible OS was hindered.

  Even if the buses are divided into multiple buses and bridged together, each local bus has a single bus master, so the CPU and local buses are considered to be a single CPU. There was the same problem as the above problem when the architecture was on the same bus.

  In addition, even if there is one or more external bus I / F in the LSI, the bus on the side to which the external bus I / F is connected is connected on the same bus, The external bus access request from another bus is frequently stopped and the desired performance cannot be obtained. In addition, there is a problem that the performance of the bus on the side to which the external bus I / F is not connected decreases when accessing the external bus I / F from another bus.

  Accordingly, the present invention has been made to solve such a problem, and an object thereof is to obtain a high-performance multiprocessor device by having an independent bus and an external bus I / F for each different architecture.

  In one embodiment of the present invention, a multiprocessor device includes a plurality of first processors, a plurality of second processors, a first bus to which the plurality of first processors are connected, and the plurality of the plurality of first processors. A second bus to which a second processor is connected, a first external bus I / F to which the first bus is connected, and a second external bus to which the second bus is connected I / F is provided on one semiconductor chip. The first processor and the second processor are controlled by different clocks and have different frequencies or phases.

  According to an embodiment of the present invention, when a plurality of multiprocessor groups are implemented on the same semiconductor chip, an independent bus and an external bus I / F are provided for each different architecture. With this configuration, since each multiprocessor group can operate almost independently, arbitration between processors of different architectures and bus contention are reduced, and a high-performance multiprocessor system can be realized at low cost and low power.

1 is a configuration diagram illustrating a multiprocessor device according to a first embodiment of the present invention. FIG. 6 is a layout diagram in the second embodiment of the present invention. FIG. 6 is a layout diagram in the second embodiment of the present invention. FIG. 6 is a layout diagram in the second embodiment of the present invention. FIG. 10 is a layout diagram in the third embodiment of the present invention. FIG. 6 is a configuration diagram showing a multiprocessor device in a fourth embodiment of the present invention. FIG. 10 is a configuration diagram showing a multiprocessor device in a fifth embodiment of the present invention. 10 is a timing chart according to the sixth embodiment of the present invention. It is a figure which shows the clock supply circuit in a prior art. FIG. 10 is a diagram illustrating a clock supply circuit according to a sixth embodiment of the present invention. FIG. 17 is a software block diagram according to Embodiment 7 of the present invention. FIG. 17 is a software block diagram according to Embodiment 7 of the present invention.

[Embodiment 1]
FIG. 1 is a configuration diagram showing a multiprocessor device according to Embodiment 1 of the present invention, and this multiprocessor device is formed on one semiconductor chip. A plurality of processors 1 to 8 are configured in parallel (first processor group) and have a SMP (Symmetric Multiple Processor) configuration. Each CPU has a primary cache (I-cache, D-cache), internal memory (U-LM), MMU (memory management), and SDI (debugger). The eight CPUs are connected to the CPU bus 10 (first bus), and the CPU bus 10 is connected to the secondary cache 12 via the CPU bus control unit 11. The secondary cache 12 is connected to the external bus 1 via a DDR2 I / F 13 (first external bus I / F).

  The CPU operates at a maximum of 533MHz. The CPU frequency is converted by a bus I / F inside the CPU and is connected to the CPU bus 10 at a maximum of 266 MHz. The secondary cache 12 and the DDR2 I / F 13 operate at a maximum of 266 MHz.

  The LSI of the present invention has an internal peripheral bus 14 (second bus) in addition to the CPU bus 10 on the same semiconductor chip. Internal peripheral bus 14 includes ICU (interrupt controller), ITIM (periodic timer), UART (Universal Asynchronous Receiver Transmitter: clock asynchronous serial I / O), CSIO (clock synchronous serial I / O), CLKC (clock Peripheral circuit 15 such as a controller), DMAC 16 (DMA controller), built-in SRAM 17, SMP configuration matrix type massively parallel processor (SIMD type massively parallel processors 31, 32, second processor group), external bus control unit 18 (second External bus I / F), CPU19 of different architecture is connected. The internal peripheral bus 14 is connected to the external bus 2 via the external bus control unit 18, and constitutes an external bus access path for connecting to external devices such as SDRAM, ROM, RAM, and IO.

  The internal peripheral bus 14 operates at a maximum of 133 MHz, and the DMAC 16, the built-in SRAM 17, and the peripheral circuit 15 also operate at a maximum of 133 MHz. The inside of the SIMD type massively parallel processor operates at a maximum of 266 MHz, is frequency-converted by the bus I / F inside the SIMD type massively parallel processor, and is connected to the internal peripheral bus 14. The CPU 19 also operates at a maximum of 266 MHz, is frequency-converted by the bus I / F inside the CPU 19 and is connected to the internal frequency bus 7. Since there is a speed difference in processing performance in this way, each processor group is controlled by a clock of a different system, and the frequency, phase, etc. are different.

  The CPU bus 10 and the internal peripheral bus 14 are connected through the secondary cache 12. Therefore, the CPUs 1 to 8 can access not only the external bus 1 from the DDR2 I / F 13 through the secondary cache 12 but also the resources of the internal peripheral bus 14 through the secondary cache 12. Therefore, since the path is long and the frequency is slow, the data transfer performance is not improved, but the CPUs 1 to 8 can access another external bus 2 through the external bus control unit 18. Each module connected to the internal peripheral bus 14 can access the external bus 2 through the external bus control unit 18, but cannot access the external bus 1.

  CPUs 1 to 8 are CPUs of the same architecture. As for the coherency of the primary / secondary cache, the contents of the primary / secondary cache memory are coherently controlled and consistent, so there is no risk of the CPU malfunctioning. Even when using a multiprocessor-compatible OS, there are only 8 CPUs of the same architecture and the secondary cache 12 on the CPU bus 10, and access to the external bus 1 is limited to access from CPUs 1 to 8, High performance can be achieved. In particular, SIMD massively parallel processors are slower than CPUs and handle large amounts of data during data processing, so they tend to occupy the bus for a long time. Is not affected on the CPU bus 10 side.

  Also, from the viewpoint of the SIMD massively parallel processor, the CPU mainly uses the path from the CPU bus 10 to the external bus 1, so there is no need to open the internal peripheral bus 14 for the CPU during data transfer, and efficiency Data transfer is possible. The effect is remarkable because it is a multiprocessor composed of multiple CPUs. In the example of the present invention, there are 8 CPUs, but 16 or 32 or more processors are for data processing such as SIMD type massively parallel processors. If the processor is on the same bus, the data processing will be delayed, so the effect of the present invention will become more prominent.

  The CPU 19 is a microprocessor that consumes less power and has a smaller area, although the operation speed and processing performance are lower than those of the CPUs 1-8. Processing that does not require arithmetic processing performance, such as starting the peripheral circuit 15, checking the timer, and power management using the CLKC, can be performed. Therefore, even if it is configured on the same bus as the SIMD type massively parallel processor, there is no problem that the performance of the SIMD type massively parallel processor decreases.

[Embodiment 2]
2 to 4 are diagrams showing layouts of the multiprocessor device according to the second embodiment of the present invention. FIG. 2 is a layout example in which each module of the first embodiment is configured on an actual silicon wafer. Fig. 3 shows the layout example of Fig. 2. CPU bus related modules (CPU1-8, CPU bus control) and internal peripheral bus related modules (SIMD type massively parallel processors 31, 32, CPU19, built-in SRAM 17, peripheral circuit 15, external bus control FIG. 7 is a diagram in which the CPU 18 and DMAC 16) are grouped into a CPU bus area 20 and an internal peripheral bus area 21, respectively. FIG. 4 is an image diagram of power supply / GND wiring 22 in the second embodiment.

  With the layout configuration shown in Fig. 2, the internal peripheral bus 14 and the CPU bus 10 can be connected in the shortest as shown in the figure. Smaller and lower cost. In addition, the number of wiring lines that intersect with signal lines other than the bus is reduced, and it is difficult for speed reduction due to wiring congestion and long-distance wiring to occur, so that a low power consumption and low cost LSI can be realized. In addition, since the area is divided for each bus area, it is easy to perform control such as power-off.

  Further, the internal peripheral bus area 21 and the CPU bus area 20 have a difference in power consumption because there is a difference in operating frequency / arithmetic processing capacity. The CPU bus area 20 having a high clock frequency and high power consumption requires low impedance wiring, and the internal peripheral bus area 21 having a low clock frequency and low power consumption may have a relatively high impedance. Low-impedance wiring in areas with high power consumption can be realized by increasing the wiring width or by reducing the wiring interval, but the area occupied by the power supply / GND wiring 22 in the wiring layer increases as a compensation. For this reason, it becomes difficult to wire other signal lines and the like, and as a result, the power consumption increases due to an increase in LSI area, an increase in cost, and an increase in wiring capacity due to bypass of signal wiring. When the respective regions are mixed, it is necessary to make the whole wiring with low impedance in order to guarantee stable operation, but the area becomes large and the cost becomes high.

  As shown in FIG. 3, when the CPU bus area 20 with high power consumption and the internal peripheral bus area 21 with low power consumption are divided, the low impedance power / GND wiring 22 need only be applied to the CPU bus area 20. For example, as shown in FIG. 4, the CPU bus area 20 may have thick wiring densely and the internal peripheral bus area 21 may have thin wiring sparsely wired. By doing so, it is possible to guarantee a stable operation while eliminating the need for unnecessary power supply wiring and reducing the cost. The same applies to the power supply terminals. As shown in FIG. 4, the power supply / GND terminal 23 of the CPU bus area 20 may be dense and the power supply / GND terminal 23 of the internal peripheral bus area 21 may be sparse.

  The line drawn on the area in FIG. 4 is a power supply or GND line, and the circle on the outer edge of the chip is a power supply or GND terminal. Although several pseudo-wide wirings are drawn, many thinner wirings are actually drawn. For example, in a manufacturing process in which the minimum width of signal wiring is 0.2 μm, wiring of 1 μm width is routed at 4 μm pitch in the CPU bus area 20, and wiring of 0.4 μm width is wired at 100 μm pitch in the internal peripheral bus area 21. . By doing so, it is possible to ensure a stable operation while eliminating the unnecessary power supply / GND terminal 23. Since the external bus is not connected to the CPU bus area 20 from FIG. 1, the number of terminals is small, and if the arrangement of the present embodiment is applied, this can be realized without much influence.

  In the present embodiment, as shown in FIG. 2, the external bus 1 and the external bus 2 are arranged apart from each other above and below the chip. Since the external bus control unit 18 or the DDR2 I / F 13 has a high driving capability, it consumes a large amount of power and is likely to cause power supply noise. However, in the arrangement according to the present embodiment, the external bus control unit 18, the DDR2 I / F 13, and the CPU, which are large current sources, are arranged apart from each other and local power concentration does not occur, so that heat generation is made uniform. The external bus control unit 18, the DDR2 I / F, and the CPU are sensitive to noise and temperature changes. However, the arrangement of the external bus control unit 18, DDR2 I / F, and CPU reduces the influence of mutual noise and heat generation.

  By disposing modules that are large in power consumption and sensitive to noise as described above, the influence of each other's noise is reduced, so that it is possible to design with a small margin for noise. Further, since the overall power consumption is made uniform and local power concentration does not occur, the power supply wiring can be simplified, and there can be no local heat generation, and the design can be performed with a small margin for temperature change. From the above, it is possible to realize a small area, low cost, low power LSI while guaranteeing stable operation.

[Embodiment 3]
FIG. 5 is a layout example of each module configured on the actual silicon wafer of the first embodiment. Compared to the second embodiment, the positional relationship between the CPU bus control module and peripheral modules, the position and size of the built-in SRAM 17, and the shapes of the CPU 19 and the secondary cache 12 are changed.

  With regard to the positional relationship between the CPU bus control module and peripheral modules, the layout using the automatic wiring tool is not a straight wiring with strictly divided areas as shown in Fig. 2, but instead of a straight line with divided areas as shown in Fig. 5. Buses are often wired. In that case, there is some overlap with the internal peripheral bus 14, but almost the same effect as in the second embodiment can be obtained. Further, as shown in FIG. 5, it may be better to arrange the CPU bus control module near the center of gravity of each CPU and the secondary cache 12, for example, the built-in SRAM 17 may be smaller than the second embodiment, and When the access frequency is low and the operation speed is sufficient, the total area can be reduced and the cost can be reduced by changing the arrangement position flexibly as shown in FIG.

  For bus wiring to the built-in SRAM 17, a buffer circuit 24 is placed where the internal peripheral bus 14 branches off as shown in FIG. By doing so, it is possible to prevent the deterioration of the speed of the internal peripheral bus 14 and the increase in power due to the increase in the wiring length of the internal peripheral bus 14. Since access to the built-in SRAM 17 has a sufficient speed, the insertion of the buffer circuit 24 is not a problem.

[Embodiment 4]
FIG. 6 is a block diagram showing a multiprocessor device according to the fourth embodiment of the present invention. The differences from the first embodiment will be described below. The CPU bus 10 and the internal peripheral bus 14 are connected through a bus bridge circuit (Bus bridge 25). Therefore, the CPUs 1 to 8 can access not only the external bus 1 from the DDR2 I / F 13 through the secondary cache 12 but also the resources of the internal peripheral bus 14 through the bus bridge 25. Accordingly, since the path is long and the frequency is slow, the data transfer performance is not improved, but another external bus 2 can be accessed through the external bus control unit 18. However, the access to the external bus 2 and the internal peripheral bus 14 through the bus bridge 25 is not subject to caching of the secondary cache 12. Each module connected to the internal peripheral bus 14 can also access the external bus 2 through the external bus control unit 18 and can also access the external bus 1 through the bus bridge.

  CPUs 1 to 8 are CPUs of the same architecture. As for the coherency of the primary / secondary cache, the contents of the primary / secondary cache memory are coherently controlled and consistent, so there is no risk of the CPU malfunctioning. Even when a multiprocessor OS is used, there are only 8 CPUs with the same architecture, secondary cache 12 and bus bridge 25 on the CPU bus 10, and access to the external bus 1 is connected to the internal peripheral bus 14. Since each module has few accesses and is almost from CPU 1-8, high performance can be achieved.

  Since other configurations and effects are the same as those of the first embodiment, description thereof is omitted.

[Embodiment 5]
FIG. 7 is a configuration diagram showing a multiprocessor device according to the fifth embodiment of the present invention. A difference from the first embodiment is that DSPs 41 and 42 are connected in place of the SIMD type massively parallel processors 31 and 32. Further, in the present embodiment, the secondary cache 12 is used for the bridge between the CPU bus 10 and the internal peripheral bus 14, but a dedicated bus bridge 25 may be used as in the fourth embodiment. Since other configurations and effects are the same as those of the first embodiment, description thereof is omitted.

[Embodiment 6]
FIG. 8 is a timing chart showing the relationship between the CPU clock (CPU clock) and the CPU bus clock (bus clock) in the first to fifth embodiments. The frequency of the CPU clock and the CPU bus clock is considered when the CPU clock is faster than the bus clock. FIG. 8 shows an example in which the frequency ratio between the CPU clock and the bus clock is 1: 1, 2: 1, 4: 1, and 8: 1. The n-divided clock (n = 1, 2, 4, 8) is a clock obtained by dividing the CPU clock according to the frequency ratio.

  In the present invention, the bus clock of FIG. 8 is used as the clock of the CPU bus 10 (see FIG. 1) instead of the n-divided clock. FIG. 9 shows a clock supply circuit when the n-divided clock is used, and FIG. 10 shows a clock supply circuit when the Sync. + bus clock is used. Both the frequency divider of FIG. 9 and the Sync. Generation of FIG. 10 are generated by CLKC. Since there is usually one CLKC in the LSI, depending on the CPU, the n-divided clock or Sync. Will be connected over a long distance, and in fact a buffer will be inserted.

  When comparing the n-divided clock and Sync., the switching frequency (frequency) is the same, but the n-divided clock needs to be exactly in phase with the CPU clock, but Sync. is not necessary Therefore, an unnecessarily large buffer and useless delay generation buffer are not required, and an LSI with a small area, low cost and low power consumption can be realized.

  As for the quality of the clock of the CPU bus 10, in FIG. 9, the bus clock in FIG. 10 is the CPU clock compared to the n-divided clock that is far from the CPU clock and the frequency divider is inserted. The branch point is close and only the AND circuit is inserted. Therefore, the phase difference (skew) from the CPU clock can be made smaller in FIG. 10, and the operation can be performed at a higher frequency. Also, a buffer for guaranteeing the hold is not necessary or required for the transfer between the CPU and the CPU bus 10. Therefore, a small area, low cost and low power consumption LSI can be realized.

  Although the relationship between the CPU and the CPU bus clock has been described in the present embodiment, the same applies to the SIMD type massively parallel processor and the internal peripheral bus 14, and the CPU 19 and the internal peripheral bus 14.

[Embodiment 7]
FIG. 11 is a block diagram of the software of the system using the first to sixth embodiments. Each processor has a device driver (driver) and an OS in the upper hierarchy. The CPUs 1 to 8 are controlled by the OS1, and the SIMD type massively parallel processors 31 and 32 and the CPU 19 are controlled by the OS2. As each OS, for example, OS1 is a non-real-time OS such as Linux (registered trademark), and OS2 is a real-time OS such as ITRON. OS1 is optimized for the CPU architecture, and on the CPU bus 10 there are only 8 CPUs of the same architecture, the secondary cache 12 and the Bus bridge 25. In addition, access to the external bus 1 is also rare because each module connected to the internal peripheral bus 14 is accessed from the CPUs 1 to 8, and high performance can be obtained. In addition, the contents of the primary / secondary cache memory are coherently controlled by OS1 to ensure consistency, and coherency issues can be optimally addressed. On the other hand, the OS2 side also has an external bus 2 independent of OS1, so there is almost no resource adjustment with OS1 and high performance can be achieved.

  FIG. 12 is obtained by additionally providing OS3 for the CPU 19 from FIG. In addition to the effects shown in FIG. 11, each OS handles only a single architecture processor, so it is more efficient.

  1 to 8, 19 CPU, 10 CPU bus, 11 CPU bus controller, 12 secondary cache, 13 DDR2 I / F, 14 internal peripheral bus, 15 peripheral circuit, 16 DMAC, 17 internal SRAM, 18 external bus controller, 20 CPU bus area, 21 internal peripheral bus area, 22 power supply / GND wiring, 23 power supply / GND terminal, 24 buffer, 25 Bus bridge, 31, 32 SIMD type massively parallel processor, 41, 42 DSP.

Claims (9)

A first multiple processors,
A second multiple processors,
A first bus to which the plurality of first processors are connected;
A second bus to which the plurality of second processors are connected;
The first and the external bus I / F first external bus connectable,
A second external bus I / F second external bus can be connected, a provided on one semiconductor chip,
Said first processor and said second processor is controlled by a separate system clock, Ri frequency or phase Do different,
The plurality of first processors includes the first external bus provided outside the semiconductor chip via the first bus and the first external bus I / F without going through the second bus. Is accessible to
The plurality of first processors can access the second external bus via the first bus, the second bus, and a second external bus I / F,
The plurality of second processors include the second external bus provided outside the semiconductor chip via the second bus and the second external bus I / F without going through the first bus. Is accessible to the
A multiprocessor device characterized by that.   2. The wiring density of power supply wiring in a processor area having a high clock frequency is increased and the wiring density of power supply wiring in a processor area having a low clock frequency is reduced among the first processor and the second processor. The multiprocessor device described. 3. The power supply terminal in the processor group area having a fast clock frequency among the first processor and the second processor is increased, and the power supply terminal in the processor group area having a slow clock frequency is reduced. multiprocessor system according to one paragraph or. One of the first processor or the second processor, CPU area and the external bus I / F fast processor of the clock frequency according to any one of claims 1 to be located remotely on a semiconductor chip 3 Multiprocessor device.   5. The multiprocessor according to claim 4, wherein a fast clock is supplied to a CPU of the processor having a fast clock frequency, and a part of the clock is generated by gating the bus clock for data processing with the external bus I / F. apparatus. A secondary cache coupled between the first bus and the second bus;
2. The multiprocessor device according to claim 1, wherein the plurality of first processors are accessible to the second external bus via the secondary cache. A bus bridge coupled between the first bus and the second bus;
2. The multiprocessor device according to claim 1, wherein the plurality of first processors can access the second external bus via the bus bridge. A secondary cache coupled to the first bus and the first external bus I / F;
8. The multiprocessor device according to claim 7, wherein the plurality of first processors can access the first external bus via the secondary cache. 9. The multiprocessor device according to claim 1, wherein the plurality of first processors have an SMP configuration, and the plurality of second processors also have an SMP configuration.

Images