BL .VL BRN R S Q E5 S Q31 9 VWS
•
0 gostou•182 visualizações
This document provides an overview of the SNAP framework, which utilizes Power CAPI technology to enable coherent acceleration between CPUs and FPGAs. Key points: - CAPI allows direct memory access between CPUs and FPGAs, avoiding overhead of device drivers and memory copies. This reduces latency significantly compared to traditional PCIe. - The SNAP framework uses CAPI to share memory coherently between applications running on CPUs and accelerators implemented on FPGAs. - It includes a kernel driver, user library, and models the hardware interface to allow co-simulation of applications and accelerators. - This framework takes advantage of features like DMA, atomic operations, and wake-ups to provide
Recomendados
Recomendados
Mais conteúdo relacionado
Mais procurados
Mais procurados (20)
Semelhante a BL .VL BRN R S Q E5 S Q31 9 VWS
Semelhante a BL .VL BRN R S Q E5 S Q31 9 VWS (20)
Mais de Yutaka Kawai
Mais de Yutaka Kawai (20)
Último
Último (20)
BL .VL BRN R S Q E5 S Q31 9 VWS
- 1. BL .VL BRN R S Q E5 S Q31 9 VWS S Q E5 6RWQ GVMRQ r 1 AWPPMV . u H YI I I IP
- 2. © 2019 IBM Corporation 2 . L U S D3 4T SJGXOTS roep U S D3 @ RROX fx • 1/ % U S1/ /II Q VGX J TWXMV @ 9 • U S @/ 3 U VOR SXW U S1/ roep • @ / roep • D3 , 4OSGQ /JJOXOTS • U S1/
- 3. © 2019 IBM Corporation 3 VMTA G7B 8U TLIYPUT
- 4. OpenPOWER Foundation Mission “Through the growing open ecosystem of the POWER Architecture and its associated technologies, the OpenPOWER Foundation facilitates its Members to share expertise, investment and intellectual property to serve the evolving needs of all end users.” Artificial Intelligence Custom Hyperscale Data Centers Hybrid Cloud Open Solutions IT consumption models are expanding Price/Performance Full system stack innovation required Moore’s Law Technology and Processors 2000 2020 Firmware / OS Accelerators Software Storage Network Full Stack Acceleration (Lower is better) IT innovation can no longer come from just the processor
- 6. Software Implementation / HPC / Research Chip / SOC I/O / Storage / Acceleration Boards / Systems System / Integration 350+ Members 35 Countries 80+ ISVs A Revolution Looks Like © 2018 OpenPOWER Foundation
- 7. OpenPOWER Ecosystem - breadth of solutions Sub USD$1,500 developer systems to TCO competitive server solutions right up to world’s fastest supercomputers. All part of the OpenPOWER Ecosystem, all run open software stacks firmware to apps.
- 8. May 20th Japan Mini Summit
- 10. OpenPOWER Announcements at North American Summit - II ● IBM releases proof of concept POWER ISA Compliant FPGA Soft Core ○ Allows anyone to experiment with the POWER ISA - researchers to hobbyists, chip manufacturers to hardware accelerator vendors ○ Micropython port also announced ○ Within just two weeks already additional FPGAs supported ○ Zephyr IoT kernel support under development ○ Linux Kernel support expected by year end Refer to : https://openpowerfoundation.org/the-next-step-in-the-openpower-foundation-journey/ POWER Instruction Set Architecture (ISA)
- 11. © 2019 IBM Corporation 11 VMTA G7B C SSPY it
- 12. © 2019 IBM Corporation 12 . L U S D3 4T SJGXOTS roep U S D3 @ RROX fx • 1/ % U S1/ /II Q VGX J TWXMV @ 9 • U S @/ 3 U VOR SXW U S1/ roep • @ / roep • D3 , 4OSGQ /JJOXOTS • U S1/
- 13. © 2019 IBM Corporation 13 1F 218 4 CC I U S D3 @ RROX t c d • B@ /VINOX IX V 3B 0 WOS WW 1NOSG 3SMOS VOSM s) g y • t 8GUGS w s U S D3 @ RROX B@ t 9OS 4T SJGXOTS • % , ( .@GS 2O MT e • 1/ % U S1/ /II Q VGX J TWXMV @ 9 − NXXUW-%%WXGXOI WIN J ITR%NTWX JFLOQ W%TU SUT VSG ,%)J% TWXMV @ 9 ( TS ( 1/ V M ( ( @ ( ( , UJL • U S @/ 3 U VOR SXW − NXXUW-%%WXGXOI WIN J ITR%NTWX JFLOQ W%TU SUT VSG ,% %0QGSINGVJF ( UJL
- 14. © 2019 IBM Corporation 14 © 2018 IBM Corporation | IBM Confidential 3 PostgreSQL Accelerated with Regular Expression Matching § Two ways for user interface: § UDF (User Defined Function) § PostgreSQL Hooks/Plugins (Standard SQL) SELECT psql_regex_capi(table, pattern, attr_id); SELECT * FROM table WHERE pkt ~ pattern; 2. 1F 2. 9I 2EHI /
- 15. © 2019 IBM Corporation 15 2. 1F 2. 9I 2EHI / © 2018 IBM Corporation | IBM Confidential 4 Overall Architecture: Multi-Threading and Multi-Engine Host Memory User Space) Buffer Cache Packet Buffer 0 Packet … Packet FPGA Result 0 Result 1 … Result N … PostgreSQL Query Results Query results to DB clients A X I I n t e r c o n n e c t Job Manager Job Queue Packet Buffer Pointer 0 … Packet Buffer Pointer N Result Buffer Packet transfer via CAPI/OpenCAPI with Virtual Address Pointers Tuples Tuples Page Thread N memcpy memcpy …… Local Configuration Bus Storage Pages Pages … Pages Pages Relations Pages Pages … Pages Pages Pages Pages … Pages Pages Tuples Tuples Page Thread 0 Pages Pages … Pages Pages Relation 0 Relation N Pages Pages … Pages Pages Pages Pages … Pages Pages Pages Pages … Pages Pages Pages Pages … Pages Pages Pages Pages … Pages Pages M RegEx 0 M RegEx 1 M … M RegEx N General Query M M Others M Packet Buffer N Packet … Packet FPGA Modules under construction FPGA Engines / User space buffers for CAPI PostgreSQL internal data structures AXI/PSL Bridge or AXI/TLX Bridge
- 16. © 2019 IBM Corporation 16 © 2018 IBM Corporation | IBM Confidential 9 Performance Evaluation: Environment Setups Host Server Romulus, 2-socket, POWER 9 22 cores, 512GB memory, normal SATA hard disk FPGA Card Xilinx VU9P CAPI CAPI 2.0 + SNAP 2.0 CAPI-Regex 8 16X1 engines, 16 packet pipelines per engine; 2 64X1 engines, 64 pipelines per engine. All running @225Mhz PostgreSQL Version 11.2, complied from source shared_buffers: the amount of memory the database server uses for shared memory buffers 4GB, 1GB max_worker_processes: the maximum number of background processes that the system can support. 176 (the max value supported by Romulus) max_parallel_workers: the maximum number of workers that the system can support for parallel queries 176 (the max value supported by Romulus) Queries CAPI-Regex in UDF mode SELECT * FROM table WHERE pkt ~ pattern Test Table Table Type Synthetic Tables Table Schema Each row contains 2 columns (ID, packet); packet is 1024-byte random string Table Size Number of rows varies between tables 2. 1F 2. 9I 2EHI /
- 17. © 2019 IBM Corporation 17 © 2018 IBM Corporation | IBM Confidential 10 Performance Comparison with CPU ü CPU version rans as it is; CAPI version runs with 8 threads on 8 16X1 regex engines with the optimal number of jobs per thread ü CAPI-regex can be ~x5 to ~x10 faster than the best PostgreSQL built-in functions (CPU multi-threads enabled) ü Max 4 threads are enabled for CPU multi-threading when table size larger than 128000 ü Buffer cache size can impact CPU version but not too much on CAPI version 512k 256k 128k 64k 32k 16k 8k 4k regex_capi BC 1GB 0.23 0.25 0.06 0.06 0.07 0.07 0.09 0.15 regex_capi BC 4GB 0.20 0.25 0.06 0.05 0.06 0.07 0.09 0.14 CPU BC 1GB 1.03 1.02 1.00 0.92 1.00 1.05 1.08 1.07 CPU BC 4GB 1.00 1.00 1.00 1.00 1.00 1.00 1.00 1.00 0.00 0.20 0.40 0.60 0.80 1.00 1.20 RelativeQueryTime Table Size (Number of 1024-Byte Lines) Query Time Comparison Between CAPI-regex and CPU regex_capi BC 1GB regex_capi BC 4GB CPU BC 1GB CPU BC 4GB 2. 1F 2. 9I 2EHI /
- 18. © 2019 IBM Corporation 18 1F . LF C IH :OIVT GXX • AOS D3 ITV • U S D3 @/ WIGQGV LO J UTOSX W HW X • DVOXX S OS C 29 ( − MNJQ U S @T VI WOR QGXOTS XTTQW − EOQOS CO GJT LTV 4 / W SXN WOW • NXXUW-%%MOXN H ITR%GSXTSHQGSINGVJ%ROIVT GXX
- 19. © 2019 IBM Corporation 19 1F . LF C IH :OIVTU XNTS • @RGQQ RH JJ J XNTS OSX VUV X V − NXXUW-%%ROIVTU XNTS TVM% • :OING Q QOSM UTVX J OX XT XN ROIVT GXX ITV − T RTJOLOIGXOTSW XT M S VOI ITJ RTWXQ UQGXLTVR WU IOLOI ITJ GUUQOIGXOTS WXGVX U ITSWTQ XI
- 20. © 2019 IBM Corporation 20 53A VMT53A
- 21. © 2019 IBM Corporation 21 . L U S D3 4T SJGXOTS roep U S D3 @ RROX fx • 1/ % U S1/ /II Q VGX J TWXMV @ 9 • U S @/ 3 U VOR SXW U S1/ roep • @ / roep • D3 , 4OSGQ /JJOXOTS • U S1/
- 22. © 2019 IBM Corporation 22 11SNAP Framework built on Power™ CAPI technology2017, IBM Corporation Memory Subsystem Virt Addr CPU Core CPU Core CPU Core CPU Core CPU Core CPU Core App External Device I/F VariablesInput Data DD Device Driver Storage Area Variables Input Data Variables Input Data Output Data Output Data • An application calls a device driver to utilize an Accelerator or any device outside the chip • The device driver performed a memory mapping operation. 3 versions of the data (not coherent). 1000s of instructions in the device driver. An application without CAPI
- 23. © 2019 IBM Corporation 23 12SNAP Framework built on Power™ CAPI technology2017, IBM Corporation Memory Subsystem Virt Addr POWER8 Core POWER8 Core POWER8 Core POWER8 Core POWER8 Core POWER8 Core App FPGA PCIE PSL Variables Input Data Output Data 1 coherent version of the data. No device driver call/instructions. • CPU unloaded since no device driver and accelerator doing the application work • The FPGA shares memory with the cores An application with CAPI
- 24. © 2019 IBM Corporation 24 13SNAP Framework built on Power™ CAPI technology2017, IBM Corporation Effect of CAPI hardware vs. PCI-E Device Driver Typical I/O Model Flow: Flow with a CAPI Model: Shared Mem. Notify Accelerator Acceleration Shared Memory Completion DD Call Copy or Pin Source Data MMIO Notify Accelerator Acceleration Poll / Interrupt Completion Copy or Unpin Result Data Ret. From DD Completion Application Dependent, but Equal to below Application Dependent, but Equal to above 300 Instructions 10,000 Instructions 3,000 Instructions 1,000 Instructions 1,000 Instructions 7.9µs 4.9µs Total ~13µs for data prep 400 Instructions 100 Instructions 0.3µs 0.06µs Total 0.36µs
- 25. © 2019 IBM Corporation 25 18SNAP Framework built on Power™ CAPI technology2017, IBM Corporation FPGA POWER8 Core Recall CAPI technology connections Proprietary hardware and designs to enable coherent acceleration Operating system enablement • little endian linux • kernel driver (cxl) • user library (libcxl) Customer application and accelerator POWER8 Processor OS App Memory (Coherent) AFU PSL PCIe CAPP cxl libcxl ▪ PSLSE models the red outlined area ▪ Re-implements libcxl api calls ▪ Models memory access ▪ Provides hardware ports to afu ▪ Enables co-simulation of AFU and App ▪ Publicly available on github
- 26. © 2019 IBM Corporation 26 19SNAP Framework built on Power™ CAPI technology2017, IBM Corporation FPGA POWER9 Core OpenCAPI technology connections Proprietary hardware and reference designs to enable coherent acceleration Operating system enablement • little endian linux • reference kernel driver (ocxl) • reference user library (libocxl) Customer application and accelerator POWER9 Processor OS App Memory (Coherent) AFU TLx DLx 25G phy 25G phy DL TL NPU (w/CAPP fcn) PSL ocxl libocxl ▪ OCSE models the red outlined area ▪ OCSE enable AFU and App co-simulation IF reference libocxl and reference TLx/DLx are used. ▪ OCSE dependencies/assumptions – Fixed reference TLx/AFU interface – Fixed reference libocxl user API ▪ Will be available to consortium members TERMS: OpenCAPI Simulation Environment (OCSE) OpenCAPI defines a Data Link Layer (DL) and Transaction Layer (TL) cited by https://www.kernel.org/doc/html/latest/userspace-api/accelerators/ocxl.html
- 27. © 2019 IBM Corporation 27 17SNAP Framework built on Power™ CAPI technology2017, IBM Corporation Feature CAPI 1.0 CAPI 2.0 OpenCAPI 3.0 OpenCAPI 4.0 Processor Generation POWER8 POWER9 POWER9 Future CAPI Logic Placement FPGA/ASIC FPGA/ASIC NA DL/TL on Host DLx/TLx on endpoint FPGA/ASIC NA DL/TL on Host DLx/TLx on endpoint FPGA/ASIC Interface Lanes per Instance Lane bit rate PCIe Gen3 x8/x16 8 Gb/s PCIe Gen4 2 x (Dual x8) 16 Gb/s Direct 25G x8 25 Gb/s Direct 25G+ x4, x8, x16, x32 25+ Gb/s Address Translation on CPU No Yes Yes Yes Native DMA from Endpoint Accelerator No Yes Yes Yes Home Agent Memory on OpenCAPI Endpoint with Load/Store Access No No Yes Yes Native Atomic Ops to Host Processor Memory from Accelerator No Yes Yes Yes Accelerator -> HW Thread Wake-up No Yes Yes Yes Low-latency small message push 128B Writes to Accelerator MMIO 4/8B only MMIO 4/8B only MMIO 4/8B only Yes Host Memory Caching Function on Accelerator Real Address Cache in PSL Real Address Cache in PSL No Effective Address Cache in Accelerator Remove PCIe layers to reduce latency significantlyComparison of IBM CAPI Implementations
- 28. © 2019 IBM Corporation 28 October 25th 2018 Power™ Coherent Acceleration Processor Interface (CAPI) 22 SNAP framework Process C Slave Context libcxl cxl SNAP library Job Queue Process B Slave Context libcxl cxl SNAP library Job Queue Process A Slave Context libcxl cxl SNAP library Job Queue Application on Host Acceleration on FPGA Software Program PSL/AXI bridge DRAM on-card Network (TBD) NVMeAXI Host DMA Control MMIO Job Manager Job Queue Quick and easy developing Use High Level Synthesis tool to convert C/C++ to RTL, or directly use RTL Programming based on SNAP library and AXI interface AXI is an industry standard for on-chip interconnection (https://www.arm.com/products/system-ip/amba-specifications) C/C++ or RTL Hardware Action HDK: CAPI PSL or BSP CAPI
- 29. © 2019 IBM Corporation 29 October 25th 2018 Power™ Coherent Acceleration Processor Interface (CAPI) 45 Scatter gather memory access Results: (Power9 – CAPI2.0 – 2.154GHz, 512MB RAM) (FPGA card: FW609 + S241: VU9P Gen3x16) (SNAP) - CAPI way saves the time for “SW gather” with relatively small penalty when K grows N=1024 blocks block size= 2kBytes Traditional way Time (µs) CAPI way Time (µs) How scattered SW gather DMA Sum Verilog HLS -RK1 309.3 183.5 492.8 171.65 173.3 -RK4 319.05 186.05 505.1 180.9 180.9 -RK16 305.1 185.7 490.8 184.6 186.95 -RK64 320.6 186.85 507.45 186.3 187.5 -RK256 318.3 185.65 503.95 218.55 215.35 -RK1024 333 189.15 522.15 236.85 224.95 -RK4096 324.4 189.35 513.75 241.15 225.55 -RK16384 307.4 185.75 493.15 240.9 224.9 0 100 200 300 400 500 600 -RK1 -RK4 -RK16 -RK64 -RK256 -RK1024 -RK4096 -RK16384 Verilog HLS Sum Time:µs More scatteredMore scattered Contiguous - Once tuned (using pragmas), HLS can compete with Verilog coding 190us to transfer 2MiB: speed = 11.04GB/s 1 2 2 R = random K is the dispersion factor of the blocks Allocate 2MB in a K * 2MB memory area → K=1 : all blocks contiguous → K=2: 2MB allocated amongst 4MB → K=4: 2MB allocated amongst 8MB
- 30. © 2019 IBM Corporation 30 c MYK%
- 31. © 2019 IBM Corporation 31 . L U S D3 4T SJGXOTS roep U S D3 @ RROX fx • 1/ % U S1/ /II Q VGX J TWXMV @ 9 • U S @/ 3 U VOR SXW U S1/ roep • @ / roep • D3 , 4OSGQ /JJOXOTS • U S1/
- 32. © 2019 IBM Corporation 32 2 EFEH 218 4 2 E HHE E E M 9 . 1 4E9 C9F © 2019 IBM Corporation 2 Proposed POWER Processor Technology and I/O Roadmap POWER8 Architecture POWER9 Architecture 2014 POWER8 12 cores 22nm New Micro- Architecture New Process Technology 2016 POWER8 w/ NVLink 12 cores 22nm Enhanced Micro- Architecture With NVLink 2017 P9 SO 12/24 cores 14nm New Micro- Architecture Direct attach memory New Process Technology 2018 P9 SU 12/24 cores 14nm Enhanced Micro- Architecture Buffered Memory POWER7 Architecture 2010 POWER7 8 cores 45nm New Micro- Architecture New Process Technology 2012 POWER7+ 8 cores 32nm Enhanced Micro- Architecture New Process Technology 2021 P10 TBA cores New Micro- Architecture New Process Technology POWER10 2020 P9 AIO 12/24 cores 14nm Enhanced Micro- Architecture New Memory Subsystem Up To 150 GB/s PCIe Gen4 x48 25 GT/s 300GB/s CAPI 2.0, OpenCAPI3.0, NVLink Sustained Memory Bandwidth Standard I/O Interconnect Advanced I/O Signaling Advanced I/O Architecture Up To 210 GB/s PCIe Gen4 x48 25 GT/s 300GB/s CAPI 2.0, OpenCAPI3.0, NVLink Up To 650 GB/s PCIe Gen4 x48 25 GT/s 300GB/s CAPI 2.0, OpenCAPI4.0, NVLink Up To 800 GB/s PCIe Gen5 32 & 50 GT/s TBA Up To 210 GB/s PCIe Gen3 N/A CAPI 1.0 Up To 210 GB/s PCIe Gen3 20 GT/s 160GB/s CAPI 1.0 , NVLink Up To 65 GB/s PCIe Gen2 N/A N/A Up To 65 GB/s PCIe Gen2 N/A N/A Statement of Direction, Subject to Change 2 Focus of today’s talk Statement of Direction, Subject to Change
- 33. © 2019 IBM Corporation 33 9 I E IE I 218 4 2 E HHE 9C M © 2019 IBM Corporation 6 Memory Signaling (8x8 OMI) Memory Signaling (8x8 OMI)PowerAXON (x48) PowerAXON (x48) PCIeGen4Signaling(x48) LocalSMPSignaling(3x30) SMPandAcceleratorInterconnect Core Core L2 Core Core L2 Core Core L2 Core Core L2 Core Core L2 Core Core L2 L3 L3 10 MB L3 Region Core Core L2 Core Core L2 Core Core L2 Core Core L2 Core Core L2 Core Core L2 L3 L3 Processor Chip Details • 728 mm2 ( 25.3 x 28.8 mm) • 8 Billion Transistors • Up to 24 SMT4 Cores • Up to 120 MB eDRAM L3 cache Semiconductor Technology • 14nm finFET • Improved device performance • Reduced energy • eDRAM • 17 layer metal stack High Bandwidth Signaling • 25 GT/s low energy differential • PowerAXON, OMI memory • 16 GT/s low energy differential • Local SMP • 16 GT/s PCIe Gen4 Open Memory Interface (OMI) • 16 channels x8 at 25 GT/s • 650 GB/s peak 1:1 r/w bandwidth • Technology Agnostic • Offered w/ Microchip DDR4 buffer (410 GB/s peak bandwidth) PowerAXON 25 GT/s Attach • Up to 16 socket glue-less SMP (4x24 SMP added to 3x30 local) • Up to x48 NVIDIA NVLINK GPU attach • Up to x48 OpenCAPI 4.0 coherent accelerator / memory attach Industry Standard I/O Attach • x48 PCIe Gen 4 at 16 GT/s • Up to x16 CAPI 2.0 coherent accelerator / storage attach Final Addition to the POWER9 Processor Family 2 TB/s Raw Signaling Bandwidth Shared by 6 Attach Protocols The Bandwidth Beast Advanced I/O (AIO)
- 34. © 2019 IBM Corporation 34 218 4 MHI CH 0 CE M I 9I M © 2019 IBM Corporation Connect all memory technologies to Power systems through OpenCAPI and OMI Why? – It is a high speed interface that allows flexibility to attach any new and emerging memory technology, including persistent memories like storage class memory (SCM) CPU P9 Switch DDR DDR NVME OpenCAPI P9 – OpenCAPI CPU P9’/P10 Switch OMI OMI NVME OpenCAPI P9’ Axone, P10 – OpenCAPI and OMI POWER Systems Memory Strategy
- 35. © 2019 IBM Corporation 35 -M: 0 CE M :HMHI C -0 © 2019 IBM Corporation • Hybrid Memory Subsystem using Low Latency NAND and DRAM – Exclusive partnership for low latency NAND media, and with Bittware for design of accelerator card 250-HMS – Low Latency NAND for capacity and persistence, with DRAM used for caching to lower average latency • Capabilities – SCM on OpenCAPI using Load/Store memory semantics – Competitive latency and bandwidth at reduced cost for systems with high capacity memory requirements • Target Applications – Primary: cost reduction on in-memory applications and databases with predominantly Sequential and mostly Read-Only processing Hybrid Memory Subsystem - HMS
- 36. © 2019 IBM Corporation 36 1F 2. HMCC I 1F 9IE II9 © 2019 IBM Corporation 9 OpenCAPI 4.0: Asymmetric Open Accelerator Attach Roadmap of Capabilities and Host Silicon Delivery Accelerator Protocol CAPI 1.0 CAPI 2.0 OpenCAPI 3.0 OpenCAPI 4.0 OpenCAPI 5.0 First Host Silicon POWER8 (GA 2014) POWER9 SO (GA 2017) POWER9 SO (GA 2017) POWER9 AIO (GA 2020) POWER10 (GA 2021) Functional Partitioning Asymmetric Asymmetric Asymmetric Asymmetric Asymmetric Host Architecture POWER POWER Any Any Any Cache Line Size Supported 128B 128B 64/128/256B 64/128/256B 64/128/256B Attach Vehicle PCIe Gen 3 Tunneled PCIe Gen 4 Tunneled 25 G (open) Native DL/TL 25 G (open) Native DL/TL 32/50 G (open) Native DL/TL Address Translation On Accelerator Host Host (secure) Host (secure) Host (secure) Native DMA to Host Mem No Yes Yes Yes Yes Atomics to Host Mem No Yes Yes Yes Yes Host Thread Wake-up No Yes Yes Yes Yes Host Memory Attach Agent No No Yes Yes Yes Low Latency Short Msg 4B/8B MMIO 4B/8B MMIO 4B/8B MMIO 128B push 128B push Posted Writes to Host Mem No No No Yes Yes Caching of Host Mem RA Cache RA Cache No VA Cache VA Cache
- 37. © 2019 IBM Corporation 37 :9 A F
- 38. © 2019 IBM Corporation 38 4 AU MW C YMS
- 39. Intel CPU 92 u D:MQN u7 Cs9QV O a au7 C IBM POWER CPU 150GB/s 150GB/s 150GB/s GPU 150GB/s 32GB/s CPU-GPU 4.7 39 92 RX T A UV P 13. X OO • M MRRo a p F 6 3 9AE DM RI F( u • 350))p F=PT )% nta5AE 9AE a TYMR c cp) , GPU GPU GPU *0
- 40. a adAWPPMVb ce 13. B h zc p c c c cfC SSPYd e a ) (/ - p c c c c D A, mh u krklb 1. 200 u a ar19 t ozp 8 3 19upxt ly F=PT )% m5AE 9AE dA5 M 9MT* 0%, e /- ) p c - ns u a /- 5A 6m )%0 p yc u b /- 5OIPTMW *%. a5INNM *%/ p yc m u b PTMYPKIsA G7B/ / b * https://www-03.ibm.com/press/jp/ja/pressrelease/53461.wss POWER9 14nm a DOMQN 3 C 7 C
- 41. RX T A UV P 13. ( l / -(( 7B8b c 7BF c l 1 )E x l 5AE1 A G7B0 C D2( ) *) rl ) 7B8 IRT , 78 BWTHR ( ( 78 7B8 IRT ) 78 BWTHR ( 78 7BF -IRT ( 78 BWTHR ( ) 78 7BF IRT - 78 BWTHR ( 78 l 1 -72 ),-94 ,()94 ( ) 94 )-72 66B )---9: (- l 1 * 94 c ” b(. 94 ) e l A5 M (- 9MT =U AWUNPRM1 ) 53A )% e / 9MT =U AWUNPRM1 ( 53A )% e 9MT =U AWUNPRM1 ( l C88 )%, C3D3 w ) :661 D4 rl CC61 .%-/D4 l 7 C/ D9491 B UOG D 7B8/ ) 7BF/ ) i ( 72 (-94 ” l 1 ) ) F l C1 85:, :5 CHWQVW - )
- 42. © 2019 IBM Corporation 42 6 WISTRON “MiHawk” 24 x NVMe = 96 lanes Gen3 PCIe = 48 lanes Gen4 PCIe = 32 lanes OpenCAPI 3.0 Image Source: Wistron
- 43. © 2019 IBM Corporation 43 1F 2. E 0 9 A 7 WISTRON “MiHawk” 24 x NVMe = 96 lanes Gen3 PCIe = 48 lanes Gen4 PCIe = 32 lanes OpenCAPI 3.0 Image Source: Wistron OpenCAPI !
- 44. © 2019 IBM Corporation 44