12.20230822_RISC-V-Summit-China-Keynote-2023-08-28-final.pdf

1、RISC-V for Digital TransformationWei-Han LienChief CPU Architect and Senior Fellow ArchitectureConfidentialAgenda2Digital TransformationScalable RISC-V based AI Scalable RISC-V processor familyChipletsConfidentialDigital TransformationHuman race is entering digital AI transformation AI revolution:Ma

2、chine intelligence replaces human brain Reshape business models,practices,and cultures for competitiveness Cloud computing,AI,IoT,and data analytics are revolutionizing the digital landscape Real-time data and streamlined processing enables agile decision-making and strategy adjustments Digital insi

3、ghts allow personalized experiences and tailored solutions,fostering customer loyalty4ConfidentialChronical of AI-driven Digital Transformation5Personal ConnectivityTransistor 1947Netscape Browser19933G WirelessiPhoneAlexNetChatGpt320001998Digital SoldConnectivityCompute200720122022Personal DeviceAI

4、 RevolutionConfidentialCompute everywhereDigital Transformation Compute Everywhere62x1012 parameters X2.5x1018 Byte data per day ChatGPT4=2-trillion parameter Data Generation=2.5 Quintillion Byte/per day Both still growing.How about power and cost?ConfidentialPervasive Computing for Digital Transfor

5、mation Massive data movement and compute everywhere Devices,edge,and cloud computing Improve compute/communication/energy costs Compute requirements Heterogeneity Scalable to meet wide-range PPA requirements Uniform architecture specification reduces complexityOpenOpen-source fosters innovations and

6、 specializationssource fosters innovations and specializations7RISC-VTenstorrent8ConfidentialTenstorrent Founded in 2016 to build the best ML training/inference chips$330M raised with 300 employees Two ML chips-Grayskull and Wormhole in production,working on third Building a high-performance RISC-V

7、processorOnly company in the world with highOnly company in the world with high-performance RISCperformance RISC-V and ML processorsV and ML processors9Jim KellerCEO,Digital Alpha processor,Apple A series,AMD Zen,Tesla Autonomous Driving systemConfidentialAI Chip Roadmap1ML Processor20212022Standalo

8、ne ML Computer2023Highly Configurable and Performant ML Chiplet2024Low Power,Low Cost ML ChipletGrayskull 12nm,276 TFLOP(FP8)Wormhole 12nm,328 TFLOP(FP8)200 GB/S Scale-out Ethernet Black Hole 6nm SiFive RISC-V X-280 Heterogenous computeGrendel CPU+ML chipletsQuasar ML ChipletNetworked ML ProcessorCh

9、ipletHeterogenousScalable AI11ConfidentialScalable Tensix ElementGrayskull:120 Tensix coresTensix coreEmbedded RISC-V processors1 Transmit1 Receive3 ComputeLicensable IP elements for scalable AIConfidentialWormhole Products(2nd Gen device for AI at scale)12nm AI Accelerator on PCIe Gen 413N300s/d(Ne

10、bula,single or dual chip config available)Modular device with 1.6TB onboard ethernetNatively scalable to an arbitrary number of devicesHigh performance at low costNebula ServerPre-built,high-density AI servers in 4U enclosures for rack systemsComprised of 32 x n300s devicesIncludes backplane interco

11、nnect,active cooling units and SDK12 PFLOP(BF8)at 6KWConfidentialScalable AI Architecture14AI scalability from 1 Tensix core to thousands of chipsSystem ChipletsChipletScale-out SystemIPsConfidentialScalable Software Stack15Fully automated path from all popular ML framework to optimized implementati

12、onHigh quality results with no manual effortSame compiler targets one chip or many thousands of chipsHardwareagnosticparallelsolutionHardwarespecificback endScalable RISC-V CPU16Confidential17Ascalon O-o-O Superscalar Processor Disruptive high-performance RISC-V processor for AI and server Projected

13、 Zen5 performance in 2024RVA-23 Advanced branch predictions8-wide decode 3 LD/ST with large load/store queues6 ALU/2 BR2 256-bit vector units2 FPU unitsConfidentialTenstorrent RISC-V O-o-O Processor Family18Higher PerformanceOpen&FreePerformanceDecode Width4-Wide Decode Sonic Boom with Vector6-Wide

14、Decode AlastorClient and Edge8-Wide Decode AscalonServer,Laptop,and HPC4-Wide Decode 3-Wide Decode 2-Wide Decode One Design and 5 IPs in a yearConfidentialCPU for AI Computation AI computations Data pre/post processing Adaptive computing resources for future AIs algorithms CPU/GPU uniform node abstr

15、action Tenstorrent overlay technology Same topological capability20Dataflow Graph MappingConfidentialAEGIS Chiplet System Architecture2416 CPU-cluster system Companion CPU cluster for AI Inter-cluster coherency Directory-base coherency system Large memory cache per DDR5-6400 channel 4 cc-NUMA 32-cor

16、e quadrants with hierarchical interconnection Ample coherent/non-coherent bandwidth for system scalability Fabric Chiplet FloorplanScalable Chiplet 25ConfidentialHeterogenous ML Processor26ConfidentialServer Chiplets27Tenstorrent Compute Everywhere28Wearable Computing29Wearable SoC with Ascalon-D2 1

17、0 mw100 mw power consumption in advance node ARM A72 high-performance superscalar processorAscalon D2Mobile Computing30Big core Ascalon 8-wide decodeLittle core Ascalon 4-wide decodeImplementation based on power-efficiency curvesComplementary DVFS states cover wide range performance/power operating

18、pointsBig/Little CoresConfidentialTenstorrent AI and RISC-V IP deliver the compute power that ADAS and IVI requireAutomotive companies can own their own silicon working with TenstorrentPower Consumption is critical:Tenstorrent technology scales from MW to mWChiplet approach reduces cost while accele

19、rating design and production schedules.AutomotiveConfidentialNetwork Packet Processing Scale out for large computation Smart NIC DPU Storage Server32Storage SKUCompute SKUConfidentialTenstorrent CPU/AI-based Video Server Host 2 x 32-core Aegis chiplet 2 x Video accelerator chiplet Video IP 10 x 4Kp6

20、0 transcodes Controller CPU Ascalon D2 cores,or TT Baby RISC coresRISCVideo CodecRISCVideo CodecCPUCPUCPUCPUD2DD2DD2DD2DD2DD2DDDR5ControllerD2DDDR5ControllerD2DDDR5ControllerD2DCPUCPUCPUCPUD2DD2DD2DD2DD2DD2DDDR5ControllerD2DDDR5ControllerD2DDDR5ControllerD2DScalable Architecture for Digital Transfor

21、mation35ConfidentialTenstorrent Scalable Architecture for Digital Transformation36 Digital transformation requires CPU/AI computing everywhere Key technology providers for wide spectrum of products for our strategy partners AI CPUWhiteboxIPCPU/AIChipletConfidentialCompute Everywhere 37Scalable CPU FamilyScalable ChipletD-6D-4D-3D-2D-8Scalable AITenstorrent RISC-V CPUs and ML technology are in a unique position+