在威龙V2 Thunderhill FPGA平台上运行数据中心和AI推理应用程序.pdf

1、Real Systems.Real Traction.Bringing High-Performance RISC-V Platforms to LifeBalaji Baktha,CEO,Ventana Micro SystemsRISC-V Europe Summit,Paris 2025Delivers a Full-Scale Compute PlatformSYSTEM IPFEATURE PARITY WITH LEADING CPU ARCHITECTURESHIGH PERFORMANCE CPUHIGH PERFORMANCE CACHE ARCHITECTURELOW LA

2、TENCYCOHERENT NOCR ASSECURITYCHIPLET TECHNOLOGYDOMAIN SPECIFIC ACCELER ATIONGENER ATIVE AISOFTWARE STACK 2Frictionless-Making it easy for customers to adopt RISC-V in their data center applications AI Acceleration is Fragmented RISC-V Unifies ItDifferent Workloads need different compute.But they all

3、 need a unified foundation.Fragmented Architectures Reinvented Software Stacks No Shared Investment No CUDA Alternative Share ISA across CPUs and accelerators Scalable to fit across any AI Tier Common toolchain and compiler infrastructure Ecosystem leverage as industry wide scaleThey All Struggle Be

4、causeRISC-V Enables Ecosystem ScaleTrainingSignal Processing/Classical MLIntegrated MatrixAttached MatrixVector/Dot ProductAI Workloads Vary Compute Must ScaleBut They All need a Unified FoundationFrom Grace Blackwell to Open Compute:Ventana Enables the RISC-V AI SuperchipRISC-V Equivalents to Nvidi

5、a Grace BlackwellCPU+AI Equivalent Using Ventana RISC-VPCle Gen5Blackwell GPU“Grace”Arm CPULPDDR5X16x PCle Gen5HBM3 DRAMNVLink/NVLinkNetwork132 Steaming Multiprocessors5th Generation Tensor CoresDDR5UCle D2D InterfacesRV VectorRV MatrixRISC-V Based Veyron E2“RUCA”ChipletsRISC-V Veyron V2 Server CHI-

6、based Coherent On-chip InterconnectPCle Gen5/6IO Hub Base Building Block4Turning Vision into Action:Real-World Execution from Compute to AICloud Compute Server PartnersAI Acceleration Partners5Veyron E2-based“RUCA”AI/HPC AccelerationRISC-V MatrixRISC-V ScalarRISC-V VectorRISC-V Veyron V2 32-core Com