Real World Results using Thousands of RISC-V Cores for AI and Beyond
summary
A 7 nm SoC integrates 1,088 tiny 64-bit vector/tensor minion cores plus 4 out‑of‑order control cores and one service processor for 1093 RISC-V processors on one chip, 160 MB on-die SRAM and up to 32 GB LPDDR4X off-chip, built into PCIe cards and racks. Measured boards run typical ML workloads at about ~20 watts per chip and cards under ~40 W, yielding ~158 inferences/sec/W and enabling ~350,000 cores per rack for datacenter-scale inference.
AI-written summary · about
chapters (9)
- 0:00Esperanto RISC-V chip
- 2:37Chip architecture
- 5:29Tiled processor array
- 7:24System-on-chip features
- 8:42Data center deployment
- 10:52Software and power
- 13:10Performance benchmarks
- 17:46Summary of capabilities
- 18:50Q&A session
related talks
13mKeynote: Taking RISC-V Intelligence Everywhere
19mDemocratizing Inference of Open-weight Models on RISC-V Manycore Acc...
40mAdvancing AI Application Development on RISC-V Platforms
22mKeynote: RISC-V at NVIDIA: One Architecture, Dozens of Applications, Billions of Processors
24mIndustrial Strength High Performance RISC V Processors For Energy Efficient Computing
19mAccelerating Edge AI with the Help of RISC-V
17mRISC-V is Ready for Powering the Era of Intelligent General Computing
40mKeynote Panel: The Future of High Perfor...