FOSSWire
Search/
Open Hardware & RISC-V

Alibaba's XuanTie Team Achieves On-Chip AI Milestone with RISC-V-Based C950 Processor

Brief written by the FOSSWire AI editor from RISC-V International How we write

1 min read 32 reads

Key Takeaway

Alibaba's XuanTie team has achieved an on-chip AI milestone by running a 27-billion-parameter large language model at over 30 tokens per second on a RISC-V-based C950 processor without a separate GPU or NPU.

Summary

The XuanTie team at Alibaba DAMO Academy has achieved a significant milestone in on-chip AI processing with their RISC-V-based C950 processor. The team was able to run a 27-billion-parameter large language model at over 30 tokens per second without the need for a separate GPU or NPU. This achievement demonstrates the potential of RISC-V-based processors for high-performance AI workloads. The C950 processor is designed to meet the RVA23 profile and features a custom matrix extension to accelerate matrix math operations.

Achieving High-Performance AI with RISC-V

  • The C950 processor features a custom matrix extension to accelerate matrix math operations.
  • The processor is designed to meet the RVA23 profile, which mandates the RISC-V Vector extension (RVV 1.0).
  • The team achieved an on-chip AI milestone by running a 27-billion-parameter large language model at over 30 tokens per second.

Why it matters

This achievement demonstrates the potential of RISC-V-based processors for high-performance AI workloads and could lead to increased adoption of RISC-V in the AI industry.

Original reference: riscv.org →