EmbeddingGemma 2: an open, lightweight multimodal embedding model
deepmind.google·6d ago
AI Engineering
Hacker News·Aug 24·5 min read
Hot Chips 2026: CUDA Targets RISC-V – By Chester Lam
TL;DR
Nvidia is expanding CUDA support to RISC-V CPUs, requiring server-grade specifications to ensure efficient GPU compute. This move enhances the compatibility of RISC-V with CUDA, particularly for machine learning applications.
✦ Why It Matters
Evaluate your RISC-V hardware against Nvidia's specifications to ensure compatibility with CUDA.
Key Takeaways
- Investigate RISC-V CPUs that comply with Nvidia's CUDA requirements.
- Consider the implications of PCIe coherency for GPU data management.
- Stay updated on SiFive's developments in RISC-V and CUDA integration.
- Evaluate the potential of RISC-V for machine learning applications in server environments.
Full Summary
CUDA, Nvidia's parallel computing platform, has primarily supported x86-64 and aarch64 architectures, but is now targeting RISC-V CPUs. To effectively run CUDA, RISC-V CPUs must meet specific requirements, including adherence to the RVA23 standard and server platform specifications that ensure reliability, availability, and serviceability (RAS).
Nvidia emphasizes the need for features like PCIe coherency to prevent data inconsistencies during GPU operations. Although RISC-V's software ecosystem is still developing, Nvidia's collaboration with SiFive aims to demonstrate a CUDA-capable RISC-V system.
However, existing RISC-V hardware may not meet these stringent requirements, limiting immediate accessibility for enthusiasts and developers. Nvidia's approach suggests a focus on high-performance server systems rather than consumer-grade hardware.