Digest · past 24h

1. NVIDIA Groq 3: SRAM, Disaggregation and Determinism in the New AI Platform
   NVIDIA's $20 billion acquisition of Groq has resulted in the integration of LPU accelerators into the Vera Rubin platform, enabling disaggregated, heterogeneous AI inference. The new Groq 3 LPX rack system combines 256 LPUs for fast, deterministic token generation, while Vera Rubin NVL72 handles flexible training and inference. The platform targets low-latency interactive AI and multi-agent workloads, promising up to 35x faster performance than Grace Blackwell NVL72 at 400 TPS per user.