LMSYS and NVIDIA: Vera CPUs cut agentic RL episode time by 11% in Miles tests
ying11231 · x · 2026-11-14
LMSYS partnered with NVIDIA to run every sandbox in its open-source RadixArk Miles RL post-training framework on NVIDIA Vera CPUs, systematically benchmarking against AMD EPYC Turin and Intel Xeon 61 — reportedly the first public end-to-end agentic RL result with all sandboxes on Vera.
- Under load, average episode time fell by 11%, returning training samples sooner and reducing GPU idle time.
- Sandboxes are built on open-source AgentENV: lightweight Firecracker VMs (more isolated than containers) that start from snapshots in under 0.1s; inference runs on SGLang on the GPU cluster.
- The core argument: as GPUs get faster, CPU speed matters more in agentic RL, since agents run commands, compile code and execute tests across dozens of turns, with hundreds of sandboxes open per run.
- Vera's advantage grows as training scales and inference accelerates. Full data and methodology on the LMSYS blog.
More from Infra
- InferenceForge claims 43% higher LLM inference throughput than best AgentX baseline — kaiwei_chang · 2026-11-14
- Redditor benchmarks a DGX Station: GB300 throughput numbers revealed — CookEasy · 2026-11-14
- Hugging Face launches llama.cpp v1 engine to run AI models locally everywhere — LinkedInNews · 2026-11-14
- China Took 59% of 600,000+ Industrial Robots Installed in 2025; Humanoid Shipments Still Tiny — ingliguori · 2026-11-14
- SGLang joins NVIDIA AI Enterprise with NIM-certified inference microservices — ying11231 · 2026-11-14
- Databricks AI agents take 2nd in Meta kernel challenge with 2.5X speedup over SGLang — Yuchenj_UW · 2026-11-14