A November 13 LMSYS blog post from the RadixArk Miles team says its open-source RL post-training framework Miles trained GLM-4.7-Flash end to end with every agent sandbox on NVIDIA Vera CPUs. With the machines already busy, episodes took 12% longer on average when sandboxes ran on AMD EPYC Turin, and within a fixed latency bar one Vera socket carried 1.3-1.5x the load of Xeon 6 and 1.4-2.1x that of Turin.
Key Takeaways
- ✓Miles ran 68 training steps and more than 2,400 episodes on Vera in both synchronous and fully asynchronous modes with no training-code changes; 530 Terminal-Bench and SWE-bench task environments were rebuilt for arm64
- ✓Controlled run (same model, tasks, GPUs; 5 steps and 160 episodes each): with sandboxes on Turin each episode took 53 s longer, including 26 s more for agent commands (60% slower) and 11 s more for the verifier
- ✓At full load each Vera core slowed only 5%-22% on Python and compile work versus 40%-70% for Turin; Vera with 176 concurrent episodes still beat Turin running just 24
- ✓Projection: with 8x faster inference, Turin episodes would take 33% longer than Vera, and 41% longer at 32x (modeled, not measured)
- ✓All figures are self-reported from a RadixArk and NVIDIA collaboration; x86 machines used AWS defaults while Vera used NVIDIA-tuned boot parameters
Key Decision Metrics at a Glance
Turn your technical choice into a development budget
Compare 40 dev plans & simulate token costs vs $20/mo subscriptions
Project Links & Resources
Direct AccessIn-Depth Technical Analysis
Background and pain points
In agentic RL the model does more than write text: it runs commands in a sandbox, compiles code, runs tests and picks its next step from the results. A training run often keeps hundreds of sandboxes open, so the CPUs behind them affect speed and scale. RadixArk partnered with NVIDIA to put every sandbox of its open-source RL post-training framework Miles on Vera CPU servers, which it calls the first public end-to-end result of this kind.
Architecture highlights
The sandbox service is built on the open-source AgentENV: each sandbox is a lightweight Firecracker VM that starts from a snapshot in under 0.1 seconds, while SGLang serves inference on the GPU cluster. Each episode, one attempt at one task, gets a fresh sandbox, the agent issues commands over many turns, and a verifier runs the tests in the same sandbox to score it. The team added support for Vera's 64K memory pages and built Firecracker for arm64.
Benchmarks and data
With each CPU socket already running 176 background tasks, episodes took 12% longer on average with sandboxes on AMD EPYC Turin, 9%-15% longer per task, while model thinking time was the same. Under a bar of 95% of commands finishing within 1.5x their standalone time, one Vera socket carried 1.3-1.5x the load of Xeon 6 and 1.4-2.1x that of Turin. These are RadixArk's own measurements.
Developer quick start
Teams running their own RL training should evaluate sandbox CPUs separately: shorter episodes mean fresher samples in asynchronous training and less idle KV cache while waiting on commands (17.9% of KV cache in use sat idle on the Turin side versus 12.1% on Vera). Miles works with many agent frameworks and sandbox backends; try it at miles.radixark.com and see the LMSYS post for full methodology and footnotes.
Benchmark side-by-side against alternatives, or calculate monthly token cost vs subscription break-even.
Discussion & Comments
0Sign in to join the discussion
Connect with AI developers to exchange benchmark insights.