RadixArk moves Miles agentic RL sandboxes onto NVIDIA Vera CPUs, cutting episode time 12% under full load
Synopsis
RadixArk, working with NVIDIA, ran every sandbox of its open-source RL post-training framework Miles on NVIDIA Vera CPUs, which it calls the first public result of end-to-end agentic RL training with all sandboxes on Vera CPU servers. In a comparison that first kept the machines busy, with the same model, tasks, GPUs and settings, episodes with sandboxes on Turin took 12% longer on average, while model thinking time was identical, so nearly all of the gap came from work inside the sandbox. With a single sandbox running, Vera performs on par with or better than current x86, and its lead grows with load.
Interpretation
In agentic RL the model runs commands in a sandbox, compiles code and runs tests, then decides its next step from the results, and a training run can keep hundreds of sandboxes open at once, so the CPUs hosting them affect training speed and scale. Terminal-Bench and SWE-bench publish only x86 images, so the team rebuilt the environments of 530 of their tasks for arm64 and checked each against its reference solution. Miles works with many agent frameworks and sandbox backends, so moving every sandbox to a different CPU architecture took no change to the training code; all the work was on the environment side.
The sandbox service is built on open-source AgentENV, where each sandbox is a lightweight Firecracker virtual machine, more isolated than a container, and starts from a snapshot in under 0.1 seconds. During training, inference is served by SGLang and training stays on the GPU cluster, and each episode gets a fresh sandbox on Vera in which an agent coordinates the attempt over dozens of turns. The team trained GLM-4.7-Flash on Terminal-Bench and SWE-bench tasks with Miles in both synchronous and fully asynchronous modes, 68 training steps and more than 2,400 episodes, all in sandboxes on Vera.
In the real training comparison, both sides used the same model, tasks, GPUs and settings for 5 steps and 160 episodes each, and with sandboxes on Turin each episode took 12% longer on average, with the four tasks taking 9%–15% longer. Model thinking time was the same on both, so nearly all of the difference came from sandbox work: Turin took 53 seconds longer per episode, of which agent commands took 26 seconds longer (60% more than on Vera), the verifier 11 seconds longer (24% more), the rest of the sandbox work 12 seconds longer, and waiting for the model only 4 seconds longer. Before comparing, the machines were made busy: each CPU socket kept running 176 other tasks, exactly filling the 176 threads of one Vera socket, and training ran on top of that.
Perspective
The comparison covers one Vera socket against one x86 socket, on AWS m8a.metal-24xl (Turin) and m8i.metal-48xl (Xeon 6), with x86 on AWS default settings and Vera on NVIDIA-tuned boot parameters; the training comparison used GLM-4.7-Flash, 8 H200 GPUs and 4 SWE-bench Verified tasks for 5 steps and 160 episodes per side.
The widening to 33% at 8 times faster inference and 41% at 32 times is a projection, not a measurement, and assumes only model thinking time shrinks while sandbox work stays as measured; the text also notes that faster inference sends commands more densely and keeps the machine busier, making the estimate conservative. The 12% training-speed gain is likewise an estimate, not a measurement, and assumes the GPUs keep up.