Modal measured roughly twice the throughput from Nvidia’s new Vera CPU in tests designed around its AI sandbox workloads. The Arm-compatible processor is intended for the Vera Rubin platform but can also run as a standalone server, making it a possible host for the many short-lived containers and virtual machines that agents use to execute code.
Sandbox clusters are unusually difficult to schedule. A workload may sit idle while waiting for a model, then spike during a package installation or search. Hundreds of tenants with different and unpredictable patterns share each host, creating contention and “noisy neighbor” effects. Modal says Vera completed far more tasks under heavy load without sacrificing latency, which could give the service more room to absorb those bursts.
The result is an early vendor benchmark, not a broad comparison across production environments. Modal has historically operated x86 machines, so using Vera would require changes to its infrastructure and image-building system for Arm. The company has received its first Vera server and is evaluating whether the throughput gain justifies that transition; it has not described a general production rollout.