Modal Tests Nvidia Vera at Up to Twice Intel GNR Task Throughput Under Concurrent Load

Modal plans to bring Nvidia’s Vera CPU to its Sandboxes service after its tests found up to twice the task throughput of an equivalent Intel GNR x86 machine under concurrent load.

Modal announced the plan Oct. 19 after receiving its first Vera server the previous week. Sandboxes currently runs on an exclusively x86 fleet, so the addition would bring Arm-based CPUs to the service. Modal has not set a date for customer access.

The benchmark ran 50 Terminal-Bench solutions and verifiers in prewarmed Docker containers. Modal removed network calls to focus on CPU work, including code compilation, file searches and Git operations. The throughput comparison measured tasks running concurrently, rather than tasks completed one at a time.

Under concurrent load, Vera delivered up to twice the task throughput of the equivalent Intel GNR machine. Task latency was equal to or sometimes lower than on the x86 machine. In separate serial runs across the 50 tasks, Vera often performed better, but the gains were neither consistent nor large.

The concurrent result matters for a service whose hosts can run hundreds of containers and virtual machines at once. CPU demand for an individual workload can rise to 10 times its baseline during work such as software installation. Sandboxes provides isolated environments where coding agents and other workloads can execute code.

Vera is an Arm-compatible CPU with 88 custom Olympus cores and up to 1.2 terabytes per second of LPDDR5X memory bandwidth. Modal must adapt its infrastructure and image-building system to support Arm before it can offer Vera-powered Sandboxes.

Modal’s result measures a different part of sandbox performance from CoreWeave’s production rollout of Vera Rubin NVL72. CoreWeave’s tests found that AI-agent sandboxes using Vera CPUs started more than three times faster; Modal measured coding-task throughput under simultaneous load.

Modal’s server is its first Vera system. Nvidia had previously delivered Vera systems to Anthropic, OpenAI, Oracle Cloud Infrastructure and SpaceXAI. AWS had also received its first Vera server. Modal’s next step is to complete the Arm support needed for Sandboxes, with no customer rollout date announced.