Modal adopts NVIDIA Arm Vera CPUs after agent-sandbox benchmark doubles throughput over x86
Modal announced on October 19 that NVIDIA’s 88-core Arm Vera CPU doubled coding-agent sandbox task throughput compared with Intel Granite Rapids x86 under high multi-container concurrency.
As coding agents spawn thousands of short-lived containers to run compilers, test suites and headless browsers, host CPU memory bandwidth has become a primary infrastructure bottleneck.
On October 19, serverless cloud provider Modal published Agentless Agent Bench—a 50-task benchmark replaying real coding-agent tool traces with LLM API wait time stripped out—and announced it is adopting NVIDIA’s 88-core Arm Vera CPU for its sandbox fleet. [1][2]
At peak multi-container density, the Vera CPU (which pairs 88 Olympus Armv9.2 cores with 1.2 TB/s of coherent LPDDR5X memory bandwidth) achieved 2.0× the completed task throughput of an Intel Xeon 6 Granite Rapids x86 host. Single-sandbox cold-start latency also improved by 25% to 40% on compilation and dependency-install workloads. [1]
The benchmark isolates host-CPU execution after LLM token generation is removed; end-to-end agent wall-clock gains will be smaller whenever model inference dominates runtime. Teams running high-concurrency evaluation harnesses should check whether their container images build cleanly for linux/arm64. [1][2]
Take it further
Inspect Modal’s 50-task Agentless Agent Bench analysis For infrastructure teams running parallel coding-agent sandboxes or RL environments. Compare memory-bandwidth scaling against x86 hosts.
Sources & details 2 sources
Discuss in your voice app
Take the reporting, visual explanations and original sources into a conversation. Copy the brief, paste it into your preferred assistant, then start its voice mode.
Paste the brief into your voice app to discuss this reporting.Reporting sources
- Modal · NVIDIA Vera CPU sandbox benchmark
Primary October 19 engineering post detailing the 50-task Agentless Agent Bench methodology and 2.0× dense throughput result.
- The Next Web · Modal’s Vera CPU sandbox deployment
Independent October 19 coverage corroborating the 88-core Vera specifications and Modal’s infrastructure rollout.
Images & figures
Under high multi-container concurrency, Vera Arm CPUs doubled sandbox task throughput over Granite Rapids x86.
Euangelion · original data figure · Source · CC BY 4.0 · Download figure