E2B Tests NVIDIA Vera for High-Density Agent Sandbox Fleets

E2B published pre-production test results on November 9, 2026, showing that four tuned runs on one dual-socket NVIDIA Vera server sustained 1,400 to 1,550 loaded sandboxes, while separate microbenchmarks measured higher processor and memory throughput than an AMD EPYC 9B45 reference.

The evaluation is not a production deployment. E2B provides isolated microVMs through an SDK, allowing agents to execute commands, work with files and return to preserved running state. Each sandbox resumes from a snapshot with memory loaded as needed. Much of that session work happens on the CPU even when the model runs elsewhere, so E2B tested both individual operations and large numbers of concurrent environments.

Processor and Memory Microbenchmarks

The Vera tests ran October 9–13, 2026, on one pre-production dual-socket server with 176 physical cores and 1,479 GiB of RAM. AMD reference measurements were taken in August 2026. In a bare-host, single-thread compute test, E2B reports 7,759 events per second for Vera and 5,131 for the AMD EPYC 9B45, or 1.51 times the reference throughput, using the median of at least three runs. Compute testing used Vera’s 64K-page kernel.

The memory test populated a 2 GiB region using 4K kernels and 2 MiB huge pages, the same huge-page backing used by E2B’s sandboxes. E2B reports 62.8 GB/s for Vera and 17.4 GB/s for AMD, equivalent to 3.61 times the throughput. This measures memory population rather than a complete sandbox startup, an important operation for a runtime that repeatedly restores snapshot-backed environments.

The comparison has narrow limits. The charts cover named systems in specific microbenchmarks and do not establish equivalent end-to-end agent speedups or a fleet-capacity advantage over AMD. E2B also says gains for individual tools and applications will vary.

Full-Stack Capacity and Host Tuning

For the full-stack work, E2B used E2B Embed, its single-machine packaging of the open-source E2B Runtime and dashboard, with the stack running directly on the host. Embed provides Firecracker microVM isolation and Python and JavaScript SDK support. The evaluation covered creation, command execution, snapshot restore, capacity and repeated creation and teardown.

E2B reports 50 ms warm, sequential sandbox starts and 3.8 ms command round trips at low load, with each figure coming from separate Vera bare-metal tests rather than fleet-load measurements. A separate churn campaign completed 22,145 create-and-destroy cycles with zero errors.

In the mixed capacity workload, each sandbox repeatedly performed a CPU burst, a 10 MiB disk write and a pause, while one in 10 ran continuously. At the upper end of the capacity range, E2B reports 95th-percentile command round-trip latency of 255 ms. When every sandbox ran continuously, a separate test sustained 550 sandboxes.

Those capacity runs reserved 512 MiB per sandbox. Their objectives required 95th-percentile latency to remain within 5× each run’s warmed burst baseline and imposed a 1% error limit per step. The result therefore describes capacity under specified responsiveness and reliability targets, rather than an absolute loading limit.

Host configuration was a major constraint. Default settings supported 320 sandboxes; network-neighbor tuning raised that to 1,000 while leaving a 1,024-device limit in place. Full tuning was required for the main capacity result. E2B describes that progression as a same-server configuration comparison, not a Vera-versus-x86 speedup.

Evaluation Limits and Planned Tests

The pre-production status leaves several unresolved issues. E2B found that x86 reference systems still completed some restore operations more quickly, while nested virtualization requires additional work. No cost or power measurements were taken, leaving operating economics outside the evaluation.

E2B’s stated next steps are a multi-node Vera cluster, cold resumes backed by object storage and an x86 comparison using the same E2B packaging and host tuning. The broader deployment case now rests on those follow-ups: whether the full-stack results hold against a comparably packaged and tuned x86 system, and whether the single-server capacity results extend across a cluster.