Modal Claims NVIDIA Vera CPUs Offer 2X Throughput Over Intel And AMD

Modal is a specialized cloud provider that offers on-demand computing infrastructure for AI developers, including isolated environments where AI coding agents can execute shell commands, compile software, manipulate repositories, and run other tools. These so-called agentic sandboxes can be unpredictable, with workloads that spend much of their time idle before suddenly demanding substantial CPU resources. Modal's business depends on running as many of these environments as possible without allowing one customer's workload to interfere with another's, and it says NVIDIA's new Vera CPU is as much as twice as fast as Intel's Granite Rapids.

To test Vera, Modal developed the Agentless Agent Bench, a collection of 50 Terminal-Bench solutions and verifiers covering tasks such as compilation, file searching, Git operations, and device emulation. The company reports up to twice the throughput of an equivalent Intel Granite Rapids (GNR) system when running these workloads concurrently. Rather than measuring how quickly a single task finishes, the throughput test measures how many verified tasks the system can complete within a fixed period as additional concurrent workloads are introduced.

Modal attributes much of Vera's advantage to memory bandwidth, echoing Intel's own earlier statements. NVIDIA's Arm-based CPU offers up to 1.2 TB/s of LPDDR5X bandwidth, compared with approximately 844 GB/s for Intel Granite Rapids and 576 GB/s for AMD's Zen 5-based EPYC platform. That's roughly 42% more bandwidth than Granite Rapids and more than twice as much as the EPYC system in this comparison. Modal says the extra bandwidth allows Vera to accommodate more concurrent sandbox workloads per core without suffering the performance degradation seen on competing systems.

The company's serial-latency results are more mixed. Modal also ran individual tasks on otherwise idle systems, measuring completion time without contention, and published results for ten representative workloads from its 50-task suite. Vera sometimes finishes first, but it doesn't win every test, and Modal itself acknowledges that its advantage is often modest. Some of the results also show striking differences between Arm and x86 systems, which could reflect varying levels of optimization across the two instruction-set architectures.

There's also some context worth keeping in mind when evaluating these findings. Modal's blog describes multiple visits to NVIDIA headquarters, including an occasion when the chipmaker's engineers confirmed that Modal's unexpected benchmark results were consistent with their expectations. The author says this was the moment he became a Vera believer. That's an understandable reaction to impressive results, but it also clearly illustrates that this is a vendor's own evaluation of a prospective hardware partner rather than an independent third-party benchmark.

That's notable because the commercial stakes are considerable. NVIDIA reportedly expects Vera CPU sales to reach at least $20 billion in its very first fiscal year, according to Reuters. For comparison, AMD's entire Data Center segment generated $16.6 billion in revenue during fiscal 2025, including both EPYC CPUs and Instinct accelerators as well as Pensando networking and some FPGA products, too. NVIDIA's target for just these CPUs would therefore exceed AMD's combined data center revenue for that year. Modal, meanwhile, plans to introduce Vera alongside its existing Intel and AMD fleet rather than replacing its x86 systems outright.

Vera's performance advantage also needs to be considered in the context of the hardware it faces. It's the newest CPU in Modal's comparison, and Intel's Granite Rapids and AMD's Zen 5 EPYC processors are general-purpose server CPUs designed to serve a broad range of workloads, from virtualization and databases to high-performance computing (HPC). Vera is very specifically optimized for the low-latency, high-concurrency workloads NVIDIA associates with AI agents and reinforcement learning. That focus makes sense for Modal's use case, but it's hardly surprising to see a newer, focused product outpacing older, more generalized products.

Vera's memory-bandwidth advantage may not last long, either. AMD's next-generation EPYC Venice processors are already arriving with up to 1.6 TB/s of memory bandwidth, matching the figure expected from Intel's forthcoming Diamond Rapids Xeons. Intel has confirmed that Diamond Rapids is coming in 2027, with its 16-channel memory configuration and second-generation MRDIMM support expected to deliver a substantial bandwidth increase over Granite Rapids. That's roughly a third more bandwidth than Vera's advertised 1.2 TB/s peak, giving both x86 vendors at least one lever to close the gap in the workloads Modal is testing.

Modal's results make a very compelling case that NVIDIA has built a CPU suited to the demands of the cloud provider's sandbox infrastructure, but don't get it twisted; Vera isn't automatically twice as fast as Intel or AMD CPUs. Instead, Vera appears to be a highly capable processor for this specific class of concurrent, agent-oriented workloads. We're keen to see more independent results emerge, especially as competing server platforms catch up on memory bandwidth.