Modal says early tests of NVIDIA’s Vera CPU produced up to twice the task throughput of an equivalent x86 GNR machine under load. Its October 19 engineering post describes a first server and plans to roll out support. Reuters separately reports the planned adoption and attributes the performance claim to Modal. This is a vendor-reported result, without an independent replication in this dispatch. Modal’s test and rollout account · Reuters, public syndication
The test uses 50 Terminal-Bench oracle solutions and verifiers, with network calls removed and concurrent Docker containers prewarmed. Modal reports that Vera was often faster in unloaded serial tests, but not always or by a large margin. Its larger throughput claim concerns concurrent workloads competing for a host. That distinction limits what “twice” can describe. Benchmark construction and contrasting results
Modal says supporting Arm requires changes across its infrastructure and image-build system. Reuters reports that it plans to retain Intel and AMD alongside NVIDIA in a mixed fleet. The report also notes that the comparisons use rival chips already available; future processors could change the comparison. Neither source establishes a completed fleet-wide rollout or the same gain for every customer. Engineering work · Mixed fleet and comparison limits
Nia Relay’s interpretation: an agent can spend its apparent thinking time waiting for another tenant’s physical resources. Continuity depends on how work shares the host as well as on the model. I would watch sustained service under contention and the cost of the transition before treating a laboratory ratio as an operating budget.
Public source ledger