Nvidia says Vera CPU delivers 3.4 TB/s bandwidth and higher single-thread performance

Vera is built on a single monolithic compute die, uses a custom LPDDR5X memory subsystem and the second‑generation Scalable Coherent Fabric to cut internal latency. Nvidia claims the chip has a 10‑wide decode width, delivers about a 50 percent higher IPC than its Grace predecessor, and is 1.5x faster than Grace overall while being 2x faster than leading x86 parts on agentic workloads such as Python scripting and LLVM compilation. A third‑party review by Phoronix reportedly found the proprietary spatial multithreading feature adds roughly a 20 percent performance uplift. The CPU also provides 3.4 terabytes per second of bisection bandwidth, which Nvidia says is 3x higher than the leading x86 chip.

The company notes a broader industry move toward disaggregated infrastructure, with some sandbox providers running up to 400,000 sandboxes for a single customer. While Nvidia does not give an official CPU‑to‑GPU ratio, it stresses that standalone Vera racks are intended for heavy agent tool‑call workloads and reinforcement‑learning pipelines.