Nvidia Vera CPUs Go Solo on Oracle Cloud: 1.2 TB/s Memory Play

Nvidia's Vera CPUs are now generally available as standalone instances on Oracle Cloud Infrastructure Superclusters — the first time the chipmaker has pushed its own CPU into hyperscale data centers without bundling it with GPUs. Nvidia and Oracle announced the move at Oracle AI World on Tuesday, alongside general availability of the Vera Rubin NVL72 rack on OCI.

Google Cloud announced Vera Rubin availability just last week, so Oracle is the second hyperscaler feather in Nvidia's cap within days. Neoclouds — Nscale, CoreWeave and Nebius among them — have been rolling out Vera instances for weeks. A standalone deployment at a hyperscaler is the different signal here.

Hyperscaler CPU fleets are overwhelmingly x86, with OCI Superclusters offering Arm only via the Ampere Altra otherwise. "It's Nvidia's first standalone CPU push into the hyperscale data center, placing Vera alongside the x86 competition regardless of the GPU compute," as the original analysis puts it. Ian Buck, the Nvidia executive behind CUDA and the company's AI infrastructure buildout, said in March when the chip was revealed that the world couldn't be served by a single, 88-core SKU. Vera, in other words, opens a door rather than upends the incumbents.

What do the benchmarks show?

Independent testing on a local Vera server — not on OCI infrastructure — shows where the chip's advantages actually sit. Nvidia rates Vera's memory system at 1.2 TB/s of bandwidth, and real numbers back that up.

On the industry-standard Stream benchmark's triad operation — the most demanding of Stream's tests — Vera more than doubles the bandwidth of a dual Intel Xeon 6980P system, even when that system pairs MRDIMMs running at 8,800 MT/s. The advantage holds across Stream's add, copy and scale operations too. The credit goes to LPDDR5X-9600 memory, a monolithic die, and a distributed mesh for caching, rather than the centralized memory control and system-level cache that AMD and Intel use.

There's a tuning caveat practitioners should know: the OpenMP "spread" policy delivers a massive bandwidth boost on Vera by binding processes as far from each other as possible. Set it up properly and the mesh pays off.

Why does this matter for agentic AI?

Two workload characteristics drive Vera's relevance. First, memory bandwidth — critical for agentic loops handling large database queries or extensive documentation. Second, Python. Frameworks like CrewAI and LangGraph, plus foundational libraries like PyTorch and TensorFlow, all run on Python, and Vera beats CPUs that look far more powerful on paper in that domain.

The broader story is that many agent workloads are regular data center workloads under high concurrency. SQLite performance, for instance, matters for any server querying a large database, whether a person, plain logic or an AI agent wrote the query. Nvidia's second-generation Scalable Coherency Fabric (SCF), at play in the Stream results, shows up again in these database queries, which are largely bound by data movement across the system.

What's the competitive picture?

AMD and Intel still lead on single-socket density, and that gap will widen: AMD's upcoming EPYC Venice CPUs pack up to 512 threads in a single socket. Vera's case rests on topology-dependent workload advantages, not raw core counts.

Oracle is moving carefully with what is a multi-gigawatt deployment. The company will also offer Nvidia's HGX NVL8, which clusters eight Rubin GPUs connected to a dual, 64-core Intel Xeon 6 host with 4TB of DDR5 memory — a telling hedge that keeps x86 in the lineup.

The standalone Vera deployment signals demand that isn't tied at the hip to a Rubin GPU. Depending on the workload, Vera may even be the better choice than the x86 guard — but real production adoption at hyperscale will be the verdict. Full technical details of the test setup arrive in a forthcoming review on Tom's Hardware, written by Senior CPU Analyst Jake Roach.