Nvidia and Oracle announced general availability of Nvidia's Vera Rubin NVL72 rack and Vera CPUs on Oracle Cloud Infrastructure Superclusters at Oracle AI World on Tuesday, giving Nvidia its first standalone CPU deployment at a hyperscaler, a week after Google Cloud made Vera Rubin generally available.
Oracle's announcement stands out because it includes standalone Vera instances on OCI Superclusters, without requiring the Rubin GPU. Neoclouds including Nscale, CoreWeave and Nebius already have Vera instances under way, but a hyperscaler offering the CPU on its own sends a different signal, according to Tom's Hardware. OCI Superclusters otherwise offers Ampere Altra for Arm-based instances, with hyperscaler CPU fleets primarily focused on x86.
Tom's Hardware, which benchmarked a Vera server for a forthcoming review (not on OCI infrastructure), found Vera more than doubling the memory bandwidth of a dual Intel Xeon 6980P system paired with MRDIMMs at 8,800 MT/s on the Stream triad test. The chip's LPDDR5X-9600 memory, monolithic die and distributed mesh for caching drive the result, and setting OpenMP to its "spread" policy boosted bandwidth further. Nvidia has said the memory system is capable of 1.2 TB/s of bandwidth.
The publication also recorded strong Python performance, relevant to agentic frameworks such as CrewAI and LangGraph, and strong SQLite results, which it said shows crossover with general data center workloads. Nvidia says Vera is a CPU built for agentic AI.
Nvidia's Ian Buck, the executive behind CUDA, said in March when the chip was revealed that the world could not be served by a single 88-core SKU. AMD and Intel still lead on single-socket density, with AMD's upcoming EPYC Venice CPUs able to pack up to 512 threads in a single socket. Oracle is approaching the deployment cautiously, the company said, and will also offer Nvidia's HGX NVL8, which pairs eight Rubin GPUs with a dual, 64-core Intel Xeon 6 host with 4TB of DDR5 memory.
Analysis
Why This Matters
- This is Nvidia's first standalone CPU at a hyperscaler, placing its own silicon alongside the x86 competition for general compute rather than only as a host for its GPUs.
- Vera's memory bandwidth and Python performance could make it relevant for both agentic AI workloads and conventional data-heavy applications.
- Oracle and Google Cloud now both offer the platform, but how much it matters depends on whether customers run Vera in production beyond GPU-bound work.
Background
Nvidia built its data center business on GPUs and accelerators, while hyperscaler fleets have been dominated by x86 processors from AMD and Intel. Vera is Nvidia's own CPU design, and its availability as a standalone instance at a hyperscaler is a first for the company. OCI Superclusters previously offered Ampere Altra as an Arm-based option. AMD and Intel retain an edge in single-socket density, with AMD's EPYC Venice line due to reach 512 threads per socket.
Key Perspectives
Nvidia: says Vera is a CPU built for agentic AI, citing a memory system rated at 1.2 TB/s of bandwidth.
Tom's Hardware: its independent testing, run on a local server rather than OCI, found Vera more than doubling the memory bandwidth of a dual Xeon 6980P system and showing strong Python and SQLite performance, which it says points to crossover with general data center workloads.
Oracle: is approaching the deployment cautiously, pairing standalone Vera instances with an eight-GPU Rubin cluster option as it brings up a multi-gigawatt deployment.
What to Watch
- Whether hyperscaler customers deploy Vera for general workloads or only alongside Rubin GPUs.
- AMD's EPYC Venice launch, which will test Vera against denser x86 sockets.
- Tom's Hardware's full review, which will publish the test setup and further benchmark results.
Checked
12 claims in this story checked against 1 source: 8 supported.
Show each claim
- SupportedNvidia’s Oracle deployment is its first standalone CPU deployment at a hyperscaler.Latest from Tom's Hardware: “It’s Nvidia’s first standalone CPU push into the hyperscale data center, placing Vera alongside the x86 competition regardless of the GPU compute.”
- SupportedIndependent benchmarks show Vera leading dual Xeon systems in memory bandwidth.Latest from Tom's Hardware: “our independent benchmarking of a local server reveals Vera's massive memory bandwidth lead over dual Xeon 6 server.”
- Partly supportedOracle unveils an eight-GPU Rubin cluster.Latest from Tom's Hardware: “The company said it will also offer Nvidia’s HGX NVL8, bringing eight Rubin GPUs together in a cluster”
- SupportedNvidia and Oracle announced general availability of the Vera Rubin NVL72 rack and Vera CPUs on OCI Superclusters at Oracle AI World on Tuesday.Latest from Tom's Hardware: “Nvidia and Oracle have announced general availability of Nvidia’s Vera Rubin NVL72 rack and Vera CPUs on Oracle Cloud Infrastructure (OCI) Superclusters at Oracle AI World on Tuesday.”
- Partly supportedGoogle Cloud made Vera Rubin generally available a week before Oracle’s announcement.Latest from Tom's Hardware: “The announcement comes hot on the heels of Google Cloud announcing general availability for Vera Rubin last week”
- SupportedOracle’s offering includes standalone Vera instances that do not require a Rubin GPU.Latest from Tom's Hardware: “Oracle’s announcement is interesting, specifically calling out standalone Vera instances on OCI Superclusters. ... placing Vera alongside the x86 competition regardless of the GPU compute.”
- SupportedTom’s Hardware found Vera more than doubled the Stream triad memory bandwidth of a dual Xeon 6980P system with 8,800 MT/s MRDIMMs.Latest from Tom's Hardware: “allows it to more than double the bandwidth available on a dual Intel Xeon 6980P system, even when it’s paired with MRDIMMs sporting speeds of 8,800 MT/s. We’re showing triad here”
- SupportedLPDDR5X-9600 memory, a monolithic die and a distributed mesh for caching contribute to Vera’s bandwidth result.Latest from Tom's Hardware: “Nvidia’s use of LPDDR5X-9600, along with a monolithic die and distributed mesh for caching, allows it to more than double the bandwidth available”
- SupportedNvidia says Vera’s memory system is capable of 1.2 TB/s of bandwidth.Latest from Tom's Hardware: “Nvidia has said the memory system is capable of 1.2 TB/s of bandwidth”
- Partly supportedIan Buck, described as the executive behind CUDA, said in March that the world could not be served by a single 88-core SKU.Latest from Tom's Hardware: “Ian Buck, Nvidia’s mind behind CUDA and the face of Nvidia’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.”
- Partly supportedOracle said it was approaching the deployment cautiously.Latest from Tom's Hardware: “Oracle is approaching cautiously, however, as any hyperscaler should be when bringing up a multi-gigawatt deployment. The company said it will also offer Nvidia’s HGX NVL8”
- SupportedThe HGX NVL8 offer combines eight Rubin GPUs with a dual 64-core Intel Xeon 6 host and 4TB of DDR5 memory.Latest from Tom's Hardware: “bringing eight Rubin GPUs together in a cluster, connected to a dual, 64-core Intel Xeon 6 host with 4TB of DDR5 memory.”