Nvidia의 Vera CPU가 이제 Oracle OCI Superclusters에서 정식 출시되었으며, Nvidia가 NVL72 구성을 넘어서면서 독립형 Vera CPU도 Oracle Cloud의 멀티기가와트 배포에 도입될 예정입니다.

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. The announcement follows Google Cloud's general availability announcement for Vera Rubin last week, giving Nvidia two major hyperscaler wins. Oracle's announcement is notable for specifically calling out standalone Vera instances on OCI Superclusters. This article includes benchmarks run on a Vera server for a forthcoming review to illustrate the chip's performance characteristics. These tests were not conducted on OCI infrastructure.

Neoclouds have been promoting Vera for several weeks, with instances already running at Nscale, CoreWeave, and Nebius, among others, so it makes sense to see Vera in GPU-dominant, AI-focused neoclouds. A standalone deployment at a hyperscaler like Oracle sends a different signal. Although hyperscalers maintain large hardware fleets and sometimes rent capacity from neoclouds, their CPU offerings are primarily x86. Outside of Vera, OCI Superclusters offer only the Ampere Altra for Arm-based instances. This is Nvidia's first standalone CPU push into the hyperscale data center, placing Vera alongside x86 competitors regardless of GPU compute.

Vera is not meant to upend the x86 old guard, at least not on its own. When the chip was revealed in March, Ian Buck, the mastermind behind CUDA and the face of Nvidia's AI infrastructure buildout, said the world couldn't be served by a single 88-core SKU. However, Vera's presence at hyperscalers like Oracle opens the door to expansion in future CPU generations.

AMD and Intel lead in single-socket density, a lead that is set to grow in the next generation, as AMD's upcoming EPYC Venice CPUs can pack up to 512 threads into a single socket. Still, Vera offers distinct advantages depending on the workload, many of which are becoming increasingly relevant in the agentic AI era.

Memory bandwidth is one of Vera's strong suits, and that is not just a marketing claim. Nvidia has said the memory system is capable of 1.2 TB/s of bandwidth, and our testing shows Nvidia's lead extends beyond theoretical numbers.

Take the industry-standard Stream benchmark as an example. Nvidia's use of LPDDR5X-9600, along with a monolithic die and a distributed mesh for caching, allows Vera to more than double the bandwidth of a dual Intel Xeon 6980P system, even when that system is paired with MRDIMMs running at 8,800 MT/s. We show the Triad results here because it is the most demanding of Stream's operations, but Vera's leads are similar across Add, Copy, and Scale. This bandwidth is critical for certain agentic loops, particularly those involving large database queries or extensive documentation. The OpenMP "spread" policy also provides a major boost to Vera's memory bandwidth. Both AMD and Intel use centralized memory control and system-level cache, while Vera distributes its cache in a mesh across the die. With OpenMP set properly, the result is much higher memory bandwidth, achieved by binding processes as far from each other as possible.

Python performance is another critical vector for agentic AI. The agentic ecosystem is expanding, but much of the AI software foundation is built on Python. Common agentic frameworks such as CrewAI and LangGraph are written in Python, and although more frameworks are being built in languages like Rust and TypeScript, Python performance remains critical to agentic loops. Vera is also strong here, outperforming CPUs that are much more powerful on paper. Popular agentic harnesses are built on Python, as are many applications that run on it, including foundational AI libraries like PyTorch and TensorFlow.

Nvidia's focus on agentic AI makes sense given the current moment, but it risks missing the forest for the trees. Oracle's standalone Vera instances shift attention to the broader implications of Vera's entry into the data center. Some aspects of performance characterize agentic AI workloads, but many agent workloads are ordinary data center workloads to begin with, simply running at high concurrency. SQLite performance, for example, matters for any server querying a large database, regardless of whether the query was written by a developer, a person, or an AI agent. Nvidia's second-generation Scalable Coherency Fabric (SCF), which we observed in the Stream results, also appears in this real-world workload, where the database queries are largely bound by data movement across the system.

This crosstalk with general data center workloads is what makes Vera so interesting in the context of Oracle's announcement. Nvidia describes Vera as a CPU built for agentic AI, but its strength in general data center workloads is notable. It is not just a CPU to buy because you need access to Nvidia's accelerators. Depending on the workload, it may even be a better choice than x86 processors.

Ultimately, how customers of these hyperscalers use, or decide not to use, Vera in production will determine its success. Vera has demonstrated its value in agentic AI loops and is showing early promise in certain general-purpose workloads. Oracle's standalone deployment indicates there is demand for Vera that is not tied exclusively to a Rubin GPU.

Oracle is approaching this cautiously, as any hyperscaler should when bringing up a multi-gigawatt deployment. The company said it will also offer Nvidia's HGX NVL8, which brings eight Rubin GPUs together in a cluster connected to a dual 64-core Intel Xeon 6 host with 4TB of DDR5 memory.

Our Vera testing in this article is part of a larger benchmarking session, and we will share the technical details of the test setup in our forthcoming review. We will publish our full results on Tom's Hardware shortly.