NVIDIA's Vera has broken cover as the company's first standalone data-centre CPU, built around 88 custom Olympus cores and aimed squarely at AI server workloads. For UK buyers weighing Vera against entrenched Xeon and EPYC estates, the detail matters more than the marketing.
View the data behind this chart
| Vera Superchip | EPYC 9755 | |
|---|---|---|
| SPEC int score | score925 | score898 |
Why a standalone NVIDIA CPU changes the conversation
Until now, NVIDIA's Grace CPUs existed mainly as GPU chaperones, built from off-the-shelf Arm Neoverse or Cortex cores and sold as part of a package. Vera is different: it's a fully custom Armv9.2 design that hyperscalers can deploy independently of NVIDIA silicon, and the early adopter list — Alibaba, ByteDance, Meta, Oracle, CoreWeave, Lambda, Nebius and NScale — signals this is being taken seriously as infrastructure, not just an accessory chip.
For UK operators, that's the real headline. NVIDIA isn't just selling GPUs with a CPU bolted on; it's making a direct pitch for the general-purpose compute budget that has historically gone to Intel and AMD, at least within AI-heavy estates.
Inside Olympus: the architecture behind the numbers
Vera's compute die is monolithic — all 88 cores on one slab of silicon reportedly fabbed on TSMC's 3nm process — while memory and I/O sit on separate chiplets, including eight LPDDR5X controllers and dedicated dies for PCIe 6.4/CXL 3.1 and NVLink Chip-to-Chip. It's a layout that echoes Amazon's Graviton 4 rather than AMD's more heavily chipletised approach, and NVIDIA argues it buys better core-to-core bandwidth and latency.
The Olympus core itself builds on Arm IP but adds a custom neural branch predictor capable of exploring two branches per cycle, plus memory renaming and value prediction schemes designed to cut pipeline stalls. NVIDIA singles out pointer-heavy, object-oriented code and branch-heavy AI-generated scripts — the kind of workloads agentic AI systems actually run — as the direct beneficiaries. If you want the full technical breakdown, it's worth reading how this fits the broader picture in our AI servers research.
Vera vs Xeon vs EPYC: what the benchmarks actually show
NVIDIA's published SPEC CPU 2026 integer results put the dual-socket Vera Superchip at 925 versus 898 for AMD's EPYC 9755 — a modest but real lead on that specific metric, with no dedicated floating-point figures disclosed yet. The first independent-adjacent benchmarks came from Phoronix, though the testing was curated by NVIDIA at its own Santa Clara facility rather than run on fully arms-length retail hardware, which is worth bearing in mind when reading the headline claims.
NVIDIA also cites a single-socket TDP of 450W with under 30W spent on memory power, over 4x memory bandwidth per core against traditional x86 chips in STREAM TRIAD testing, and a 1.6x geometric-mean gain over its own previous-generation Grace CPU. AMD hasn't stayed quiet: it's pointing to its forthcoming 256-core Zen 6 'Venice' part, which the company estimates could beat Vera by 3.3x at rack level — a reminder that today's numbers are a snapshot, not a settled verdict. Buyers should compare with Intel Xeon 6 CPUs before assuming Vera's lead generalises beyond these specific tests.
Where Vera actually fits in a GPU-centric rack
NVIDIA is pitching Vera for two roles: as the head node managing GPUs in Vera Rubin systems, and as a dedicated host for AI agents, which run on CPUs rather than GPUs even when the underlying models don't. The dual-socket Vera CPU Superchip links two chips over NVLink-C2C at 1.8 TB/s, delivering 176 cores, 352 threads and 2.4 TB/s of aggregate memory bandwidth — roughly double AMD's 2024-era Turin EPYCs. NVIDIA's rack-scale reference designs stack 128 of these Superchips into a single liquid-cooled rack, and the Vera Rubin NVL72 platform pairs 36 Vera CPUs with 72 Rubin GPUs.
That framing matters for procurement: Vera is being sold as a platform CPU for accelerator-dense racks, not a like-for-like swap into a general-purpose fleet. Anyone weighing this against a Blackwell deployment should read our take on NVIDIA Blackwell vs Rubin systems before committing budget.
What UK buyers should actually do next
If your workload is agentic AI orchestration, real-time inference pipelines or SQL-heavy analytics feeding LLM applications, Vera's memory bandwidth and branch-prediction advantages are genuinely relevant — NVIDIA's own partner data cites 3x faster large-scale SQL analytics with Starburst and up to 6x lower latency with Redpanda versus leading x86 chips. For everything else — standard virtualised estates, database servers, general compute — the existing Xeon and EPYC ecosystem still has the deployment maturity, software compatibility and support depth that most UK IT teams rely on.
Before switching platforms, it's worth running the numbers through a GPU server needs for LLM inference calculation to see whether Vera's rack-scale benefits actually translate to your workload, and to weigh the financing implications via server financing options. Teams planning a first deployment should also look at what it takes to build your UK on-prem AI cluster before assuming Vera slots in without infrastructure changes.
- 01The Register — NVIDIA's Vera CPU and the Olympus cores that power it: a deep dive · 1 August 2026
- 02ServeTheHome — Diving deeper on NVIDIA's Vera CPU: architectural details and SPEC CPU 2026 benchmarks · 1 July 2026
- 03Tom's Hardware — NVIDIA spills the beans on Vera CPU: SPEC benchmarks revealed, Olympus architecture detailed · 1 July 2026
- 04Tom's Hardware — NVIDIA's Vera CPU tested in common Linux benchmarks · 1 July 2026
- 05NVIDIA Developer Blog — NVIDIA Vera CPU delivers high performance, bandwidth and efficiency for AI factories · 1 June 2026
- 06Tom's Hardware — AMD fires back at NVIDIA claiming 256-core Zen 6 'Venice' beats Vera by 3.3x in rack-level performance · 15 July 2026
- 07NVIDIA — Vera Rubin NVL72 platform · 1 March 2026