The Nvidia Vera CPU is a data center processor unveiled at GTC on March 16, 2026, purpose-built for agentic AI, reinforcement learning, and AI factory workloads. What makes it genuinely unusual is not its spec sheet — it is the strategy: Nvidia is producing exactly one model, one SKU, and CEO Jensen Huang has stated the company expects to build a multi-billion-dollar CPU business from that single configuration alone.
Key Takeaways
- The Nvidia Vera CPU features 88 custom Olympus cores based on Armv9.2-A, delivering 176 threads via NVIDIA Spatial Multithreading.
- Its LPDDR5X memory subsystem delivers up to 1.2 TB/s bandwidth — twice the bandwidth at half the power of traditional CPUs.
- NVLink-C2C provides 1.8 TB/s coherent CPU-GPU bandwidth, roughly 7x what PCIe Gen 6 offers.
- A 256-chip liquid-cooled rack sustains over 22,500 concurrent CPU environments at full performance.
- Partner availability is expected in the second half of 2026, with full production already underway.
What the Nvidia Vera CPU actually is
The Nvidia Vera CPU packs 88 custom NVIDIA-designed Olympus cores on an Armv9.2-A architecture, supporting NVIDIA Spatial Multithreading for 176 total threads. Unlike conventional hyperthreading, this SMT implementation uses physical resource partitioning rather than time-slicing — a meaningful architectural distinction that Nvidia says preserves per-thread performance under load.
The 162MB L3 cache spans all 88 cores in a single unified domain, eliminating the NUMA latency issues that plague multi-chiplet designs from competitors. Connectivity runs through a second-generation NVIDIA Scalable Coherency Fabric mesh derived from Arm’s CMN-700. Each core features a 10-wide instruction decode front-end, a neural branch predictor capable of two branches per cycle, and a custom prefetch engine tuned for graph database analytics. Vera is also the first CPU to support FP8 precision — a data type previously associated only with AI accelerators.
Memory is LPDDR5X, supporting up to 1.5 TB capacity — three times what the Grace predecessor offered — with bandwidth reaching 1.2 TB/s. The NVLink-C2C interconnect ties CPU to GPU at 1.8 TB/s of coherent bandwidth, which dwarfs PCIe Gen 6’s ceiling by a factor of seven. I/O includes PCIe 6.0 with 16 lanes and CXL 3.1 support.
How Vera compares to AMD EPYC, Intel Xeon, and Grace
Against AMD EPYC and Intel Xeon in the AI-optimized data center segment, Nvidia claims Vera delivers 1.5x performance-per-sandbox, 3x memory bandwidth per core, and 2x energy efficiency. These are vendor-reported figures without independent benchmark confirmation, so treat them as directional rather than definitive — but the architectural reasoning behind the efficiency claims is sound given the LPDDR5X subsystem’s power profile.
The comparison to Nvidia’s own Grace predecessor is more concrete. Grace shipped with 72 Arm Neoverse cores and 144 threads. Vera steps that up to 88 cores and 176 threads, with Nvidia claiming roughly 2x overall performance and 50% faster single-threaded execution. The memory capacity jump from Grace to Vera — supporting up to 1.5 TB versus Grace’s configuration — is arguably the more significant upgrade for AI workloads that need to keep enormous model weights resident in memory.
Where AMD and Intel still compete on breadth — offering multiple SKUs across different core counts, TDPs, and price bands — Nvidia is making the opposite bet. One chip, one configuration, maximum volume. It is a strategy that works when your target customer is a hyperscaler building thousands of identical AI factory racks, not an enterprise buyer optimizing for a specific workload at a specific budget.
The single-SKU strategy and what it means for AI infrastructure
Nvidia’s decision to ship exactly one Vera model is a deliberate departure from how the CPU market has operated for decades. AMD and Intel each maintain extensive product families with dozens of variants tuned for different power envelopes, core counts, and price points. Nvidia is treating the CPU layer the way it treats its top-end AI accelerators: one flagship, full production, no compromise configurations.
A 256-chip liquid-cooled Vera rack sustains over 22,500 concurrent CPU environments at full performance, with 3.4 TB/s bisection bandwidth and over 90% peak memory bandwidth maintained under load. Those figures describe a system built for the relentless, parallel demands of agentic AI inference and training pipelines — workloads where consistency matters as much as peak throughput. The rack-level design also sidesteps the thermal management complexity that comes with air-cooled deployments at this density.
Full production is already underway, with partner availability targeted for the second half of 2026. No pricing has been disclosed, which is typical for data center silicon at this stage — hyperscalers negotiate volume deals that bear little resemblance to any public list price.
Is the Nvidia Vera CPU coming to desktop or consumer PCs?
No. The Vera CPU is designed exclusively for data center and AI factory deployments. Nvidia has not announced any consumer or workstation variant, and the LPDDR5X memory subsystem and liquid-cooling requirements make a desktop version implausible in the near term.
How does Vera differ from the Grace CPU it replaces?
Vera upgrades Grace’s 72 Arm Neoverse cores to 88 custom Olympus cores, increases thread count from 144 to 176, triples memory capacity support to 1.5 TB, and doubles claimed overall performance. The shift from Neoverse to custom Olympus cores also adds FP8 precision support, which Grace did not offer.
When will the Nvidia Vera CPU be available?
Full production is underway as of the March 2026 GTC announcement. Partner availability is expected in the second half of 2026, though Nvidia has not specified which partners or regions will receive initial allocations.
The single-SKU gamble is either Nvidia’s most confident move yet or its most exposed. If hyperscalers standardize on Vera-based AI factory racks at the scale Jensen Huang is projecting, producing one chip in massive volume is exactly the right call — lower manufacturing complexity, higher yields, simpler supply chains. The Nvidia Vera CPU is not trying to compete with AMD EPYC on versatility. It is betting that the AI infrastructure buildout is large enough, and homogeneous enough, that versatility does not matter.
Edited by the All Things Geek team.
Source: Tom's Hardware
