
KEY POINTS
- Nvidia plans to cover most use cases with a single Vera SKU that has 88 cores clocking up to 3.3 GHz.
- Executive Ian Finder says rivals have long over-segmented their CPU lineups.
- An analyst says Nvidia may be quietly binning chips into variants for customers such as hyperscalers.
Nvidia is preparing to cover most datacenter CPU needs with one product, according to The Register. Ian Finder, who leads development of the company's datacenter CPUs, argued that the large number of SKUs from Intel and AMD is unnecessary for the vast majority of use cases.
Finder spoke to the outlet on the sidelines of Supercomputing 2026 this week. He said there had been what he called an overskewing for many years on the part of other vendors. Nvidia's first datacenter CPU, Grace, arrived in 2023, but it was mostly deployed in AI head nodes and high-end network cards, plus a few standalone deployments by hyperscalers such as Meta.
Its successor, Vera, is meant for a wider role. Besides appearing in Vera-Rubin superchips, it will be sold in standalone CPU racks as well as storage and context memory systems. This week Nvidia announced a rack with 192 Vera CPUs aimed at high-performance computing workloads that do not map well to GPUs. On paper, one SKU with 88 cores running at up to 3.3 GHz powers all of these products.
Finder explained that most datacenter CPUs have not been heavily loaded, so designers have tended to optimize for unloaded latency. Chips that lean on many cores for throughput often see higher latency under load. Optimizing for loaded latency, in turn, usually costs throughput. AMD's recent Epyc generations show the split, with dense cores for throughput and frequency-optimized cores for latency.
Analyst Ian Cutress of More than Moore told the outlet that congestion on connectivity buses raises memory latency as more cores become active. He said Nvidia is right that designers sometimes have to choose between core count and core latency. He also suggested Nvidia is confident enough in its market understanding to think one chip will serve most buyers, and that the company may want to capture that bulk segment.
Nvidia tried to avoid the trade-off by building a balanced design that holds low latency and high throughput with all 88 cores busy. One compromise is a less powerful vector unit, which matters less for buyers who also have GPUs for vectorized code. Finder said such choices allow performance per core similar to frequency-optimized x86 parts, with throughput per watt similar to density-optimized ones, for work such as software engineering agents, databases and analytics.
The Register noted that Nvidia is not alone in this thinking, since cloud providers have long skipped the highest core-count Intel and AMD parts in favor of a more rounded mix. AMD, meanwhile, has argued that agentic workloads are many different workloads, so a broader portfolio gives customers more choice.
Vera also offers some flexibility of its own. It is one of few Arm chips supporting simultaneous multithreading, which splits each of its Olympus cores so the chip behaves more like 176 narrower cores. Cutress also suspects Nvidia may already bin defective chips into lower-core variants, such as a 72-core part for a hyperscaler, without advertising them as separate SKUs.
