Nvidia bets a single 88-core Vera CPU can cover most datacenter workloads

Nvidia's head of datacenter CPU development says rivals have offered too many chip variants, and Vera is meant to deliver both low latency and high throughput in one design.

· 2 min read

a close up of a computer with green lights
Photo: Tyler / Unsplash

Nvidia is pursuing a one-chip strategy for datacenter processors, according to a report from The Register. Ian Finder, who leads development of Nvidia's datacenter CPUs, said most customers do not need the long list of variants that Intel and AMD release with each generation.

Speaking to the outlet at Supercomputing 2026 this week, Finder described what he called an "overskewing" by competitors over many years. Intel Xeon and AMD Epyc generations typically include dozens of models, covering high core counts, high clock speeds and low-power edge parts.

Nvidia is still a newcomer in this market. Its first datacenter CPU, Grace, arrived in 2023 but was mostly used in AI head nodes and high-end network cards, plus a few standalone deployments at hyperscalers such as Meta. Its successor, Vera, is meant for a wider role. It will appear in Vera-Rubin superchips and also in standalone CPU racks, storage systems and context memory systems. This week Nvidia announced a rack holding 192 Vera CPUs for high-performance computing work that does not suit GPUs well. On paper, one SKU with 88 cores running up to 3.3 GHz powers all of these products.

Finder explained the reasoning in terms of latency. Most datacenter CPUs have not run under heavy load, so designers tuned them for latency when idle. Chips that depend on many cores for throughput often see higher latency once they are busy. Optimizing for latency under load usually costs throughput. The Register pointed to AMD's recent Epyc generations, which include dense Zen-C cores for throughput and frequency-optimized cores for latency.

Analyst Ian Cutress of More than Moore agreed with the underlying point. He said memory latency rises as more cores run because of congestion on the connecting buses, so designers sometimes must choose between core count and per-core latency. He also suggested Nvidia may simply want to capture the bulk of the market.

Vera aims to keep latency low and throughput high even with all 88 cores active. One compromise is a less powerful vector unit, which matters less when GPUs handle vectorized code. Finder argued the design can match the per-core performance of frequency-optimized x86 chips and the throughput per watt of density-optimized ones, for work such as software engineering agents, databases and analytics.

Nvidia's approach has precedent, the report notes. Cloud providers have long skipped the highest core-count Intel and AMD parts in favor of balanced chips, since they cannot predict customer workloads. AMD, however, has argued that agents are many workloads rather than one, and that a broader lineup gives customers more choice.

Vera also has flexibility built in. It is one of few Arm chips with simultaneous multithreading, and enabling it splits each of the 88 Olympus cores so the chip behaves like 176 narrower cores. Cutress also suspects Nvidia sorts chips with defective cores into lower-tier variants for customers, such as a 72-core version for a hyperscaler, without marketing them as separate SKUs. For now, the report says, Nvidia sells just one.

Related