Somewhere in Nvidia's marketing department, someone is very pleased with themselves for naming a CPU after a Roman goddess and giving its cores names lifted from Mount Olympus. Subtlety was never the point here — dominance was.
88 Cores, One Very Custom Silicon Diet
At Hot Chips 2026, Nvidia pulled back the curtain on Vera, its in-house Arm server CPU, revealing 88 custom "Olympus" cores spread across 176 threads on a single monolithic die — a notable departure from the chiplet approach AMD and Intel favor. Each chip carries 164MB of unified L3 cache, a beefy neural branch predictor Nvidia says can resolve two branches per cycle with "zero penalty," and a memory subsystem built around LPDDR5X delivering 1.2 TB/s of bandwidth across up to 1.5TB of capacity.
Nvidia's own SPEC CPU 2026 benchmarks claim Vera delivers up to 1.8x the performance of AMD's EPYC Turin 9755 on agentic AI workloads — the caveat being that "Nvidia's own benchmarks" is doing a lot of work in that sentence, and independent verification hasn't caught up yet.
Building the Chip That Babysits the Chip
Vera isn't trying to replace Nvidia's GPUs — it's the CPU built to keep them fed. As AI shifts from "answer one prompt" to "run an autonomous agent that plans, calls tools, and writes code for minutes at a stretch," the bottleneck increasingly isn't raw GPU throughput, it's how fast a CPU can shuttle instructions, manage state, and keep an agent's reasoning loop from stalling out.
That's the quiet subtext here: Nvidia is no longer content selling the engine, it wants to sell the whole car, chassis included. A custom CPU purpose-built for agentic workloads is Nvidia betting that "AI agent" compute looks different enough from traditional workloads that owning both halves of the stack is worth the R&D spend.
When even the CPU has main-character energy, you know the AI hardware arms race has entered a new act.
Whatever silicon ends up under the hood, the real challenge for most businesses is still turning that horsepower into an AI integration that actually helps customers — that's the part we specialize in, so come tell us what you're trying to build.
Source: TechSpot