Nvidia spent years known for a well-defined formula: designing extremely high-performance GPUs and leaving general-purpose computing in the hands of partners like Intel and AMD. With Vera, that equation changed quietly but profoundly. The company detailed the architecture of its new Vera CPU and, above all, of the Olympus core that powers it — a general-purpose processor that Nvidia neither bought nor licensed off the shelf, but designed from scratch to solve a specific problem: sustaining agentic-AI workloads, which are heavy on logic, branches and pointer manipulation, very different from the dense math that dominates model training.
The most striking technical detail is the core count. Each Vera CPU houses 88 Olympus cores on a single monolithic die, a choice that directly contrasts with the strategy of x86 rivals, which have been moving toward multi-chip designs — so-called chiplets — in search of scale and cost reduction. Olympus is a core compatible with the ARMv9.2 ISA, but it is not one of the ready-made cores Arm licenses to manufacturers. It is Nvidia's own design, developed in-house and optimized with a clear goal: maximizing single-thread execution performance. That translates into a ten-instruction-per-cycle decode front end and a core with very deep out-of-order execution, the kind of engineering traditionally associated with the most complex server processors.
The rationale behind this design is directly tied to Nvidia's vision of the future of computing. In agentic-AI scenarios, the system does not spend its time only multiplying matrices; it executes trillions of small logical decisions, follows code paths full of conditional branches, manipulates linked data structures and waits for responses from other components. That profile is essentially the opposite of what a GPU does well. GPUs are machines of massive parallelism, excellent when the work is uniform and predictable; agentic AI, however, is dotted with dependencies, latency bubbles and sequences of decisions that demand a fast and intelligent core on its own. Olympus was born precisely to fill that gap, letting Nvidia offer a complete system — GPU, CPU and interconnect — designed under the same architectural vision.
There is a strategic dimension that goes beyond engineering. By designing its own core instead of licensing Arm's designs, Nvidia gains control over its product roadmap, margins and differentiation, but also assumes the costs and risks of maintaining a world-class CPU design team, something few companies on the planet have the resources to sustain. The choice of a monolithic die with 88 cores, in contrast to rivals' chiplets, suggests the company prioritizes performance and cache coherence in a single silicon, even if that imposes manufacturing-yield and cost challenges. For datacenter customers, the signal is clear: Nvidia wants to be a supplier of complete AI systems, not just accelerators, contesting territory that historically belonged to Intel and AMD in the server-processor field.
The open point is how this design will behave in practice against competitors. Specification numbers tell only part of the story; real performance in agentic workloads will depend on factors such as sustained frequency, energy efficiency, the quality of the interconnect with the GPUs and the software ecosystem. The big question is whether the bet on single-thread performance and a monolithic core will prove superior to Intel's and AMD's modular designs, or to Arm-licensed alternatives that are also gaining ground in the server market. If the strategy works, Nvidia consolidates a practically unbeatable position in the AI datacenter. If it runs into practical limitations, the market will get a valuable lesson about the limits of an ambitious, integrated design. What is certain, for now, is that the company has stopped being only the world's most valuable GPU maker to also become a CPU architect — and Olympus is living proof of that transformation.
Sources: The Register, Tom's Hardware, NVIDIA Technical Blog, TechRadar Pro
✓ Independent sources cross-checked and verified before publishing
