Nvidia Vera CPU gains up to 39% with software tuning, Nvidia says at SC26
Nvidia told attendees at Supercomputing 2026 that its Vera processor, built on custom Olympus cores, needs architecture-specific code to reach its full potential. Tuned kernels lifted performance by nearly 39% in some demanding workloads.

Key takeaways
- Nvidia said at SC26 that architecture-specific software tuning can raise Vera CPU performance by nearly 39% in certain demanding workloads.
- Vera uses Nvidia's proprietary Olympus cores with 88 cores, 176 threads and six vector pipelines, replacing the licensed Neoverse V2 cores used in Grace.
- In a matrix multiplication benchmark, an Olympus-tuned kernel lifted Vera to 7.53 instructions per clock, beating Grace by nearly 67% per core and per clock.
- Nvidia positions Vera for agentic AI and similar workloads, not as a general-purpose server CPU for tasks such as web hosting.
In this article
What Nvidia announced at SC26
Nvidia used the Supercomputing 2026 conference to argue that its Vera processor still has considerable untapped performance. According to Tom's Hardware, the company said that code tuned specifically for the chip's architecture can lift performance by nearly 39% in certain computationally heavy workloads.
Vera is built around Nvidia's own Olympus cores, which are Arm-based designs developed in-house rather than licensed from Arm itself. The chip works both as a standalone data-centre processor and as the host CPU for Nvidia's Vera Rubin NVL72 server systems.
The message from Nvidia was straightforward: existing software already runs well on Vera, but it does not always exploit everything the hardware offers. Adaptation is needed to unlock the full speed of the custom cores.
How Olympus differs from Grace
Nvidia's previous data-centre processor, Grace, used off-the-shelf Neoverse V2 cores designed by Arm. Vera instead relies on Olympus, a proprietary microarchitecture built to maximise performance from a single core while also sustaining high total throughput.
To achieve this, Olympus carries a wide front end with a 10-wide decoder, a neural branch predictor and a deep out-of-order execution engine. It also has six vector execution units, four load pipelines and two store pipelines, supported by large private caches and advanced prefetch mechanisms.
Nvidia chose 88 high-performance cores with support for its own spatial multithreading technology, giving 176 threads in total, rather than packing in as many cores as possible. Hannah Coutand, who leads Vera CPU marketing at Nvidia, told Tom's Hardware that "CPU design is really just a game of trade-offs". She explained that filling a die with many cores leaves no room to maximise instructions per clock.
What the benchmarks showed

Nvidia's researchers measured the two architectures using a double-precision general matrix multiplication workload built on an assembly kernel from OpenBLAS, an open-source linear algebra library. OpenBLAS includes a kernel tuned for the Neoverse V2 cores in Grace, but no kernel written specifically for Olympus.
Grace, running the same code, managed 4.51 instructions per clock and 13.03 FP64 operations per cycle. That put Vera roughly 20% ahead on identical software.
On this particular workload, measured per core and per clock, Vera beat its predecessor by nearly 67%.
The hardware itself explains part of the gap. Neoverse V2 has four 128-bit SIMD pipelines capable of 16 FP64 operations per clock, while Olympus has six such pipelines that can handle up to 24 FP64 operations per cycle, a 50% increase in peak throughput.
Software was the other part. The Neoverse V2 kernel used 81.4% of Grace's peak FP64 throughput but only about 66.5% of Vera's. After adjusting matrix tile organisation, register allocation and instruction scheduling to suit Olympus, the researchers reached 90.6% of Vera's peak FP64 throughput, according to Nvidia's tutorial published at SC26.
Not a general-purpose processor
Nvidia is explicit that Vera is not aimed at every server task. Coutand said the chip "is not a general-purpose CPU" and noted that a web server for a cloud data centre would be better served by other choices.
The company originally designed the Olympus microarchitecture with data processing applications in mind. In practice, its defining traits, namely high per-core speed, high memory bandwidth and low memory access latency, turned out to match the needs of agentic AI workloads as well, according to Coutand.
She added that Vera has what she described as great fungibility across many workloads, provided those workloads value the things the chip does well. Nvidia therefore positions Vera primarily for AI-related computing rather than as a drop-in replacement for conventional server processors.
Why the tuning gap matters
Software written for existing Arm-based data-centre processors will still run on Vera and benefit from its capabilities without any changes. The SC26 results, however, show that at least some applications must be fine-tuned for the new microarchitecture before they can use its execution resources fully.
It remains unclear how much performance Olympus-specific optimisation could add in workloads other than matrix multiplication. The 39% figure comes from Nvidia's own demonstration, so independent confirmation is not yet available.
As Nvidia pushes Vera into the standalone server CPU market rather than using it only alongside its data-centre GPUs, it must persuade performance-critical software makers to optimise for its chips, as Intel and AMD have done for decades. Nvidia's long experience building the CUDA ecosystem for its GPUs could help it drive similar optimisation efforts for Vera as adoption grows.
Sources
Spotted a mistake? See our corrections policy. Read how News by OnDevice works in our editorial policy.



