Nvidia claims its Vera CPU can gain up to 39% performance boost via software optimization at SC26 — architectural fine-t

Semiconductors

Nvidia Says Software Tuning Can Lift Vera CPU Performance 39%

Nvidia says Olympus-tuned software can lift its Vera CPU up to 39%, with per-clock DGEMM results nearly 67% above Grace once kernels are rewritten for the new cores.

By
Rebecca Stone
Filed
Channel
Semiconductors
Read
3 min read

Nvidia claims software optimizations tailored to its Vera CPU can deliver up to a 39% performance uplift in computationally intensive workloads. The company presented the figure at Supercomputing 2026 (SC26), and it comes with a caveat: existing code written for Arm-based server chips often leaves Vera's custom Olympus cores underutilized until developers rework it.

Vera runs old software faster out of the box. But Nvidia's own benchmark results, published in an SC26 tutorial, show the gap between what the silicon can do and what off-the-shelf kernels actually extract from it — a gap the company will have to close as it pushes beyond GPU-companion CPUs into the standalone server market.

What's inside Olympus?

Unlike Grace, which uses off-the-shelf Arm Neoverse V2 cores, Vera runs Nvidia's proprietary Olympus microarchitecture. Each core packs a wide front end with a 10-wide decoder, a neural branch predictor, a deep out-of-order execution engine, six vector execution units, four load pipelines, two store pipelines, large private caches and advanced prefetch mechanisms.

Nvidia deliberately chose 88 high-performance cores — 176 threads with its proprietary spatial multithreading — over maximizing core count per socket. Single-thread speed was a primary design goal.

"CPU design is really just a game of trade-offs," said Hannah Coutand, who leads Vera CPU marketing at Nvidia, in an interview with Tom's Hardware Premium. "You could maximize for a lot of cores, but literally when you maximize the die for lots of cores, you do not actually have space to maximize for IPC."

Nvidia positions Vera primarily for agentic AI workloads, though Coutand says Olympus was originally conceived with data processing in mind. Either way, the same three attributes matter: high per-core performance, high memory bandwidth and low memory access latency.

"We are very clear in our strategy with Vera and our positioning with Vera that it is not a general-purpose CPU," Coutand said. "If you want a web server for your cloud data center, there are probably other choices that are better suited for that than Vera."

How big is the optimization gap?

Nvidia's researchers benchmarked a double-precision general matrix multiplication (DGEMM) workload built on an assembly kernel from OpenBLAS, the open-source linear algebra library. OpenBLAS ships a kernel tuned for Neoverse V2, the cores inside Grace — but not for Olympus. Using the standalone miniKernelGEMM benchmark with a 128 × 128 matrix that fits in both processors' L1 caches, the numbers split cleanly:

  • Grace, running the Neoverse V2 kernel: 4.51 instructions per clock and 13.03 FP64 operations per cycle.
  • Vera, running that same unmodified kernel: 5.43 IPC and 15.69 FP64 operations per cycle — roughly 20% faster.
  • Vera, running a kernel rewritten for Olympus: 7.53 IPC and 21.75 FP64 operations per cycle — nearly 67% ahead of Grace on a per-core, per-clock basis.

The hardware explains part of the story. Neoverse V2 offers four 128-bit SIMD pipelines, good for 16 FP64 operations per clock; Olympus has six, reaching up to 24 — a 50% increase.

Utilization tells the rest. The Neoverse V2-optimized kernel uses 81.4% of Grace's peak FP64 throughput but only about 66.5% of Vera's. By adjusting matrix tile organization, register allocation and instruction scheduling to suit Olympus, Nvidia's researchers pushed utilization to 90.6%.

What happens next?

The SC26 results prove at least some applications will need Olympus-specific tuning to exploit Vera's execution resources fully, and it remains unclear how much can be gained in workloads beyond matrix multiplication. As Nvidia chases the standalone server CPU market, it inherits the optimization burden Intel and AMD have shouldered for decades. Its CUDA ecosystem experience, though, could shorten that road considerably.

Source: Tom's Hardware

Share this article:

More from Rebecca Stone

Rebecca Stone

Show full bio

Correspondent covering technology and product news in Semiconductors, chip manufacturing, hardware, AI compute at Chip Dispatch.

1,092 articles

Related articles

« Previous articleNext article »