NVIDIA Vera CPU with Olympus cores delivers unprecedented ARM performance against Intel and AMD

NVIDIA Vera CPU Emerges as Serious Challenger to x86 Dominance

NVIDIA has released its first independent benchmark results for the NVIDIA Vera CPU, revealing performance that positions the ARM-based CPU as formidable competition to established x86_64 processors from Intel and AMD 1

1

. The chip, now in full production, features 88 custom Olympus cores designed specifically for agentic AI workloads and represents NVIDIA's most ambitious entry into the standalone CPU market 2

2

.

Phoronix testing revealed the NVIDIA Vera CPU delivered a 63% performance improvement over NVIDIA's Grace CPU, its predecessor that relied on off-the-shelf Arm Neoverse-V2 cores 5

5

. More significantly, the processor outperformed AMD EPYC 9575F—a 64-core Zen 5 chip running at 5 GHz—by 10% and beat Intel Xeon 6980P, a 128-core Granite Rapids processor, by 55% 3

3

.

Early Adopters Signal Market Confidence in NVIDIA's New In-House Processor

During his Computex keynote in Taipei, NVIDIA CEO Jensen Huang named Anthropic, OpenAI, SpaceX, and Oracle Cloud Infrastructure among the first major customers adopting Vera 2

2

. These AI labs represent the compute-intensive workloads that define current data center buildouts, making their early commitment a significant validation of NVIDIA's data center CPU strategy 4

4

.

James Bradbury, head of compute at Anthropic, stated: "Scaling compute is an important accelerant for the growth of models. We're excited to see Vera emerge as a promising part of the ecosystem when solving for agentic workloads" . Oracle Cloud Infrastructure will be the first hyperscaler to deploy Vera at scale, with broader availability across major clouds expected in the second half of 2026 2

2

.

Olympus Cores Deliver Custom Architecture for AI Agents

Unlike Grace, which used standard Arm Neoverse cores, Vera features NVIDIA's custom-designed Olympus cores compatible with the Armv9.2 instruction set architecture 1

1

. The 88-core design supports 176 threads via spatial multi-threading and includes double the L2 cache of Grace at 2MB per core, plus a 164MB unified L3 cache 1

1

.

The processor pairs these Olympus cores with LPDDR5X memory delivering up to 1.2TB/s of memory bandwidth while consuming less than 30 watts of memory power, compared to over 100 watts for traditional DDR5 systems 3

3

. In STREAM TRIAD testing, Vera sustained 90% of its peak memory bandwidth and delivered over 4x the memory bandwidth per core compared to traditional x86 CPUs 3

3

.

Targeting Agentic AI Workloads Where CPUs Matter Most

NVIDIA positioned Vera explicitly for agentic AI workloads—tasks where AI agents plan and execute actions rather than simply answering prompts 2

2

. These workloads involve branch-heavy runtimes, sandboxed code execution, data processing, and orchestration that stress CPU capabilities 3

3

.

The benchmarks focused on workloads relevant to AI factories: code compilation, file compression, video transcoding, Python and Java execution, and database management 1

1

. In Linux kernel compilation, a single-socket Vera completed the task in just 20 seconds—the fastest result Phoronix measured—delivering 2x faster compilation per core compared to a 128-core competitor 3

3

.

Michael Larabel, founder of Phoronix, concluded: "NVIDIA's Vera CPU with its in-house-designed Olympus CPU cores ends up packing a heavy-hitting punch with competitiveness to Intel/AMD x86_64 CPUs that I have never seen out of any other ARM or non-x86_64 processors" 1

1

.

Linux Support and Ecosystem Readiness Ahead of Launch

Vera arrives with upstream Linux kernel support already in place. Linux 7.1+ includes key driver support, and the processor works with ARM64 server distributions like Ubuntu and Fedora 1

1

. NVIDIA will continue providing Base OS, its modified Ubuntu version with Vera-specific patches 1

1

.

Compiler support arrived early, with GCC 16.1+ and LLVM Clang 21+ adding Olympus optimization in 2025—months before the hardware ramp 1

1

. This proactive upstreaming mirrors Intel's historical approach and contrasts with AMD's Zen 6 compiler support, which arrived later in the development cycle 1

1

.

Production Ramp and Competitive Landscape

Vera is now in full production, with first units hand-delivered in May 2025 before the public Computex announcement 2

2

. System manufacturers building standalone Vera CPU systems include Dell Technologies, HPE, Lenovo, and Supermicro, along with ASUS, Foxconn, GIGABYTE, and others 4

4

.

The tested pre-production unit ran at a 450-watt socket TDP, though power consumption and frequency monitoring were restricted during this initial benchmarking phase 1

1

. Performance per watt metrics remain to be demonstrated as production units ship in enclosed server chassis designs 5

5

.

Vera faces competition from AMD's next-generation EPYC Venice based on Zen 6 architecture, already in mass production for a second-half 2026 release, plus Intel's Diamond Rapids platform and emerging data center offerings from Qualcomm and Arm 5

5

. The processor also serves as the host CPU for the Vera Rubin platform, which NVIDIA has called its largest product launch 2

2

.

NVIDIA's entry into standalone CPUs represents a strategic shift. With Grace CPUs reaching nearly 2.5 million shipments to date, Vera aims to capture growing demand for processors optimized for inference and AI agents rather than traditional server workloads 4

4

. Whether NVIDIA can compete with Intel and AMD on volume shipments, pricing, and sustained performance will determine if these early benchmarks translate into market share gains when broader deployment begins later in 2026.