Newsletter Subscribe
Enter your email address below and subscribe to our newsletter
Enter your email address below and subscribe to our newsletter

NVIDIA’s Vera CPU has posted its first independent benchmarks, with testing by Phoronix showing the 88-core Arm processor delivering what the publication describes as the best performance ever seen on an Arm chip — including a 10% advantage over AMD’s high-frequency EPYC 9575F on a geometric mean basis.reddit
The results mark a turning point for NVIDIA’s ambitions beyond GPUs, positioning the company as a direct competitor to AMD and Intel in the data center CPU market just months before the chip enters broad commercial availability in the second half of 2026.
The Vera CPU integrates 88 custom “Olympus” cores built on the Armv9.2-A architecture, supporting 176 threads via NVIDIA’s Spatial Multithreading technology. It replaces the off-the-shelf Arm Neoverse V2 cores used in its predecessor, the Grace CPU, with a ground-up custom design that NVIDIA claims delivers roughly double Grace’s performance.moorinsightsstrategy
That generational leap comes from multiple sources: a 46% improvement from the Olympus microarchitecture itself — including higher vector throughput and larger caches — plus another 27% from the increased core count. The chip pairs with LPDDR5X memory via SOCAMM2 modules delivering up to 1.2 TB/s of bandwidth, roughly three times the per-core memory bandwidth of competing x86 processors.computerbase
Third-party testing by Redpanda, which evaluated the chip on Kafka-compatible streaming workloads, found Vera delivered 5.5 times lower latency than AMD EPYC Turin and 2.5 times faster performance than Intel Xeon 6 Granite Rapids, along with 73% higher throughput in ring-shuffle SQL tests.redpanda
The benchmarks arrive as NVIDIA moves to build a complete AI infrastructure stack. The company has already shipped initial Vera units to Anthropic and OpenAI, with Oracle Cloud Infrastructure confirmed as the first hyperscaler deploying at scale. CoreWeave is set to offer standalone Vera CPU access, with broader availability across AWS, Google Cloud, and Azure expected later this year.byteiota
NVIDIA CFO Colette Kress has indicated the company expects nearly $20 billion in CPU-related revenue for 2026, encompassing standalone Vera servers and processors integrated within its Vera Rubin and Grace Blackwell platforms.reddit
Analysts at GF Securities expect NVIDIA to further showcase Vera’s advantages at Computex 2026, predicting the company will demonstrate 1.5 times faster speeds, twice the overall performance, and four times the rack density compared to x86 alternatives.wccftech