Nvidia Vera CPU: Full Specs, Benchmark Results, and the Rosa Follow-Up

Nvidia's CPU ambitions have moved beyond the Grace chip. The Vera CPU, now in full production, is a purpose-built data-center processor for the agentic AI era. Based on early independent benchmarks and Nvidia's own published data, it appears to offer generational leaps over both the previous Grace CPU and competing x86 architectures from Intel and AMD. This article covers the verified technical specifications, the first benchmark results from Phoronix and Redpanda, analyst commentary from Wedbush and Citi, and everything officially known about the follow-up Rosa CPU and the Feynman platform.

What Wedbush and Citi Said

Wedbush Securities analyst Matt Bryson sees Vera as a significant growth lever beyond Nvidia's GPU business. Bryson highlighted that Vera's high core count could help it outperform x86-based CPUs and noted that Nvidia pitching Vera CPUs to Chinese customers is a "sensible" move given tight CPU supply . In a January 2026 note, Wedbush described the Vera Rubin platform as a "10x leap toward the era of agentic AI" and confirmed it had entered full production .

Citi analysts expressed surprise at the early unveiling of the Vera Rubin platform at CES 2026. They stated there is upside to $500 billion in AI demand from the Blackwell/Rubin roadmap .

Vera CPU — Technical Specifications

Nvidia designed Vera from the ground up for AI factories. Key specifications from Nvidia's developer blog and detailed hardware analyses :

Specification	Detail
Cores	88 custom Nvidia Olympus cores
Threads	176 (via Nvidia Spatial Multithreading, physical resource partitioning)
ISA	Arm v9.2
Instruction Decode Width	10-wide
Branch Predictor	Neural; 2 taken branches per cycle
L2 Cache (per core)	2 MB
L3 Cache (total)	164 MB
Memory	LPDDR5X (SOCAMM2 modules), up to 1.2 TB/s bandwidth
Memory Power	<30 W (vs. >100 W for traditional DDR5)
Interconnect	PCIe Gen 6, CXL 3.1
TDP	450 W (single-socket pre-production system)
Process Node	TSMC 3nm (N3)
Die Design	Single monolithic die with all 88 cores, no NUMA eccentricities
FP8 Support	Native FP8 precision for AI workloads

Nvidia also states that Vera delivers up to 50% faster sandbox performance versus competitive platforms and 80% faster sandbox environment performance than traditional CPU infrastructure . CEO Jensen Huang highlighted the new Vera chip, targeting a $200 billion market .

Vera CPU — Benchmark Results

The first widely reported independent benchmarks came from Phoronix on May 26, 2026, testing a single-socket Vera system at Nvidia's Santa Clara headquarters . These were Nvidia-hosted, curated tests, meaning results may not fully generalize to all third-party workloads .

Key results from the Phoronix geometric mean:

• 1.55x faster than Intel's flagship Xeon 6980P (Granite Rapids)
• ~10% ahead of AMD's EPYC 9575F (64 Zen 5 cores)
• 1.6–1.63x faster than Nvidia's own prior-generation Grace CPU
• 90% STREAM TRIAD efficiency — sustained memory bandwidth utilization
• 20-second Linux kernel compilation time
• 1.5x overall performance vs 128-core x86 CPUs

Independent validation from Redpanda: In their own benchmarks comparing Vera against five other systems, Redpanda found Vera delivered the lowest streaming latencies across the board, the best interconnect scaling, and the fastest build times with up to 73% higher throughput than AMD EPYC “Turin” .

Caveat: As Tom's Hardware noted, these were curated tests at Nvidia's Santa Clara facility, so results may not generalize to all workloads .

Rosa CPU — What Is Known

Rosa is the confirmed follow-up to Vera, named after Nobel Prize-winning medical physicist Rosalyn Sussman Yalow. It belongs to the Feynman generation and is expected in 2028 .

Confirmed from Nvidia:

• Core architecture: New Rigel core, based on Arm v9.2 ISA
• Goal: Higher per-core performance than Vera's Olympus core while maintaining the same die area
• Key improvements: Better instruction delivery, larger L2 cache, more efficient memory handling
• Platform: Pairs with the Feynman GPU (next-gen data center GPU architecture) and connects with BlueField-5 DPU and ConnectX-10 SuperNIC
• Target workload: Optimized for agentic AI
• Cadence: Nvidia shortened the CPU development cycle from ~4 years to ~2 years

Rumored/Reported (not officially confirmed by Nvidia):

• Process node: May use TSMC N2 (2nm) or possibly skip straight to A16 (1.6nm-class) with Super Power Rail backside power delivery
• Consumer PC variant (Rosa Feynman Spark) reportedly planned for 2030

Summary: Rosa is a same-die-area performance-per-core upgrade over Vera's Olympus core, built on Arm v9.2 Rigel architecture, launching in 2028 as part of the Feynman platform. Process node details (2nm vs A16) are still speculative.