What Do the Vera Benchmarks Actually Show?

According to Phoronix, which conducted the tests on a pre-production NVIDIA DGX Vera system, the Vera CPU achieved a SPEC CPU 2017 integer score of 12.4 — compared to 5.4 for Intel's Xeon Platinum 8592+ and 6.1 for AMD's EPYC 9654. In floating-point, Vera scored 14.1, versus 5.9 for Intel and 6.8 for AMD. The key metric, however, is all-core performance: Vera sustained 95% of its peak performance when all 72 cores were active, while Intel and AMD systems typically drop to 60-70% under similar loads due to thermal and power constraints. NVIDIA's blog post attributes this to Vera's new Grace architecture, which uses a unified memory fabric and a novel power delivery system.

Why Does All-Core Performance Matter for AI Factories?

The shift to agentic AI — where AI agents orchestrate complex, multi-step workflows — creates a new CPU requirement. According to NVIDIA's blog, agentic AI demands "fast cores, massive memory bandwidth and the ability to sustain high performance when all cores are active." This is because agents must coordinate GPU clusters, manage data pipelines, and run inference scheduling in real time. A CPU that throttles under load becomes the bottleneck. Phoronix noted that Vera's all-core performance is "unprecedented" for a server CPU, calling it "packing a heavy-hitting punch." However, the benchmarks are limited to SPEC CPU 2017, which tests raw compute, not AI-specific tasks like GPU orchestration or memory-bound inference scheduling.

How Does Vera Compare to Intel and AMD?

What Are the Limitations of These Benchmarks?

First, the tests are synthetic. SPEC CPU 2017 does not measure how a CPU performs when orchestrating GPU clusters, managing memory for large language models, or running inference scheduling — the actual workloads of an AI factory. Second, the system was pre-production hardware; shipping units may differ. Third, power consumption data was not disclosed. According to Phoronix, "NVIDIA did not provide power measurements for this preview," making it impossible to assess efficiency. Finally, the benchmarks compare Vera to Intel and AMD CPUs from 2023-2024; Intel's upcoming Granite Rapids and AMD's Turin may narrow the gap. Until real-world AI pipeline benchmarks are published, the evidence supports only a synthetic advantage.

Who Gains and Who Loses If Vera Delivers in Production?

If Vera's performance translates to real AI workloads, NVIDIA gains a moat in the AI factory market: customers who buy Vera CPUs will likely pair them with NVIDIA GPUs, creating a lock-in effect. Intel and AMD lose their last stronghold in the data center — the CPU — as NVIDIA expands beyond GPUs. Cloud providers like AWS and Azure, which currently offer Intel and AMD instances, may face pressure to offer Vera-based instances or risk losing AI workloads. However, the biggest loser could be AMD, which has positioned EPYC as the AI data center CPU but cannot match Vera's all-core retention. According to an analyst note from Moor Insights & Strategy, "If Vera delivers on these benchmarks, AMD's EPYC loses its primary value proposition for AI workloads."

1. NVIDIA will release production benchmarks for Vera in Q4 2026 showing 40-50% improvement in end-to-end inference throughput for agentic AI workflows.
2. Intel will announce a partnership with a major cloud provider (likely AWS) in Q3 2026 to develop a custom AI CPU, attempting to counter Vera.
3. AMD will lose at least 10% of its data center CPU market share to NVIDIA by Q2 2027 if Vera's production performance matches benchmarks.

• Vera's all-core performance retention is the metric that matters most for AI factories, not raw single-thread speed.
• The benchmarks are synthetic; real AI pipeline performance remains unproven and is the critical unknown.
• NVIDIA is using CPU benchmarks to extend its moat from GPUs to the entire AI infrastructure stack.
• Intel and AMD have 18-24 months to respond, but architectural changes take 3-5 years to bring to market.
• The biggest risk to Vera is not competition but NVIDIA's own execution: shipping a production CPU that matches benchmarks.