EPYC 9996 Posts 5,470 SPEC vs 1,459 Estimate for Nvidia Vera

By Marcus Bennett

A two-socket AMD EPYC 9996 system scored 5,470 on SPEC CPU2017. Nvidia's Vera CPU shows 1,459 in the same table — but that number is AMD's estimate, not a filed result. Intel's Xeon 6980P sits between them at 2,510. That 5,470 figure is the only top-line number in this three-way fight backed by a submitted SPEC result, a list price ($14,904) and shipping hardware.

AMD declared general availability for its 6th Gen EPYC 9006 Series on November 10, 2026, with OEM and ODM systems shipping by November 12. Nvidia's Vera reached general availability on Oracle Cloud Infrastructure on October 27, five days earlier. Intel's Xeon 6 remains the incumbent — and, as Intel's newsroom confirmed on November 13, the host CPU inside Nvidia's own DGX Rubin NVL8 systems. Nvidia sells an Arm CPU and still ships an x86 host in a flagship box.

What does the data actually support?

Nvidia has published neither a list price nor a SPEC score for Vera. The only Vera number in AMD's comparison is AMD's own estimate, and the earlier fight over Nvidia's 1.8x claim versus AMD's rebuttal already showed how far vendor-chosen workloads drift apart.

The arithmetic in AMD's table holds together. Dividing 5,470 by 2,510 gives 2.18x over the Xeon 6980P, matching AMD's "up to 2.2x" claim. Dividing by the 1,459 Vera estimate gives 3.75x, close to AMD's stated 3.7x. Per-dollar, 0.367 against 0.180 works out to roughly 2.0x. The ratios are internally consistent — but only against a table AMD built.

The per-core column tells a subtler story. Vera's estimated 8.29 per core sits between the Xeon 6980P's 9.80 and the 6990E+'s 5.30, which suggests Nvidia's Olympus cores are respectable individually. The gap to EPYC's 10.68 per core comes mainly from core count: 88 cores per socket for Vera against 256 for the 9996.

Why the Vera row is an estimate, not a result

AMD's footnotes expose the construction. For SPECrate 2026 integer, AMD compares a 2P EPYC (256C at 400 W, GCC 16.1, estimated 1,810) against a Vera ETL256 rack row (88C at 450 W TDP, LPDDR5X-9600, GCC 15.2, estimated 925 per 2P). Compilers differ by a full major version, and compiler choice moves SPEC integer scores. AMD, not Nvidia, produced the Vera numbers, so nobody can check them against a filed result.

Nvidia's own claims are workload-specific and name no x86 baseline: up to 1.8x faster single-threaded performance, 80% faster agentic task completion, and "twice the efficiency and 50% faster than traditional CPUs" for AI-factory workloads.

Specs that are actually published

• Vera CPU: 88 Arm-compatible Olympus cores per socket, 450 W TDP, up to 1.5 TB LPDDR5X, NVLink-C2C at up to 1.8 TB/s coherent to Rubin GPUs.
• EPYC 9996: 256 Zen 6 cores per socket, 512 threads, 600 W default power, DDR5 and MRDIMM-12800 at up to 1.6 TB/s per socket, PCIe Gen 6.
• Xeon 6980P: 128 P-cores per socket, 500 W, $13,955 as cited by AMD.

Vera's memory plays a specific agent-workload trick: while an agent waits for a tool to return, idle KV cache parks in CPU memory instead of occupying GPU memory. The coherent 1.8 TB/s link lets CPU memory act as an extension of the GPU pool. That is the architectural reason Nvidia built its own CPU.

Pricing: what a buyer can quote

AMD footnoted its prices "as of 11/05/2026" and described the $14,904 figure as a 1,000-unit price. The EPYC 9965 lists at $11,988; the Xeon 6990E+ at $14,995 and the 6980P at $13,955, both as cited by AMD. Nvidia published no price for the Vera CPU, the Vera CPU Rack, Vera Rubin NVL72 or HGX NVL8. A buyer evaluating Vera must request a quote, which alone pushes smaller teams toward chips they can price on a website.

Deployments and open questions

Vera's first systems went to Anthropic, OpenAI, Oracle Cloud Infrastructure and SpaceXAI, per Nvidia's product page. Meta said on October 27 it intends to expand Vera deployments. Noetra announced a physical-AI factory with 13,750 Vera CPUs and 27,500 Rubin GPUs. AWS is a collaboration, not a live instance. On the AMD side, Azure and Oracle 6th Gen instances are "coming soon," though 5th Gen EPYC already hosts Google Cloud TPU 8tx.

One corporate loose end: on November 12, Jensen Huang resigned as Nvidia's president, CEO and board member, effective immediately. Colette Kress holds the interim CEO role. No source links the change to Vera's schedule. Nvidia reports Q3 fiscal 2027 results on November 17 — the first chance to hear management address supply and demand.

Verdict: EPYC 9996 wins the comparison that can be completed. Take Vera when buying Rubin racks anyway; keep Xeon 6 where x86 compatibility rules. Until Nvidia files a SPEC result or publishes a price, 3.7x is AMD's best case and 1.8x is Nvidia's.