AMD says its newest server chip beat Intel’s flagship Xeon by more than double, and edged out a CPU built by a company that has never sold one before: Nvidia. On September 18, 2026, AMD published benchmark claims putting its sixth-generation EPYC “Venice” processor ahead of Intel’s Xeon 6980P by 2.37x and ahead of Nvidia’s 88-core Vera CPU by 2.24x on the SPEC CPU 2026 Integer Rate test. The numbers are AMD’s own, not independently verified, but they mark the first time a chipmaker has framed Nvidia’s CPU business as a direct rival worth beating in a public benchmark chart.

The claims landed two months after AMD launched Helios, its rack-scale AI platform, and one month after Nvidia declared its own Vera Rubin NVL72 system to be in full production. Both companies are racing to lock in data center customers before 2027 budgets get set, and the CPU fight inside those racks has become almost as loud as the GPU fight. For a market that used to be a two-horse race between AMD and Intel, Nvidia’s arrival as a CPU vendor changes the math for anyone buying server hardware this cycle.

What AMD Actually Claimed on September 18

AMD’s benchmark disclosure centered on two chips from its Venice generation. A 256-core EPYC 9996 posted a score AMD says was 2.37x higher than Intel’s 96-core Xeon 6980P and 2.24x higher than Nvidia’s 88-core Vera CPU, using the SPEC CPU 2026 Integer Rate suite maintained by the Standard Performance Evaluation Corporation. A separate, higher-clocked 96-core Venice part came in about 20% faster per core than the Vera chip in the same test family, according to AMD’s published figures.

Core counts matter here because they explain most of the gap. AMD’s 256-core part has nearly triple the cores of Nvidia’s 88-core Vera chip, so a rate-based benchmark that scales with parallel throughput favors AMD by design. The per-core comparison is the more interesting number: a 20% single-thread advantage over a brand-new Arm-based CPU from a company that has spent two decades building GPUs, not general-purpose processors, is a harder thing to wave away. Neither AMD nor Intel nor Nvidia has published a third-party audit of these results, and SPEC.org itself does not certify vendor self-submitted comparisons before they’re publicized.

EPYC Venice and Helios: AMD’s Full-Stack Pitch

Venice is the CPU half of a bigger platform AMD calls Helios, which it launched on July 23, 2026. Helios pairs 72 Instinct MI455X GPUs with 18 sixth-generation EPYC Venice processors, AMD’s Pensando networking silicon, and its ROCm software stack in a single rack-scale unit. AMD has already sold Helios hardware at volume: Oracle took delivery of 50,000 MI450-family GPUs under the platform earlier this year, a deal covered in detail after it closed.

The MI455X accelerator that anchors Helios carries 432GB of HBM4 memory and roughly 23.3 TB/s of bandwidth, per AMD’s own listings, with a claimed 20.1 petaflops of dense FP8 performance across a full 72-GPU rack. AMD says Helios shipments begin near the end of Q3 2026, with OpenAI expected to bring systems online in Q4. None of the MLPerf-style independent benchmark results for MI455X had been published as of this week, so those throughput figures remain AMD’s alone for now.

Nvidia’s Vera CPU: A New Front in an Old Fight

Nvidia’s Vera chip is the more unusual story here. Nvidia built its business on GPUs and has spent the last few years buying its way into networking, storage, and now general-purpose compute. Vera is an Arm-based CPU designed to sit alongside Nvidia’s Rubin GPUs in the same way AMD’s EPYC sits alongside Instinct, and Nvidia began production shipments of the combined Vera Rubin NVL72 system in August 2026. That system pairs 72 Rubin GPUs, each carrying 288GB of HBM4, with Vera CPUs in a single liquid-cooled rack, and Nvidia says it moved the platform to full production during September, a milestone detailed in an earlier report on the launch.

At Nvidia’s AI Infrastructure Summit on September 15, the company said Vera Rubin NVL72 could deliver up to 30 times the throughput per megawatt of its previous GB300 NVL72 platform. That figure describes the whole rack, GPUs and CPU and networking together, not the Vera chip in isolation, which is part of why AMD’s narrower CPU-to-CPU comparison three days later landed as pointed. Server buyer HPE has already leaned into the Vera platform: its stock hit a 52-week high earlier this year on enthusiasm for Nvidia CPU-based systems, a sign that Wall Street is already pricing in Nvidia as a legitimate third CPU vendor.

Intel’s Xeon 6980P Sits in the Middle

Intel is the company with the most to lose from this particular fight, since it’s the incumbent both AMD and Nvidia are measuring themselves against. Xeon 6980P is built on Intel’s 18A process and tops out at 288 cores in its highest-density configuration, a launch AMD itself was on the losing end of a few months earlier when Intel claimed a 30% lead over EPYC in a different workload, a story covered when it first surfaced. That earlier Intel claim and this week’s AMD claim aren’t directly comparable since they use different benchmark suites and configurations, but the back-and-forth shows how much vendor-published numbers now drive the public narrative around server CPU performance, ahead of any independent lab result.

Intel has not responded publicly to AMD’s September 18 figures as of this article’s publication. The company is in the middle of its own transition, having pushed its Nova Lake-S consumer line to a later Razor Lake refresh and raised prices roughly 10% ahead of a planned March 2027 increase, moves that suggest Intel is managing cost pressure across its lineup rather than racing to match every benchmark claim from rivals chip for chip.

ROCm 10: AMD’s Software Argument

Hardware claims only matter if the software stack can use them, and AMD used the same week to push its software story too. AMD shipped ROCm 10 on August 27, 2026, and says the new release delivers up to 3.3x higher inference throughput and 2.4x higher training throughput than ROCm 7 on identical hardware. That’s a big enough jump that, if it holds up in independent testing, it would matter more to buyers than the raw CPU benchmark gap, since most AI workloads are bottlenecked by software efficiency and memory bandwidth rather than integer core count alone.

Developers who want to check what ROCm version a Helios-based system is running can do it directly from the command line:

rocminfo | grep -i "ROCm Version"
rocm-smi --showdriverversion

ROCm’s push matters because software lock-in, not raw silicon, has historically been Nvidia’s biggest advantage over AMD. CUDA has a decade-long head start with developers, and AMD closing that gap on paper doesn’t mean production teams will migrate workloads overnight. Anyone auditing a rack for core count and topology on the Linux side can pull the raw layout with a standard utility:

lscpu | egrep 'Model name|Socket|Core|Thread'

Server CPU Spec Comparison, September 2026

The table below lines up the three chips at the center of this week’s benchmark dispute, using the specs each vendor has published.

ChipVendorMax CoresArchitectureClaimed SPEC INT Rate Result
EPYC 9996 (Venice)AMD256x86 (6th-gen EPYC)Baseline (2.37x Xeon, 2.24x Vera)
Xeon 6980PIntel96x86 (Intel 18A)2.37x behind AMD’s 9996 (per AMD)
Vera CPUNvidia88Arm2.24x behind AMD’s 9996 (per AMD)
Venice (96-core variant)AMD96x86 (6th-gen EPYC)~20% ahead of Vera per-core (per AMD)

Every figure in that table originates with AMD’s own September 18 disclosure. Intel and Nvidia have not published counter-benchmarks as of this week, and SPEC CPU 2026 Integer Rate results submitted by a single vendor are not the same thing as results verified by a neutral lab.

AI Accelerator Comparison: The Bigger Picture

The CPU fight doesn’t happen in isolation. It sits inside a much larger contest over rack-scale AI systems, where GPUs and memory bandwidth matter more to most buyers than integer throughput. Here’s how the current generation of accelerators compares on paper.

AcceleratorVendorMemoryBandwidthDense FP8 (claimed)
MI455X (Helios)AMD432GB HBM4~23.3 TB/s20.1 PFLOPS (72-GPU rack)
Rubin (Vera Rubin NVL72)Nvidia288GB HBM4/GPU~22 TB/s (system)Not disclosed per-rack
MI355XAMD288GB HBM3e8 TB/s5.03 PFLOPS (8-GPU HGX)
Blackwell B200Nvidia180GB HBM3e~8 TB/s4.5 PFLOPS (8-GPU HGX)
H200Nvidia141GB HBM3e4.8 TB/s2.0 PFLOPS

The pattern is consistent across both tables: every headline number comes from the vendor that built the chip. That’s normal for a launch week, but it means buyers evaluating these platforms for 2027 deployments are still working from marketing decks rather than lab reports. Memory has become the real constraint behind all of it, and pricing pressure from the ongoing HBM shortage, which has already pushed chip prices up sharply in China, is shaping how aggressively any of these vendors can scale production even if their benchmark claims hold up.

Why Nvidia Building CPUs Changes the Market

Nvidia entering the CPU market isn’t just a product story, it’s a margin story. Nvidia has historically sold GPUs at high margins while treating the CPU as someone else’s commodity part bolted onto the rack. Vera changes that: Nvidia now captures more of the bill of materials on every Vera Rubin system it ships, and it controls the CPU-to-GPU interconnect end to end instead of depending on AMD or Intel silicon to talk to its own chips efficiently.

That vertical integration is also why AMD picked this specific fight. Losing CPU market share to Intel is a familiar, survivable problem for AMD. Losing it to Nvidia, a company that already dominates the GPU side of every AI rack AMD is trying to sell into, is a different kind of threat. AMD’s stock has had a strong run this year, according to AMD’s investor relations filings, crossing the trillion-dollar market cap mark after a 9.6% single-day surge, and keeping that momentum depends on convincing data center buyers that AMD’s full stack, CPU and GPU and software together, beats buying everything from Nvidia in one box.

Historical Context: From Two-Horse Race to Three

For most of the last twenty years, the server CPU market was a duel between Intel and AMD, with Arm-based challengers like Ampere and cloud-native chips such as AWS Graviton picking off niche workloads at the margins. That started shifting when hyperscalers began designing their own silicon, and it shifted again when Arm-based server share hit a record 15.3% earlier this year, a milestone tracked in detail as it happened. Nvidia’s Vera chip is the most consequential Arm entrant yet, not because of its raw specs but because Nvidia has the GPU relationship and the balance sheet to force its CPU into deals that would otherwise have gone to AMD or Intel by default.

The Arm-versus-x86 market split is now worth roughly $53 billion to $34.6 billion in AI servers specifically, according to the estimates in that earlier coverage, and every chip discussed in this piece sits somewhere on that line: AMD and Intel defending x86, Nvidia pushing Arm from the top of the stack down.

What Cloud Providers and Enterprises Should Watch

Buyers evaluating 2027 capacity commitments now have to model three CPU vendors instead of two, and the software question matters as much as the hardware one. A workload tuned for CUDA and x86 doesn’t move to an Arm-based Vera system for free, and a workload tuned for ROCm on EPYC doesn’t move to Xeon for free either. Migration cost, not raw benchmark scores, will decide a lot of these purchasing decisions over the next year.

Procurement teams should also weigh how much of each claim is rack-level versus chip-level. Nvidia’s 30x throughput-per-megawatt figure describes an entire Vera Rubin NVL72 system, not the Vera CPU alone, while AMD’s 2.37x figure is a narrower, chip-to-chip integer benchmark. Comparing the two directly, without adjusting for scope, will produce a misleading picture of which platform actually costs less to run at scale.

Predictions for Late 2026 and 2027

  • Independent SPEC CPU 2026 results for EPYC Venice, Xeon 6980P, and Vera will start appearing from third-party labs within the next two to three months, and they’re likely to narrow AMD’s claimed gap somewhat once workload tuning differences are accounted for.
  • Intel will publish a counter-benchmark before the end of 2026 rather than let AMD’s numbers stand unanswered, given how directly the September 18 claims targeted Xeon 6980P.
  • Nvidia will keep pricing Vera as part of a bundled rack deal rather than selling it as a standalone CPU, making apples-to-apples chip comparisons harder for buyers through at least 2027.
  • AMD’s MI450-generation shipments, including the Oracle deal, will expand to at least one more hyperscaler by early 2027 as Helios production ramps past its Q3 2026 start.
  • Software portability, specifically how fast ROCm 10’s claimed gains hold up under real production workloads, will matter more to enterprise buyers by mid-2027 than any single integer benchmark published this month.

The Caveat That Applies to All of This

Every number in this article traces back to the vendor that built the chip. AMD published its own Xeon and Vera comparisons. Nvidia published its own throughput-per-megawatt claim for Vera Rubin NVL72. Intel has published its own core-count and process-node figures for Xeon 6980P in the past. None of these companies submitted results to a neutral third-party lab before making them public, and SPEC.org’s published methodology doesn’t require that step before a vendor discloses a number. Treat every figure here as a marketing claim until an independent benchmark house, not the chipmaker itself, reproduces it.

Frequently Asked Questions

What is AMD EPYC Venice?

EPYC Venice is AMD’s sixth-generation EPYC server CPU family, built to pair with AMD’s Instinct MI455X GPUs inside the Helios rack-scale platform that launched July 23, 2026.

Does Nvidia actually sell a CPU now?

Yes. Nvidia’s Vera chip is an Arm-based CPU built to sit alongside Rubin GPUs in the Vera Rubin NVL72 system, which Nvidia moved to full production in September 2026.

Are AMD’s benchmark claims against Xeon and Vera independently verified?

No. As of this article’s publication, the 2.37x and 2.24x figures come solely from AMD’s own SPEC CPU 2026 Integer Rate submissions. Neither Intel nor Nvidia has published a rebuttal, and no independent lab has confirmed the results.

What is ROCm 10 and why does it matter?

ROCm 10 is AMD’s latest GPU software stack, released August 27, 2026. AMD says it delivers up to 3.3x higher inference throughput and 2.4x higher training throughput than ROCm 7 on the same hardware, a claim aimed directly at closing AMD’s long-standing software gap with Nvidia’s CUDA ecosystem.

How does the MI455X compare to Nvidia’s Rubin GPU?

Both are HBM4-based accelerators aimed at the same rack-scale AI market. MI455X carries 432GB of HBM4 per GPU at roughly 23.3 TB/s, while Rubin carries 288GB per GPU in the Vera Rubin NVL72 system. Neither company has published independently verified head-to-head throughput numbers yet.

When will Helios and Vera Rubin systems actually ship to customers?

AMD says Helios shipments begin near the end of Q3 2026, with OpenAI expected to bring systems online in Q4 2026. Nvidia says Vera Rubin NVL72 entered full production in September 2026, with cloud providers first in line for allocation.

Why does core count matter so much in this comparison?

SPEC CPU 2026 Integer Rate scales with parallel throughput, so a 256-core chip has a built-in advantage over an 88-core or 96-core rival on that specific test. AMD’s more telling number is its roughly 20% per-core lead over Vera on a similarly sized 96-core Venice part, which controls for core count more directly.