Everyone talks about GPUs when they talk about AI. AMD just published a detailed argument for why the CPU sitting next to that GPU matters just as much — and why their chip wins that argument decisively.
AMD has fired a carefully aimed shot at NVIDIA with a new technical blog making the case that agentic AI infrastructure requires a fundamental rethink — and that when you evaluate AI servers the way enterprises actually deploy them (at the rack level, under real power constraints), AMD EPYC delivers higher deployable CPU throughput, x86 software continuity and a standards-based path to dense, AI-supporting infrastructure — available today on shipping platforms.
The headline claim: AMD EPYC 9965 (“Turin,” 192-core) delivers a 2.37x normalised geometric mean advantage over NVIDIA Vera (88-core “Olympus”), with Intel Xeon 6980P (“Granite Rapids-AP,” 128-core) turning in 1.46x over NVIDIA Vera. And that’s before AMD’s next-generation Venice chip arrives.
Table of Contents
The Benchmark Numbers at a Glance
| Processor | Cores | Rack-Level Throughput vs. NVIDIA Vera |
|---|---|---|
| AMD EPYC 9965 “Turin” | 192 cores | 2.37x (current, shipping) |
| Intel Xeon 6980P “Granite Rapids” | 128 cores | 1.46x (current, shipping) |
| NVIDIA Vera “Olympus” | 88 cores | 1.0x (baseline) |
| AMD EPYC “Venice” (next-gen) | 256 cores | 3.30x (projected) |
All results based on modelled 100kW rack configurations using publicly available and internal benchmark data. AMD notes these are intended to provide directional comparison rather than direct measured rack benchmarks.
Why the Rack Level Is the Right Way to Measure
This is the central argument in AMD’s blog — and it’s a more sophisticated point than a typical benchmark press release.
Customers do not deploy benchmark headlines; they deploy racks constrained by power, cooling, floor space, software compatibility and operational readiness. A chip that looks impressive in isolated single-socket tests can look very different when you ask how many cores you can actually deploy within a 100kW rack power envelope.
AMD offers over 27,000 cores per rack with current Turin configurations and is projected to offer over 36,000 cores with its next-generation Venice lineup. NVIDIA’s Vera CPUs pack approximately 22,500 cores in rack-scale solutions — giving AMD the density and value advantage, offering 2.18–2.90x cores per socket at 1.18–1.41x system power.
The core density gap is real. But AMD also addresses single-threaded performance — the metric density-focused comparisons often ignore. AMD is expected to deliver a 27% performance-per-core advantage with its 64-core Venice CPUs against NVIDIA’s Vera 88-core configurations. Even the high-core-count 96-core Venice models are projected to drive single-core performance gains of 11% over Vera.
The Agentic AI Shift: Why CPUs Suddenly Matter Again
AMD’s deeper argument — beyond the benchmarks — is about the architecture of agentic AI itself, and why the CPU tier has become strategically critical in a way it wasn’t during the first wave of generative AI.
Generative AI’s first wave was built around a fairly simple pattern. The AI system of choice for the next few years will not be a single “AI box.” It will look more like a distributed system — GPU racks for dense model compute, fast networking, and agentic CPU racks for orchestration, processing data, and tool execution. At this point, a balanced architecture will matter more than ever. If the CPU tier is undersized, GPUs wait. If the orchestration layer is not designed for concurrency, cost and complexity rise.
In practical terms, every agentic AI deployment — whether it’s a Claude Code agent writing software, a financial analysis agent processing documents, or a drug discovery agent running protein simulations — generates a cascade of CPU-bound work alongside the GPU inference. Each agent needs multiple CPUs to handle all the tasks the agent creates: data pre- and post-processing to maximise GPU efficiency, hosting the agent framework and coordinating all tasks, and task execution across standard enterprise platforms such as databases, storage, compute, and search spawned by multiple agents.
For more on how agentic AI is reshaping the entire technology infrastructure stack, read our AI and technology coverage on TechnoSports and our breakdown of Anthropic’s recursive self-improvement report — which quantifies exactly how much CPU-bound agentic work is already happening at the frontier.
The x86 Continuity Argument: AMD’s Most Practical Card
Beyond raw performance, AMD’s blog leans heavily on a point that enterprise IT departments will find compelling: this is infrastructure customers can build today on standard x86 platforms, not a future architecture.
NVIDIA’s Vera is an ARM-based CPU — a capable chip, but one that requires enterprises to commit to a different software ecosystem, potentially recompile workloads, and accept dependency on a vendor whose primary business is GPU sales and who now competes directly in the CPU space.
Most enterprises can already take full advantage of the x86 software ecosystem to add agents, and there is no reason to move off of it to use closed proprietary architectures and thus be at the mercy of vendors that control them.
That is a pointed argument — and one that AMD knows resonates deeply with the enterprise procurement teams who have spent decades building x86-native infrastructure.
The Methodology Caveat: What AMD Isn’t Saying Loudly
To its credit, AMD is transparent about its methodology — though the caveats deserve attention.
AMD’s Venice results are derived from an estimated 1.7x scaling factor over the EPYC 9965, along with internal testing. AMD’s methodology ends with: “These results are intended to provide directional comparison rather than direct measured rack benchmarks.”
Independent analysts have noted that you cannot simply scale the performance of a single node up in a linear fashion — interconnects, as well as thermal and power limitations, will become a factor as you scale up.
The Turin numbers are on shipping hardware and are more reliable. The Venice projections are AMD’s own estimates based on scaling factors. Both are worth noting separately when evaluating the claims.
The Business Context: AMD’s Data Centre Business Is Exploding
This benchmark blog doesn’t exist in a vacuum. AMD posted record revenue of $10.3 billion in Q1 2026 — a 38% increase versus the previous year — driven mainly by strong demand in the data centre and AI segments. Data centre segment revenue was $5.8 billion, up 57% year-over-year, driven by strong demand for AMD EPYC processors and the continued ramp of AMD Instinct GPU shipments.
The agentic AI CPU story is not just a marketing positioning exercise — it is the narrative behind AMD’s fastest-growing business segment, and the company is clearly investing in making the technical case alongside the commercial momentum.
For the latest AMD product launches and semiconductor industry news, follow our technology coverage on TechnoSports.
Sources: AMD Official Blog | Tom’s Hardware | WCCFTech | All benchmark results based on AMD’s modelled 100kW rack configurations.
