AMD claims EPYC lead in agentic Artificial Intelligence workloads

AMD is using rack-level benchmarks to argue EPYC CPUs will remain central to agentic Artificial Intelligence infrastructure. The claims target Nvidia’s Vera platform and Intel’s Xeon lineup as data centers rebalance around CPU-heavy orchestration work.

AMD released new server CPU benchmarks positioning its EPYC processors as a key platform for agentic Artificial Intelligence workloads, where software agents coordinate databases, web services and orchestration tools across constrained data center racks. The company argues that rack-level throughput, power use, cooling demands and physical density matter more than isolated chip-level performance as infrastructure shifts toward agent-driven systems.

In a modeled 100-kilowatt rack scenario, AMD’s current-generation EPYC 9965 processor delivered roughly 2.37 times the rack-level throughput of Nvidia’s forthcoming Vera CPU platform. The lead widens with AMD’s next-generation EPYC Venice chip, built on the Zen 6 architecture, which is projected to stretch that advantage to 3.3 times over Vera. Against Intel’s Xeon 6980P Granite Rapids-AP processor, the EPYC 9965 turned in a 1.6x throughput lead. Even Intel’s current silicon showed a 1.46x advantage over Nvidia Vera in the same set of benchmarks.

AMD’s test suite focused on the types of workloads that support agentic Artificial Intelligence, including SPEC CPU 2017, a SPECjbb2015-derived Java workload, NGINX and WRK web-serving tests, redis-benchmark, Memcached and memtier_benchmark, and a TPC-C-derived OLTP proxy on MySQL. Rack density was another focus: current EPYC Turin systems can pack more than 27,000 CPU cores into a single rack, while the upcoming Venice generation is projected to exceed 36,000 cores. Nvidia’s Vera-based racks, by comparison, top out around 22,500 cores.

Single-threaded performance also tilts in AMD’s favor according to the company’s projections. A 64-core Venice part is expected to deliver a 27% per-core performance advantage over Nvidia’s 88-core Vera configuration. Even the higher-core-count 96-core Venice models are projected to hold an 11% single-core lead.

The benchmarks arrive during a pullback in semiconductor stocks. AMD shares fell 3.92% on Wednesday, caught in a broader tech sell-off that has trimmed the stock by roughly 9.3% over the past five trading sessions. Despite the pullback, AMD remains up more than 113% year-to-date and has surged roughly 293% over the past twelve months. AMD management has pegged agentic Artificial Intelligence infrastructure as a $120bn market opportunity over the next several years, while Nvidia has suggested it could reach $200bn.

60

Impact Score

Hades variant affects 23 PyPI package versions

The Mini Shai-Hulud Hades variant is targeting PyPI packages tied to bioinformatics and Artificial Intelligence themes. Socket researchers say the malware uses Python startup hooks and compiled extensions to run a JavaScript stealer.

DiffusionGemma rethinks text generation with diffusion

DiffusionGemma applies diffusion-style denoising to text, trading autoregressive token-by-token decoding for iterative canvas refinement. Its design combines encoder guidance, bidirectional denoising, scheduling, and entropy-based sampling.

NVIDIA shows RTX Spark platform at Computex 2026

NVIDIA presented RTX Spark in Taipei as a Windows on Arm platform spanning gaming, creator, and Artificial Intelligence workloads. Microsoft also detailed Windows 11 optimizations built specifically for the new NVIDIA silicon.

AWS enterprise processor targets Artificial Intelligence inference

AWS’s Annapurna Labs-designed enterprise server processor uses a chiplet architecture for cloud infrastructure and Artificial Intelligence inferencing. The design combines Arm compute resources, cache coherency, and high-bandwidth interconnects for AWS deployments.

Contact Us

Got questions? Use the form to contact us.

Contact Form

Clicking next sends a verification code to your email. After verifying, you can enter your message.