AMD has introduced its 6th Generation EPYC server processors, based on the Zen 6 architecture, with SP7-based systems now shipping from server manufacturers and cloud deployments planned by Microsoft Azure and Oracle Cloud Infrastructure. The new generation, code-named Venice, expands the server CPU portfolio to address agentic AI, GPU-host infrastructure, enterprise applications, cloud computing and high-performance computing. Meta is validating the platforms ahead of planned large-scale deployment.

The launch reflects the expanding role of CPUs in AI data centers, where processors orchestrate AI agents, manage GPU resources, execute code and support databases and services surrounding large language models. AMD is offering configurations emphasizing core density, frequency, memory bandwidth and power efficiency. The EPYC 9996 flagship provides 256 cores and 512 threads per socket at a 600W default CPU power rating. A two-socket EPYC 9996 system has a published SPECrate2017 integer score of 5,470 versus 2,510 for a two-socket Intel Xeon 6980P system. AMD also estimates up to 3.7 times NVIDIA Vera system performance; that comparison relies on AMD projections rather than a published Vera SPEC result.

AMD is extending EPYC beyond individual servers with Helion agentic AI racks. Helion Liquid targets high-density liquid-cooled deployments alongside Helios; Helion Air targets existing air-cooled facilities; and Helion Air Inference combines CPU execution and local GPU inference. AMD says certain configurations support up to 73,728 threads per rack. High-frequency EPYC configurations, reaching 5.3 GHz, target GPU-host workloads with AMD-projected LLM offload-throughput improvements associated with PCIe 6.0 bandwidth.

“AI is significantly expanding the role of the CPU across every layer of the data center,” said Dan McNamara, senior vice president and general manager, Compute and Enterprise AI, AMD. “With 6th Gen AMD EPYC, customers can standardize on the proven x86 foundation while choosing a workload-optimized solution to run more agents, keep accelerated infrastructure fed, and power the business-critical applications they depend on.”

Analysis: CPUs Become a Critical Layer of Agentic AI Infrastructure

Although GPUs and specialized AI processors execute computationally intensive portions of neural networks, CPUs remain responsible for application execution, model-serving infrastructure, data preparation, networking and orchestration. Agentic AI adds CPU-intensive code execution, tool invocation, database queries and isolated execution environments. High-frequency CPUs target latency-sensitive GPU-host operations; high-core-count processors target concurrent agents and virtualization.

The EPYC 9996 provides 512 hardware threads per socket, but actual agent throughput depends on memory, workload complexity and software. AMD’s theoretical agents-per-watt comparisons use thread resources as a proxy rather than directly measured application throughput. Its 35 kW rack analysis suggests that processor and memory power capping could increase theoretical agent density by up to 18%, but this estimate includes fractional server counts and extrapolated performance.

PCIe bandwidth is another consideration: CPU-memory model-weight offloading can be limited by the CPU-to-GPU link. AMD’s up-to-1.8× LLM throughput statement projects PCIe 6.0 results from PCIe 5.0 tests; it is not a measured production PCIe 6.0 improvement. Azure has announced HDv2 preview instances and other planned services, Oracle expects sixth-generation EPYC in OCI Compute E7, and Meta is validating platforms. AMD has not disclosed the eventual deployment volumes.