AMD launched its Epyc 9006 server processors to target enterprise data centers running agentic artificial intelligence workloads outside public clouds.
The market is shifting from cloud-based model training toward executing tools, application programming interface calls, and database queries directly on server processors, according to Dan McNamara, senior vice president and general manager of compute and enterprise AI at AMD.
“The shift in the market to agentic AI means moving workloads from model training on cloud GPUs to executing tools, API calls, and database queries on server CPUs,” McNamara said during an opening address at the “AMD in Depth” conference in Las Vegas.
IDC figures show the number of simultaneously running agents worldwide will reach 2.5 billion by 2030, which marks an eightyfold increase from current levels, McNamara told attendees.
Zen 6 Core Density
The Epyc 9006 processor family, code-named “Venice,” succeeds the Epyc 9005 “Turin” line by introducing Zen 6 and Zen 6c cores that scale up to 256 cores per socket.
AMD maintains hyperthreading across all versions, allowing systems to reach 512 threads per socket, whereas competitor Intel disabled hyperthreading on its Xeon 6+ chips. Supermicro launched its H15 server range using the sixth-generation processors, claiming up to 33 percent more cores and 1.8 times the token throughput per second on demanding models.
“Running across server energy consumption, this means we are executing practically three times more agents per watt than our competitor,” McNamara said.
The top-tier 256-core and 512-thread version draws 600 watts of power. By comparison, Intel’s Xeon 6+ chip with 288 cores and 288 threads consumes 450 watts.
Performance figures yield an energy efficiency ratio of 0.85 watts per thread for the Epyc 9006, contrasted against 1.56 watts per thread for the Xeon 6+.
6th Gen AMD EPYC SP7 processors are now available. Built on the new “Zen 6” architecture, these CPUs provide the performance and flexibility to support growing agentic AI deployments as well as demanding enterprise, cloud and HPC workloads. Systems are shipping from server… pic.twitter.com/482aI7VsnY
Zen 6 and Zen 6c cores share identical features but use different manufacturing nodes, with Zen 6c relying on a more advanced production line. Each circuit packs either 12 Zen 6 cores with 48 megabytes of shared L3 cache or 32 Zen 6c cores sharing 128 megabytes of L3 cache.
Higher transistor density limits Zen 6c circuits to a clock speed of 4.1 gigahertz, while Zen 6 circuits reach 5.3 gigahertz.
AMD Splits Epyc 9006 into Four Commercial Versions
AMD splits the Epyc 9006 lineup into four distinct commercial versions known as SP7, SP7-X, SP8, and Verano. MSI prepares server platforms utilizing both SP7 and SP8 architectures, with initial SP7 system deliveries slated for the fourth quarter of 2026.
Enterprise servers utilize the SP8 version, featuring 128 PCIe 6.0 lanes and a power draw capped at 400 watts.
“With the H15 portfolio now available, Supermicro and AMD offer clients a faster path to agentic AI infrastructure,” said Michael McNerney, senior vice president of marketing at Supermicro.
The Verano version arrives later in 2027 to address hardware sizing limits by housing six core circuits alongside two empty slots reserved for accelerators such as GPUs or network controllers.
Helion Reference Rack Designs Standardize Hardware Deployments
AMD introduces reference rack designs called Helion to standardize hardware deployments for enterprise and cloud operators.
