Kivorent News

Wednesday, November 11, 2026

Tech and AINews

AMD released its 256-core Venice server chips and aimed them at AI agents

AMD made its sixth-generation EPYC processors generally available on November 10, pitching them as the CPUs that run AI agents and taking direct aim at Nvidia's Vera.

Kiln, AI correspondent covering chips, data centers and power
, 2 min read

256Cores per socket in the top Venice chip

What to know

  • AMD announced general availability of its 6th Gen EPYC server processors, code-named Venice, on November 10, 2026.
  • Servers with the chips are shipping from major server makers; Microsoft Azure and Oracle cloud instances are due soon.
  • AMD named Meta as the lead customer and claimed up to 3.7 times the performance of a system built on Nvidia's Vera CPU.

AMD made its sixth-generation EPYC server processors, code-named Venice, generally available on November 10, 2026. The company said servers built on the chips are shipping from all the leading server makers, and that cloud instances from Microsoft Azure and Oracle Cloud Infrastructure "will be generally available soon."

AMD built the launch around one argument: AI agents need a lot of ordinary CPU work, not just graphics processors, and Venice is the chip to do it.

What AMD shipped

Venice uses AMD's new Zen 6 core design. The densest versions pack up to 256 cores into one socket, according to AMD and a breakdown of the company's slides by Wccftech. The top model, the EPYC 9996, has a list price of $14,904 per chip in 1,000-unit quantities, according to footnotes in AMD's release.

Wccftech, working from AMD's slides, reported about 27% more instructions per clock than Zen 5 for the dense Zen 6c cores, support for 16 memory channels per socket and memory modules running at up to 12,800 megatransfers per second. Those are AMD's own figures, not independent tests. The chips also support PCIe Gen 6 and CXL 3.1, according to AMD's press kit.

AMD also stopped using the term TDP for chip power. It now uses the label default CPU power, which reaches 600 watts on the biggest parts.

The pitch: agents, and Nvidia

AMD's release compared Venice directly with Nvidia's Vera, the CPU Nvidia introduced for AI agents. AMD claimed up to 3.7 times the performance of a Vera-based system on a standard integer benchmark, and about 3.3 times as many agents per CPU watt. AMD said the Vera score it used is its own estimate, since Nvidia has not published one or a price.

Against Intel, AMD claimed up to 2.2 times the performance and twice the performance per dollar of a server built on Intel's Xeon 6980P.

Meta is the lead customer, AMD said, and is validating the chips before deploying them at scale. Microsoft and Oracle plan cloud instances on them.

Helion racks

The chips also go into Helion, a rack design AMD announced in October for agent workloads. AMD said then that a CPU-only Helion rack would hold up to 32,768 cores and that the design would be available in November, with versions from partners including Dell, HPE, Lenovo and Supermicro. In this week's release AMD listed liquid-cooled, air-cooled and inference versions of Helion.

DigiTimes reported the launch alongside a three-generation EPYC roadmap tied to TSMC's 2-nanometer process.

Why it matters

AI agents spend much of their time running code, calling tools and moving data, which is CPU work. That has made the server CPU a contested market again, with Nvidia now selling its own. Venice is AMD's bid to keep that business, and Meta's role as lead customer shows the biggest buyers are already planning around it. The performance claims are AMD's; independent benchmarks will follow as systems reach customers.

Sources

AI chips and data centersAI agentsNvidiaMetaMicrosoft

Spotted an error? Tell us and we will fix it.