AMD Venice EPYC Targets 1.9x Throughput for Agentic AI

By Grace Kim

AMD says its sixth-generation EPYC 9006 “Venice” processors deliver 1.9 times the throughput of Turin EPYC 9005. The gain supports a larger CPU role in agentic AI.

What does agentic AI change in the datacenter?

The EPYC 9006 series carries the “Venice” name and uses AMD’s sixth-generation Zen microarchitecture. One cited configuration reaches 5.3 GHz with 96 cores. Turin’s cited configuration reached 5.0 GHz with 64 cores.

AMD says memory bandwidth reaches 1.6 TB/s through 16 DDR5 channels. Turin offered 614 GB/s through 12 channels. I/O bandwidth doubles as the interface moves from PCI Express Gen 5 to Gen 6.

Generative AI’s reasoning and mixture-of-experts model waves favored GPUs and pushed CPUs into supporting roles, but agentic systems require a different balance across accelerator and general-purpose compute.

Dan McNamara, AMD’s senior vice president and general manager for compute and enterprise AI, described the infrastructure shift in May.

“Instead of answering one prompt, an agent breaks a goal into steps, decides what to do next, calls multiple models, queries databases, connects with APIs, runs enterprise applications, checks permissions, retrieves memory, validates output and then loops back through the process again.”

He said agentic systems need CPU strength for orchestration, policy, security and execution. They also require host processors. Sandboxes must run code generated by AI models.

AMD argues CPUs handle these branching tasks better than GPUs. The current GPU-to-CPU ratio runs from 1:4 to 1:8. The company expects agentic deployments to bring it to 1:1.

That shift favors AMD, Intel and other CPU suppliers. It does not make AMD a challenger to Nvidia’s dominant GPU franchise. AMD’s data-center CPU share stands at about 46%. The company expects the mix change to support CPU revenue, although it offered no capacity or revenue forecast.

Some analyst firms forecast agentic AI to grow nearly 43% annually through 2034. The estimate is not an AMD capacity, shipment or revenue commitment.

ChatGPT’s late-2022 launch supercharged the GenAI era. Nvidia’s valuation has soared into the trillions during successive generative-AI waves.

The midterm election result, characterized as a Democratic “Blue Wave,” could bring more U.S. AI regulation. A recent scandal involving Nvidia co-founder and CEO Jensen Huang creates another uncertain variable. Their effects on the datacenter buildout remain unclear.

How does Venice compare with competing processors?

AMD traces its server-CPU expansion to first-generation “Naples” EPYC in 2018. Fourth-generation “Genoa” arrived in 2022. That launch opened a broader roadmap of optimized Turin, Bergamo and Siena designs.

Madhu Rangarajan, AMD corporate vice president for compute and enterprise AI products, described a portfolio of more than 1,000 “instances.” He said that breadth gives customers a consistent offering across different deployment requirements.

Mike Clark, AMD corporate fellow, senior vice president and lead Zen architect, said the company did not foresee agentic AI’s rise. He long expected the Core Complex Die design to play a critical role. AMD invested heavily in Zen 5 and continued that work in Zen 6. Clark said the investment now pays off.

AMD reports competitive throughput and per-core performance against Nvidia’s Vera, Intel’s sixth-generation Xeon 6980P and Arm’s AGI processors. These figures come from AMD’s own comparisons.

Rangarajan called concurrent-agent capacity an important measure. In a separate 256-core, 400-watt Venice configuration, AMD claims 3.3 times Vera’s agents per watt. The company claims 1.8 times the agents per dollar of Intel’s sixth-generation Xeon. It also claims 1.5 times as many agents per Helion AI rack as fit in a Vera rack. AMD reports better density and efficiency than Vera.

Against fifth-generation, 192-core EPYC 9965, AMD says Turin outperformed Intel’s 128-core Xeon 6980P. It also beat Amazon’s 192-core Graviton5. The 256-core EPYC 9996 extends the lead, according to AMD.

Rangarajan said agentic AI drives CPU demand across three networks beyond general-purpose servers:

• Search clusters for agent searches.
• Enterprise tools such as CRM and ERP, with more agents running EVM workloads.
• Database access for agent operations.

Unlike earlier prompt-and-answer systems, agent pipelines combine model calls, database access, application execution, permission checks and iterative validation, forcing servers to balance accelerator throughput with general-purpose compute.

What do AMD’s agent benchmarks establish?

Rangarajan said: “Ephemeral tools are the more interesting part of agentic AI because there is no well-defined benchmark there.” He described them as short-lived, isolated runtimes. Agents create them during a session. They are destroyed automatically.

AMD selected group graph, document parsing, coding agent and compilation workloads as representative. Rangarajan said results were strong at each stage of that pipeline. He warned that different agent workflows could combine stages differently.

Because agents can combine the selected stages in different proportions, AMD’s results indicate a favorable workload profile but cannot predict performance across every framework, application or production environment.

What comes next?

“Verano” and “Venice-X” processors are due next year. If agentic deployments pull the GPU-to-CPU mix from 1:4–1:8 toward the expected 1:1, Venice can strengthen AMD’s position against Intel. The follow-on processors will test whether that advantage holds across real-world workload mixes.