On November 10, 2026, AMD announced general availability of its 6th Gen EPYC SP7 products, built on the Zen 6 architecture. Systems based on the processors have also begun shipping from major server vendors. As AI agents take on a growing range of tasks, from calling models to executing code, AMD is positioning the new EPYC generation for three uses: running general-purpose services, serving as the host CPU that supports GPUs, and powering isolated execution environments that safely run generated code and similar workloads.
However, the performance multiples AMD presented include not only comparisons based on published benchmark scores but also figures based on the company's own internal estimates. Now that SP7-based systems are available to deploy, the size of the multiples is not the only thing that matters. It is also necessary to distinguish what kind of workload each number measures and under what conditions each comparison holds.
6th Gen EPYC SP7 Begins Shipping, While Cloud Availability Varies
In its announcement at noon US Eastern Standard Time on November 10, AMD said that platforms based on the 6th Gen EPYC SP7 are shipping from major OEMs and ODMs. This moves the new CPU generation from the stage of evaluating specifications and development plans to the stage of selecting servers to deploy. Cloud availability, on the other hand, differs by service, and not everything can be used immediately.
Microsoft Azure's HDv2 is currently in preview. AMD also mentions HXv2 and ND MI455X v7, but not all instances, including these, are generally available. General availability on Azure and Oracle Cloud Infrastructure (OCI) is described as planned for the future. OCI Compute E7 is also planned to use the 6th Gen EPYC.
Meta is evaluating the 6th Gen EPYC as an early customer and is preparing for large-scale deployment. That does not mean large-scale deployment has been completed. Organizations installing servers in their own data centers need to check the shipping systems, and those using the cloud need to confirm which availability stage their target instance is in.
Zen 6 Claims Up to 31% IPC Gain, With CCD Configuration Changes Contributing
AMD claims IPC improvements for Zen 6 of 20% for enterprise and cloud workloads and 31% for AI workloads. IPC (Instructions Per Cycle) is a metric of how many instructions a CPU executes per clock cycle. Even when measured at a fixed clock frequency, however, the entire performance difference does not necessarily come from improvements inside the CPU core.
An explainer from ServeTheHome examines in detail the configuration differences included in this comparison. According to the outlet, AMD compared Zen 5 and Zen 6c at a fixed clock frequency, and Zen 6c has L3 cache capacity equal to that of the larger cores. There are also differences in the configuration of the CCD (Core Complex Die), the chiplet that integrates CPU cores and cache.
Whereas the previous generation split 16 cores across two CCDs, the new generation places 16 cores on a single CCD. ServeTheHome explains that this change also contributes to the performance gain by reducing CCD-to-CCD communication that passes through the I/O die. In other words, the fixed-clock comparison reflects not only improvements to the CPU cores themselves but also the effects of the cache configuration and more efficient inter-core communication.
The 20% and 31% IPC gains should therefore be understood as performance differences under the comparison conditions AMD set. ServeTheHome's article does not independently verify these figures; it analyzes the configuration differences involved in the comparison. The same IPC number means something different depending on whether you are evaluating improvements inside the CPU core or the performance gain of the product as a whole.
Claims of Up to 3.7x Gains, but Evaluation Methods Differ Significantly
AMD's performance comparisons fall broadly into three methods. For general-purpose integer workloads, it uses published benchmark scores. For LLM inference, it estimates performance from PCIe transfer bandwidth. For isolated execution environments for AI agents, it uses CPU thread count as a proxy metric. All are presented as performance multiples, but what is actually measured or estimated differs, so they cannot be compared on the same basis.
When evaluating the performance of a host CPU supporting GPUs, where AI model weights are placed is especially important. The LLM inference comparison used a configuration in which part of the model weights is offloaded from the GPU to CPU-side memory and transferred over PCIe as needed. The effect of changing the CPU or interconnect differs between cases where data transfer constrains performance and cases where processing completes within the GPU.
Based on footnotes ii, iii, and iv in AMD's official November 10 announcement, the performance multiples and how they were calculated are summarized below.
| Target workload | Performance multiple claimed by AMD | Evaluation method and comparison conditions | Caveats |
|---|---|---|---|
| General-purpose integer processing | EPYC 9996 up to 2.2x vs. Xeon 6980P, up to 3.7x vs. Vera | SPECrate2017_int_base in a 2-socket configuration. EPYC 9996 scored 5470, Xeon 6980P 2510, and Vera 1459 as an AMD internal estimate. Published scores are as of November 5; the Vera estimate is as of October 10 | Published measured scores were not used for all comparison targets |
| LLM inference (part of weights offloaded to CPU-side memory) | Up to 1.8x output tokens/second vs. previous-gen EPYC 9575F | 96 cores at 5GHz, Qwen3-30B-A3B, BF16, one NVIDIA B200, vLLM. Extrapolated from measured PCIe Gen5 x16 values to PCIe Gen6 x16 transfer bandwidth | An estimate for conditions where PCIe transfer is the bottleneck; does not apply to inference in general that completes within the GPU |
| Isolated execution environment for AI agents | About 3.3x agents per CPU watt vs. Vera, about 2.5x vs. Xeon 6980P | Theoretical estimate assuming EPYC at 256 cores and 400W, Vera at 450W, and Xeon at 500W. Available CPU thread count used as a proxy for number of agents | Does not measure the number of tasks actually completed or processing speed |
In the general-purpose integer comparison, published benchmark scores exist for the EPYC 9996 and Xeon 6980P, while the score for NVIDIA Vera uses an AMD internal estimate. The 3.7x figure therefore cannot be taken as the result of measuring all CPUs under identical conditions. Note also that the 2-socket EPYC 9996 configuration totals 512 cores, and the 600W figure AMD cites refers to the standard CPU power of the CPU alone. It differs from the power consumption of the entire 2-socket configuration or of the data center, including facilities, so the measurement targets must be aligned when comparing performance and power consumption.
The LLM inference comparison used 1024 input tokens and 256 output tokens, measured three times with the median used. The test environment was a 2-socket configuration with SMT disabled. According to AMD, under these conditions PCIe receive bandwidth reached 89–95% of its limit, while host-side DRAM read bandwidth stayed below 10% of its limit. In other words, the main constraint was not CPU-side memory read performance but data transfer to the GPU over PCIe.
Based on this result, AMD calculated the up-to-1.8x gain by factoring in the expanded transfer bandwidth from moving to PCIe Gen6. This multiple is therefore an estimate built on existing measurements, not a value measured across a complete PCIe Gen6 system. The 6th Gen EPYC lineup also advertises operating frequencies of up to 5.3GHz, but this LLM comparison was run at 96 cores and 5GHz. It cannot be read as the performance at 5.3GHz.
For the isolated execution environment for AI agents, too, the roughly 3.3x and 2.5x figures AMD presented are not the result of measuring actual task-processing performance. They are theoretical comparisons calculated by using available CPU thread count as a proxy for the number of agents that can run concurrently and dividing by CPU power consumption. A multiple based on thread count therefore does not translate directly into an improvement in tasks completed or processing speed. AMD itself explains that actual processing capability varies with the workload, the model used, and memory capacity, as well as software, execution management, and system configuration.
GPU Hosts and AI Agents Call for Different CPU Selection Criteria
When considering adopting the 6th Gen EPYC, it is necessary to identify the factor that constrains performance for each use. For workloads where data transfer with the GPU is the bottleneck, where model weights are placed and which interface they are transferred over matters more than the CPU's standalone compute performance. The LLM inference comparison suggests the need to align the conditions under which weights are offloaded to CPU-side memory and to evaluate PCIe transfer bandwidth as part of the picture. For isolated execution environments for AI agents, on the other hand, concurrent execution performance must be verified, including not just the available CPU thread count but also the memory used by each agent, the model, and the execution management mechanism.
Dan McNamara, senior vice president and general manager of AMD's Compute and Enterprise AI business, said, "AI is significantly expanding the role of the CPU at every layer of the data center." Indeed, when AI agents carry out work, the CPU handles many tasks beyond model inference, including fetching the data needed before and after inference, executing generated code, and communicating with existing applications and databases. The CPU running general-purpose services, the host CPU supporting GPUs, and the CPU running generated code in isolation each require different performance characteristics.
The three "Helion" configurations AMD introduced also reflect these differences in use and installation environment. Helion Liquid is designed for dense compute environments placed alongside Helios, while Helion Air targets deployment in existing air-cooled data centers. Helion Air Inference showcases a configuration that combines CPU processing with local GPU inference in the same rack. This shows that overall system design, including cooling method and placement relative to GPUs, matters as much as CPU performance.
With 6th Gen EPYC SP7 systems now at the shipping stage, it will become possible to evaluate performance and operating conditions on real workloads, not just the published benchmark multiples. In particular, how often model weights are offloaded to CPU-side memory, the transfer bandwidth, and the number of concurrently running AI agents may vary greatly by configuration. Verifying under conditions that match the deployment environment, including compatibility with existing cooling facilities and rack configurations, will be the key to judging how much of the 6th Gen EPYC's performance can be put to use.