The CPU Behind the Agents: Benchmarking Intel Xeon 6 for Agentic AI Workloads
Authors:
October 14, 2026
An independent Terminal Bench 2 evaluation of AWS m8i (Intel Xeon 6), m7i (4th Gen Intel Xeon Scalable), and m8a (5th Gen AMD EPYC).
Agentic AI shifts meaningful work onto the CPU. Between model responses, autonomous agents run shell commands, manipulate files, compile code, execute tests, and manage containers. At fleet scale, these events determine how fast tasks complete, how many agents a server can sustain, and what that work costs. All of this work either runs on the CPU or is managed by services that do.
Signal65 independently tested multiple public cloud instance sizes and CPU types using an agentic record and replay framework. By using both Intel Xeon and AMD Turin CPU types and two different instance sizes, we were able to measure the impact of these CPU choices on the performance, and more importantly the cost, of agentic workloads.
The instance choices included AWS m8i (Intel Xeon 6, Granite Rapids), m7i (4th Gen Intel Xeon Scalable, Sapphire Rapids), and m8a (5th Gen AMD EPYC, Turin) using an agentic record and replay tool. The testing used Terminal Bench 2 with recorded inference responses to eliminate LLM variance between runs, emphasizing on-instance execution: the agent’s commands, containerized (sandboxed) task environments, and orchestration. Each test was repeated three times at different concurrency levels with up to 100 processes tested.
Testing found the following highlights:
- Clear generational differences: Compared with prior-generation Intel m7i instances, newer m8i instances completed tests 6.9 – 11.0% faster and at 2.2 – 6.6% lower compute cost.
- Intel and AMD have similar economics: When comparing price-performance, m8i test cost was at parity or better at four of five tested concurrency levels, with an approximately 5.5% cost advantage at concurrency 20, m8i's most cost-efficient tested operating point. m8a instances completed tests faster at an approximately 15% higher hourly rate.
- Strong scale-up utilization when orchestration supplies enough work: On m8i.16xlarge, as our testing raised the number of concurrent processes, average CPU utilization increased to approximately 54%, with peaks near 100%, compared to average utilization rates of less than 10% using the standard Terminal Bench single-process runs.
For this replay workload, m8i is the clear upgrade for existing m7i deployments and offers competitive economics against m8a, while m8a remains the latency leader. The scale-up result also shows that larger systems can support much more concurrent agent work, but only when the orchestration layer supplies enough independent work.
Research commissioned by:
