You can now see exactly how a frontier LLM performs on AMD data-center GPUs. The measurements cover realistic multi-turn coding sessions.
AMD published a technical article measuring Kimi K3 inference on its Instinct MI355X GPUs using the TokenSpeed system. On eight MI355X GPUs, TokenSpeed reached 216 output tokens per second per user at concurrency one in a measured multi-turn, code-oriented SWE-smith workload. The article was updated on 2026-10-10, indicating recent refinements to the results.
That matters because teams buying or building inference servers rarely get vendor-grade, workload-specific numbers they can trust for capacity planning. These figures and the tuning guidance let engineers estimate how many users a rack will support and what latency to expect for multi-turn code generation.
TokenSpeed's approach is largely about software and tuning: it optimizes the GPU pipeline so a single user keeps high throughput across several back-and-forth exchanges. Think of it as tuning an engine for stop-and-go city driving rather than a long highway run.
Remember this is a vendor-run measurement. Your mileage will vary with prompt mix, concurrency, and integration choices, and running the reported configuration still requires rack-scale GPU infrastructure.
For infrastructure and platform teams, AMD's article turns an abstract capability into actionable numbers. The next test will be independent benchmarks and different workloads. Will they confirm these gains, or reveal trade-offs?
