How AMD and Embedded LLM scaled Kimi K3 on MI355X to match B300 throughput per GPU
Original titleScaling Kimi K3 on AMD Instinct MI355X: Matching NVIDIA HGX B300 Throughput per GPU on AgentX
AMD and Embedded LLM report that an eight-GPU MI355X system serving Kimi K3 with vLLM delivers 24.4% more total tokens per second per GPU than NVIDIA HGX B300 at 50 tokens/s/user P90 interactivity on the AgentX benchmark.
Under the benchmark's modeled ownership cost, it delivers 87.4% more tokens per dollar.
The post explains the optimizations behind these results, including decode context parallelism, MoE output deduplication, kernel fusion, MLA decode padding, and KDA parallelization.
The post traces optimizations for serving Kimi K3 on AMD MI355X, linking vLLM serving choices to GPU kernels and hardware limits, useful for diagnosing bottlenecks in other inference stacks.
Source: vLLM Blog · vllm.aiPublished · added here