Kimi K3 on MI355X: vLLM inference results
What happened
An October 16 report covered vLLM serving of Kimi K3 on AMD Instinct MI355X, with optimization and benchmark results. On eight MI355X cards, throughput was measured on the AgentX benchmark at an interactivity target of 50 tokens/s per user at P90. That setup reached about 7,567 tokens/s per GPU, against 6,083 tokens/s per GPU on an NVIDIA HGX B300, a 24.4% edge for MI355X. The report gives the configuration and test conditions but does not disclose the specific optimization steps.
Written by AI from the coverage · updated 1 hour ago
Coverage
Follow the reports to see the story from different sides.
- AI HOT · Tips & opinionsvLLM 解析 Kimi K3 在 AMD Instinct MI355X 上的推理优化，单 GPU 吞吐超 NVIDIA HGX B300 24.4%
在 AgentX 基准上，八卡 AMD Instinct MI355X 用 vLLM 服务 Kimi K3，在 50 tokens/s/user P90 交互性下每 GPU 吞吐约 7,567 tokens/s，比 NVIDIA HGX B300 的 6,083 高出 24.4%。
Heat over time
Not enough continuous observations to draw a trend yet.