RealFeedApp Kimi K3 on 16 GB300 GPUs delivers 840 tokens/s
Researchers used agents to write GPU kernels and fuse operators, achieving 840 tokens/s on 16 NVIDIA GB300 GPUs and increasing P90 interactivity by 1.8x on the AgentX workload.
- Field
- applying language models
- What they did
- Researchers used coding agents to automatically write complex GPU kernels and optimize execution code for the Kimi K3 model on NVIDIA GB300 chips.
- Why it matters
- This approach achieves over 840 tokens/s decode speed and higher interactivity, which speeds up the use of language models.
#vllm#kimi k3#gpu kernels#agent assisted#nvidia gb300
