Unverified50% confidenceFactExact time
DeepInfer's publicly disclosed PD disaggregation optimization splits the Pre-fill stage and the Decode stage onto different hardware: an 8-card server handles Pre-fill, and a 64-to-128-card supernode handles Decode, with a GPU ratio of about 1 to 1.
1
Sources
50%
Confidence
Long-term
Relevance
10/31/2026
First Seen
Sources
DeepSeek Open-Sources DeepInfer: The Cost-Saving Tricks Are There, the Scheduling Isn't
bilibili硬哥拆解10/31/2026
Related Entities
Related Claims
UnverifiedThe Prefill (prompt processing) phase is compute-intensive and relies on the GPU's parallel matrix computation capability; the Decode (generation) phase is memory-bandwidth-intensive and depends more on VRAM bandwidth.73% similarVerifiedThe Prefill stage of LLM inference is compute-intensive with GPU utilization close to 100%, while the Decode stage is memory-bandwidth-intensive with the bottleneck being HBM VRAM bandwidth.73% similarUnverifiedThe Decode stage is a memory-bandwidth-intensive task, with GPU utilization typically only 20–40%.72% similarUnverifiedThe prefill stage is a compute-intensive operation primarily limited by GPU compute, while the decode stage is a memory-bandwidth-intensive operation primarily limited by VRAM bandwidth.71% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1120182API
curl https://kongchang.com/api/v1/knowledge/claims/1120182MCP
get_claim(id=1120182)