Unverified50% confidenceEventExact time
The original team's four-node mixed-GPU cluster throughput decayed by about 18% over 36 hours while GPU utilization remained steady; the cause was diagnosed as KV block fragmentation leading to monotonically rising RSS and drifting TTFT p95.
1
Sources
50%
Confidence
Medium-term (~90 days)
Relevance
10/18/2026
First Seen
Valid until: 1/16/2027
Sources
Related Entities
Related Claims
UnverifiedGPU utilization (DCGM_FI_DEV_GPU_UTIL) measures the proportion of time during a sampling period in which at least one kernel is running on the GPU.70% similarUnverifiedThe trade-off of the hybrid offloading mechanism is that generation speed will be slower than in pure GPU mode.67% similarUnverifiedTraditional inference runs the Prefill stage and Decode stage serially on the same group of GPUs, causing the two types of workloads to interfere with each other and making it difficult to optimize GPU utilization67% similarUnverifiedA GPU warp typically consists of 32 threads; when threads within the same warp take different branches, warp divergence occurs, causing some threads to idle and wait, reducing parallel efficiency.66% similarUnverifiedConverting from FP32 to FP16 can achieve a 2–4x throughput improvement on the Tensor Cores of modern GPUs.66% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1063646API
curl https://kongchang.com/api/v1/knowledge/claims/1063646MCP
get_claim(id=1063646)