Unverified50% confidenceFactExact time
With a batch size of 1, GPU parallel compute unit utilization is very low, because only one token participates in matrix operations at a time and a large number of SMs sit idle.
1
Sources
50%
Confidence
Long-term
Relevance
10/27/2026
First Seen
Sources
Related Entities
Related Claims
Unverified单Token生成的算术强度约为1 FLOP/Byte，导致GPU计算单元利用率不到1%79% similarUnverifiedThe cost of the unified memory architecture is lower memory bandwidth than the dedicated VRAM of high-end discrete GPUs, so throughput is usually below that of data-center GPUs.70% similarUnverified自回归推理是memory-bound（内存带宽受限）问题，因为batch size通常为1时算术强度极低，GPU大部分时间在等待数据搬运70% similarUnverifiedGPU的warp切换几乎零开销，因为GPU为每个warp在硬件寄存器文件中预分配了独立的寄存器空间70% similarUnverifiedGPU每个SM的共享内存通常仅有48-96KB70% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1104076API
curl https://kongchang.com/api/v1/knowledge/claims/1104076MCP
get_claim(id=1104076)