Unverified50% confidenceFactExact time
若HBM带宽不足，GPU算力利用率（MFU）可能跌至10%以下
1
Sources
50%
Confidence
Long-term
Relevance
7/19/2026
First Seen
Sources
Related Claims
UnverifiedHigh-end GPUs (NVIDIA H100/H200) have significant supply-side bottlenecks, and the pace of cost reduction is far slower than that of bandwidth or storage in the internet era.70% similarVerifiedA100 GPU的HBM带宽约2TB/s，而片上SRAM带宽可达19TB/s，相差近10倍65% similarUnverifiedThe HBM memory bandwidth of the NVIDIA H100 is approximately 3.35 TB/s, and its utilization rate in autoregressive generation tasks is often below 10%.64% similarVerified分层加载导致推理速度受限，其根本原因是 PCIe 总线带宽（PCIe 4.0 理论带宽约 32GB/s）远低于 GPU 内部显存带宽63% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/559603API
curl https://kongchang.com/api/v1/knowledge/claims/559603MCP
get_claim(id=559603)