Unverified50% confidenceFactExact time
100% GPU utilization can mask issues such as low kernel occupancy, memory bandwidth bottlenecks, CPU-GPU data pipeline blocking, and excessive communication overhead; this phenomenon is known as false saturation.
1
Sources
50%
Confidence
Long-term
Relevance
8/7/2026
First Seen
Sources
Related Claims
Unverified过度融合可能导致单个内核占用过多寄存器，降低GPU占用率（Occupancy）从而削弱并行度72% similarUnverifiedGPU上下文切换需保存完整GPU状态（包括寄存器文件、共享内存和L1缓存），数据量可达数百MB，而CPU上下文切换仅需保存几KB72% similarUnverifiedCPU offload places some Transformer layers in system main memory to be computed by the CPU, with data transferred over the PCIe bus (PCIe 4.0 x16 theoretical bandwidth is approximately 32 GB/s), while GPU memory bandwidth is typically in the range of hundreds to over a thousand GB/s, making the bandwidth gap the fundamental cause of the performance bottleneck.72% similarUnverified传统LLM推理中，KV Cache的内存管理导致60%-80%的GPU显存被浪费71% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/702284API
curl https://kongchang.com/api/v1/knowledge/claims/702284MCP
get_claim(id=702284)