̽»¨ÊÓƵ

The “utilization illusion” illustrates high GPU SM usage appears to indicate efficiency but actually masks low overall throughput. Delays such as kernel launch latency, CPU–GPU control jitter, and synchronization barriers cause the hardware to stall, reducing token output despite steady utilization metrics. The key insight is that runtime coordination, not raw compute power, is often the real performance bottleneck.