Faster GPU, Same Token Latency: Why CPU Launch Overhead Bottlenecks LLM Inference
blaizedsouza · x · 2026-10-15
A thorough explainer of a classic LLM interview question: move your model to a faster GPU, profiling shows kernels finishing sooner, yet per-token latency barely improves. Where did the speedup go?
- An LLM forward pass is a long sequence of ops (attention, matmul, normalization, sampling, data movement), not one giant GPU operation
- Without CUDA Graphs, the CPU prepares and submits each op individually, adding small launch overhead on every call
- Decoding repeats this op sequence per token, so the overhead compounds: some kernels run only microseconds, leaving the GPU idle waiting for the CPU to submit the next one
- Those gaps in the profiler trace are the missing speedup — the bottleneck is CPU submission, not GPU compute
The takeaway: buying more FLOPS won't speed up decoding; you need CUDA Graphs or batched submission to eliminate launch overhead. Links to the author's follow-up article on traditional ML vs LLM inference.
More from Infra
- NVIDIA and Sweden's Wallenberg launch Sferical One, a shared AI supercomputer — nvidia · 2026-10-16
- Next infra frontier: self-evolving systems that adapt and self-verify — _shreya_s · 2026-10-16
- Cerebras pushes disaggregated inference: heterogeneous prefill pools feeding wafer-scale decode — ai · 2026-10-16
- d-Matrix GA's Corsair rackscale, demos first public Attention-FFN disaggregation at 845 tok/s — ai · 2026-10-16
- Actual's inference engine hits 10k tok/s on NVIDIA DGX Station, GB300 case study — natesiggard · 2026-10-16
- TSMC Q3 revenue hits $46.7B, up 41% YoY; Q4 guidance of $52-54B tops estimates — Beth_Kindig · 2026-10-16