Chipference

Roofline Pareto model of LLM decode across chip generations

Chart

X axis

Y axis

Compare against

Generation gains (× vs next chip down) Log Y (equal ratios look equal) Speed of light (dashed)

Workload

Input tokens

Output tokens

Count prefill GPUs

Parallelism searched

Tensor parallel Pipeline parallel DP attention

Chips

Hover the chart to compare chips at one point. Click to pin.

Assumptions