Compare
Pick 2–4 records of the same type and read their fields side by side. Values are copied verbatim from the records; the table only aligns them. Fields where the records disagree are shaded, so a contested value cannot hide behind a neighbour that happens to agree.
Start from a question
Flagship NVIDIA vs AMDThe three parts most serving decisions actually come down to.3 recordsSame silicon, different form factorIdentical die, three memory configurations — the datasheet differences are the interesting part.3 recordsDecode engines, consumer siliconDifferent answers to different questions; read avoid_for, not the README.3 records4-bit weight-only formatsAll 4-bit, none of them comparable — the schemes differ in more than the grid.4 recordsScale-up fabricsCompare per-direction bandwidth, not the vendor aggregate.2 recordsMoE models with an explicit active-parameter splitparams_b vs active_params_b is the whole serving problem for these.2 records
Build your own
The row set is fetched per type, so picking accelerators downloads only the accelerator matrix. Selection lives in the URL — copy the address bar to share a comparison.