MACH X

Dense-model decode on a single card is Mach X's known weak axis — llama.cpp's SYCL kernel edges the Qwen3.6-27B decode loop (15.2 vs 23.2 tok/s). It's the one place a competitor is faster, and it's flagged in amber above rather than buried. Every other bar on this page is a win, and the fix (a faster dense-decode GEMV) is the next lever, not a mystery.

Speed only counts if the output is right. “Bit-exact” is the strongest correctness claim there is: the engine produces output that is byte-for-byte identical to a trusted reference — the same tokens, down to the same bits — not merely “close.” It proves the speed didn't come from quietly cutting numerical corners.

Bit-exact ✓

Dense models vs llama.cpp. Per-layer activations, greedy token stream, and perplexity all match the llama.cpp oracle bit-for-bit (Llama-3.x cosine 0.99998–1.0).

Bit-exact ✓

Multi-GPU layer-split vs single-GPU. Splitting a model across two B70s changes nothing about the result — the output is byte-identical to running it on one card.

Equivalent ≈

Tensor-parallel. Summing each card's partial results in a different order flips the last few bits, so this path is validated by cosine ≈ 0.999999 + matching perplexity rather than bit-identity — numerically the same answer, not the same bits.