Head to head
A
B
Table (click a header to sort; click a row to scroll to its card)
Clips
Picks
Starred arms with their numbers, as markdown; copy this back into the session.
Scene: spider (2D anime, 90 f T2VA, 25 steps, 1376x768, shot cut at 3.0 s) (all scenes). Same seed everywhere; the reference is the exact W4A8 arm (ref_w4a8_exact). Most arms here FAKE-QUANTISE the activations of a region of the shipped W4A8 model (weights stay int4 g16); the gold card is a REAL NVFP4 checkpoint rendered on the native fp4 kernels, weights and activations both. Latent mean |dx| counts dose, not damage (kvi8r "perfect" = 0.24; one NVFP4 block = 0.35; FP8 everywhere = 0.75 vs NVFP4 everywhere 0.78 at 3x lower per-op error). Scene-specific ruler notes: PSNR falls monotonically with latent |dx| (19.6 to 11.7 dB); motion-curve corr is bimodal - ~0.97 means the two big events (frame 32, the shot cut at 71/72) land on the same frames, ~0.5 means one slid a frame, below 0.3 means events moved several frames. Wall times on fake arms carry the fake-quantiser's overhead (2-10% slower than the reference); the real card's wall time is the regime's actual speed (0.84x here).
Click a clip to play it with sound; hover plays muted; d on a card swaps in the amplified |arm - ref| view. Star candidates, send two to A/B, flip with f, choose which side you hear, wipe with the slider. Cost columns marked [PROJ] are projections from GEMM microbenches; everything else is measured.
Starred arms with their numbers, as markdown; copy this back into the session.