SAME autoencoder: VRAM, latency, accuracy

Eight configurations — {SAME-S, SAME-L} × {16-bit, fp8} × {chunked, single-shot} — measured across 32 latent lengths from L=1 to L=8192 (244 points). Accuracy is a full round trip against real music — seven mastered tracks from the 806k-track corpus (rock, jazz, classical, electronic, hip hop, piano, folk) concatenated to 762 s, because L=8192 is longer than any single song and tiling one track would be artificially self-similar. Never noise: the autoencoder is content-dependent and noise fabricates numbers that do not hold on music. 16-bit means fp16 on SAME-L and bf16 on SAME-S. Hover any point for its value.

Encoder — VRAM scratch reserved

TensorRT commits a context's scratch from its PROFILE CEILING, not from the shape you bind — so these are flat in L. That is the whole point of the low band. SAME-L's two precisions reserve identical scratch, so those lines overlap exactly.

SAME-L · 16-bit · chunkedSAME-L · 16-bit · single-shotSAME-L · fp8 · chunkedSAME-L · fp8 · single-shotSAME-S · 16-bit · chunkedSAME-S · 16-bit · single-shotSAME-S · fp8 · chunkedSAME-S · fp8 · single-shot

Encoder — latency

Best of three, after warm-up. Single-shot lines stop at L=4096: that is the wide band's ceiling. Chunked keeps going because it windows.

Encoder — accuracy vs the original music, content held fixed

The SAME 190 s of real music at every point — reassembled from L-sized blocks, so only the processing length changes. This is the apples-to-apples question the TFLite rung table asks, and the answer is the same: flat. From L=16 up the whole curve moves 0.002 dB on SAME-L and 0.020 dB on SAME-S. The fall below L≈16 is real but is not a seam: 0.09 s blocks encoded independently lose context at every boundary.

Decoder — VRAM scratch reserved

Decoder — latency

Decoder — accuracy vs the original music, content held fixed

Same construction, decoder varying. Chunked and single-shot land on the same value to three decimals at every length, which is the direct evidence that the windowed overlap is exact.