This file owns the measured speed results that used to make the main README hard to scan. Times are complete product walls unless a row explicitly says it is denoiser-only. Lower is better. Ratios compare the native compiler with the matched framework baseline on the same GPU and workload unless explicitly labeled native-versus-native.
| Model and product boundary | GPU | Native C++ | Matched baseline | Native result |
|---|---|---|---|---|
| FLUX.2 [klein] Base 9B, prompt to PNG | RTX 5080 | 52.809 s | 99.242 s | 1.879x faster |
| Krea 2 Turbo, prompt to PNG | RTX 3090 Ti | 26.58 s | 59.14 s | 2.225x faster |
| MiniMax-H3 FL2VA, prompt to MP4 | RTX 3090 Ti | 89.77 s | 81.095 s | 1.107x slower |
The model rows are different workloads and hardware targets. Compare native with baseline within a row; do not compare raw seconds across rows.
xychart-beta
title "FLUX.2 Base 9B on RTX 5080 — lower is better"
x-axis ["Native C++", "Stock ComfyUI"]
y-axis "Seconds" 0 --> 105
bar [52.809, 99.242]
xychart-beta
title "Krea 2 Turbo on RTX 3090 Ti — lower is better"
x-axis ["Native C++", "Framework baseline"]
y-axis "Seconds" 0 --> 65
bar [26.58, 59.14]
xychart-beta
title "MiniMax-H3 FL2VA on RTX 3090 Ti — lower is better"
x-axis ["Native C++", "Stock ComfyUI"]
y-axis "Seconds" 0 --> 85
bar [89.77, 81.095]
The frozen workload uses the official undistilled Base 9B transformer, creator
generalized-time Euler schedule, creator F32 VAE, 1024x1024 output, 50 steps,
guidance 4, seed 20260901, and the literal prompt
A cat holding a sign that says hello world.
| Measurement | Native admitted plan | Stock ComfyUI |
|---|---|---|
| Prompt to PNG, three-run range | 52.793–52.842 s | 99.242 s |
| Prompt to PNG, median | 52.809 s | 99.242 s |
| Denoise wall, measured range | 47.767–47.930 s | included in product wall |
| Streamed transformer weights per step | 0 B | framework-managed |
| Resident transformer weights | 9,107,591,168 B | framework-managed |
| Peak prepared native bytes | 11,349,488,128 B | 15,641 MiB peak NVML |
Native saves 46.433 seconds, or 46.788% of baseline latency. That meets the approved near-55-second target but is not labeled a 2x result.
The admitted denoiser plan applies generic H4096 F32 signed ConvRot W8A8 to 112
linears, uses group-64 INT8 weight-only storage for the nine remaining
linears, and uses group 32 for final_layer.linear.weight. Only the packed
single-stream input projections use activation clip ratio 0.999; every other
activation ratio remains 1.0. Exact attention uses cuDNN SDPA heuristic A.
The 14,800 MiB first-consumer residency plan keeps every transformer weight
resident during denoising.
The unchanged final-token gate passes against the frozen native creator oracle:
| Gate | Result |
|---|---|
| Cosine | 0.9988074500 |
| Relative L2 | 0.0488599297 |
| Maximum absolute error | 1.73046875 |
| Norm ratio | 1.0007087464 |
| Nonfinite values | 0 |
The decoded PNG passed full-resolution visual inspection for anatomy, fur,
exposure, color, and legible hello world text. Its SHA-256 is
600a201c6a76dc4d23ea19c9603f79ec765f344400460c6e441e161d07f5cde1.
The practical ComfyUI comparator matches transformer, prompt, seed, geometry, steps, guidance, schedule class, and F32 VAE. Its packaged Qwen conditioner is FP8-mixed while native conditioning is BF16; that physical precision difference is disclosed. The native executable links no libtorch or libpython and invokes no Python worker.
The complete warm-page-cache workload is 1024x1024, 4096 image tokens, 512 text
tokens, 28 MMDiT blocks, BF16 production math, eight Euler steps, CFG disabled,
and mu=1.15. Prompt, seed, initial latent, schedule, checkpoint, and
conditioning boundary are fixed.
| External stage | Native C++ | Creator / ComfyUI / PyTorch BF16 |
|---|---|---|
| Tokenizer + Qwen3-VL | 2.28 s | 9.05 s |
| TextFusion | 1.29 s | 7.52 s |
| Prepared 8-step denoise | 19.54 s | 34.26 s |
| Qwen-Image VAE + PNG | 3.43 s | 8.31 s |
| Prompt to PNG | 26.58 s | 59.14 s |
That is a measured 2.225x complete-chain speedup. The native run recorded zero filesystem input, so this is a warm-page-cache result. A separate cold-filesystem native diagnostic took 55.58 seconds, dominated by 27.42 seconds of Qwen weight page-in.
All eight denoiser states and the final latent are bit-identical to the frozen
creator trajectory. The VAE gate passes at cosine 0.99999339 and relative L2
0.00363512 after clamping. The inspected PNG SHA-256 is
eea79ee7d84a703235481b0e1859ca087fa20d40304aa0243d9929da8333fbfd.
The framework comparator is ordinary eager ComfyUI/PyTorch without
torch.compile or Inductor fusion. The strict comparator uses the creator
padding mask and matching post-TextFusion boundary. Stock ComfyUI's Krea
frontend omitted that mask at main blocks in the checked revision, so its
product timing is retained but is not a numerical-parity arm.
The matched contract is two keyframes, 832x480, 124 delivered frames at 24 fps,
207 audio latents, 1,256 conditioner tokens, sequence length 16,880, and seven
res_multistep/simple evaluations. Both measurements cover a literal prompt
through the saved H.264/AAC MP4.
| Complete-wall measurement | Time |
|---|---|
| Native Compiler ConvRot INT8, exact attention on evaluations 1-3, INT8 attention on 4-7 | 89.77 s (runs 89.77, 102.44, 88.54; median) |
| Native Compiler ConvRot INT8 + INT8 attention on every evaluation (2026-09-01, retired for quality) | 78.88 s |
| Matched stock ComfyUI ConvRot INT8, INT8 attention on every evaluation | 81.095 s |
| Strict 2x ceiling | <40.548 s |
Native is 1.107x slower than the comparator, 8.675 seconds or 10.7%, and 49.2 seconds above the strict 2x ceiling; no H3 speed result is claimed. The comparator has not been measured at matched exactness.
Why the recipe changed (2026-09-03): the seven-evaluation latents were
decoded with the INT8 attention route on every evaluation, with exact cuDNN
attention on every evaluation, and with exact attention on the first two,
three, and four evaluations, all from identical inputs, and gated with
difquality against the exact decode (bars PSNR 30 dB, SSIM 0.90, audio SNR
20 dB, eight sampled frames):
| attention route | worst frame PSNR | worst SSIM | audio SNR | denoise cost |
|---|---|---|---|---|
| INT8 on all 7 (former recipe; also the comparator's route) | 24.5 dB | 0.850 | 1.7 dB | 0 |
| exact on 1-2 | 27.1 dB | 0.908 | 3.1 dB | +7 s |
| exact on 1-3 (recipe) | 32.3 dB | 0.958 | 12.9 dB | +10.5 s |
| exact on 1-4 | 37.5 dB | 0.978 | 14.7 dB | +14 s |
| exact on all 7 | reference | +24.4 s |
The INT8 route's per-evaluation error is amplified along the sampler
trajectory; the comparator produces the same off-reference output because it
dispatches the same INT8 kernel. The video bars pass from three exact
evaluations; the audio stays under the 20 dB bar and was accepted on a
listening review, which the recorded difquality verdict does not replace
(it stays FAIL on the audio bar). Each exact evaluation costs about 3.5 s
(10.6-10.8 s versus 7.2-7.6 s).
The native chain includes presentation, Qwen3-VL vision/text conditioning,
both keyframe encodes, seven denoiser evaluations, video/audio decode, and mux.
The saved 832x480 MP4 contains 124 H.264 frames at 24 fps plus 32 kHz stereo AAC
and passed visual inspection. Its SHA-256 is
cfb0c6b8110ff2dda9a0a240e57b5a16476cff0b3c5c41872f131e2ebc21e64f.
The saved MP4 from the current recipe (run 2 of the protocol) has SHA-256
63c2551ef666c130df09341f599c94fa9974afd295683d9cc336654d1d2dccea. Decoded
parity with exact attention is as tabulated above; parity with the stock
competitor is not the bar any more, since the competitor's output is the
INT8-only one.
An exact-output QKV/RMSNorm/RoPE fusion improved isolated hot denoiser timing but regressed complete wall to 84.49 seconds warm and 92.05 seconds cold. It was removed; local kernel wins do not override the complete-wall result.
dif_exact_stream_attention is a project-owned exact dense non-causal
attention backend for Opcode::Attention (noncausal BF16 [S,H,128] /
[B,S,H,128], full heads, no bias): flash-style FP32 online softmax, K/V
tiles streamed through 48 KiB of shared memory with cp.async double
buffering, Q fragments and the FP32 output accumulator register resident,
zero global workspace. Both Q·Kt and P·V use BF16 operands with FP32
accumulation. This preserves V's BF16 exponent range and prevents the FP16
tile-sum overflow found with Q=K=0, S=32, V=4096 (the previous kernel
returned infinity instead of 4096).
The measurements below are historical results from the earlier FP16 P·V kernel. They do not establish performance or end-to-end parity for the corrected BF16/FP32 kernel; those full-model gates must be remeasured.
Isolated H3 DiT geometry [16880,56,128] BF16, warm, 20 iterations, three
interleaved sessions, real-data inputs (constant-fill inputs understate time
by ~13% through power draw): exact-stream 67.50–67.77 ms vs cuDNN SDPA
69.97–70.10 ms (1.035x) vs native FlashAttention-2 74.87 ms. Numerics vs
cuDNN: cosine 0.9999967, rel L2 2.55e-3, max abs 4.88e-4, 0 nonfinite —
tighter than the cuDNN-vs-FlashAttention exact noise floor (3.10e-3).
Against an f64 oracle at the full geometry: cosine 0.9999986, rel L2
1.69e-3.
Whole-system A/B on the RTX 5080 T2VA chain (832x480x124, S=15,283, 20 schedule points, 19 evaluations, ConvRot INT8 resident 28): denoiser wall 105.56/105.86 s (cuDNN) vs 103.43/103.52 s (exact-stream), attention GPU time 54.55/54.83 s vs 52.35/52.41 s — a reproducible 2.2-second (2.1%) complete-denoise saving with byte-identical planned VRAM (13,878,443,008 bytes) and bit-identical repeat runs per backend. End-to-end latent divergence from cuDNN (cosine 0.913 after 19 evaluations) is smaller than exchanging cuDNN for FlashAttention-2 (0.907); decoded MP4s from both backends passed visual inspection. The frozen RTX 3090 Ti FL2VA row above is unchanged; its fixtures and GPU gate cannot run on this machine.
This is a native-versus-native partial-cold comparison, not a new ComfyUI comparison or a warm-generation claim. The frozen T2VA workload is 832x480x124 at 24 fps, 439 text tokens, 19 Euler evaluations, BF16 conditioning, all 50 ConvRot INT8 projection blocks (28 GPU-resident), and exact cuDNN attention. Both arms regenerate conditioning and noise and include both decoders and mux. The declared ConvRot/text/VAE cache files start cold; prepared programs/kernel caches are retained. These original measurements did not explicitly evict the original transformer shards used for the remaining non-quantized weights, so they are not an all-model-files-cold result. The current exact recipes include those shards for subsequent measurements. The complete process group has a 22 GiB RAM cap, no swap, and a 300 W GPU limit.
| Complete prompt-to-MP4 measurement | Before (4de52a1) |
Cache fix |
|---|---|---|
| Raw samples, seconds | 255.929, 267.175 | 170.087, 180.291, 174.542 |
| Median, seconds | 261.552 | 174.542 |
| Peak process-group RAM, including file cache | 22.000 GiB | 13.121 GiB |
The measured ratio is 1.4985x, a 33.27% latency reduction. This does not meet a strict 1.5x threshold and does not meet the requested 2x target. The third candidate run uses the final prepare-time keep/evict policy ordering; the first two use the same lazy-upload path before that ordering fix.
The conditioner had been rereading approximately 50 GB of weights to warm a cache that could never fit. That also prevented the denoiser's reusable streamed tail from remaining cached, causing about 172 GB of denoiser disk reads. Warming now requires whole-working-set headroom and respects eviction of host copies of GPU-resident weights. Final-run denoiser disk reads fell to 28.3 GB. No attention, GEMM, precision, schedule, or checkpoint changed.
The final candidate is byte-identical to the control at token IDs,
conditioning, both initial noises, video/audio latents, raw/decoded video,
WAV audio, and final MP4. The MP4 SHA-256 is
249383056146ab47f3152ac85c979a2a7c351e2432e44f748c87e2b578c0a59d;
decoded frames were inspected. Release core, tensor-I/O, target, and benchmark
tests pass. Replay uses
h3-t2va-convrot-exact.json with
difbench run --drop-file-cache; its checkpoint/program prerequisites and
machine-specific resident-layer count must be supplied.
The cache mechanism is shared and adds no SM120-only kernel requirement. RTX 3090/Ti execution is not yet revalidated; its speedup must not be inferred from this T2VA result. Local image-mode checks follow below. Sol-Attn was screened separately on real captured H3 Q/K/V: useful speeds had significant numerical departures, and no sparse attention replacement was admitted.
Native-versus-native, one full-cold sample per arm, using the same 832x480, 124-frame, 19-evaluation, BF16-conditioner/ConvRot-INT8/exact-attention contract. These recipes explicitly reset the original transformer shards as well as the derived packs, text encoder and VAEs. All prompt/image processing, noise, denoising, both decoders and MP4 saving are timed.
| Pipeline | Before | Cache fix | Measured speedup |
|---|---|---|---|
| FL2VA, first frame | 272.082 s | 181.095 s | 1.502x |
| FL2VA, last frame | 271.287 s | 185.133 s | 1.465x |
| FL2VA, first + last | 288.026 s | 195.028 s | 1.477x |
| Ref2VA, one image reference | 236.898 s | 184.401 s | 1.285x |
All four pairs are byte-identical through raw/decoded video, WAV and final MP4: 79 byte comparisons and 12 zero-error numerical checks passed, with no nonfinite latent values. Contact sheets were inspected. Candidate process-group RAM stayed below 15.84 GiB under the unchanged 22 GiB cap; no OOM occurred. The four selected release suites also pass.
Replay fixtures are h3-fl2va-keyframe-convrot-exact.json (first or last),
h3-fl2va-first-last-convrot-exact.json, and
h3-ref2va-image-convrot-exact.json under perf/recipes/; see
preparation and scope.
Ref2VA uses its own denoiser pack and bundle, not FL2VA weights. This is
same-precision native regression evidence, not creator-BF16 versus INT8 parity.
2x is not reached. These individual 480p pairs do not establish a speed guarantee at other resolutions or for larger reference sets. Reference-video/ audio input encoding and actual RTX 3090/Ti execution remain unverified.
- Freeze checkpoint, prompt, seed, resolution, steps, guidance, scheduler, and output gate before comparing plans.
- Call a timing end-to-end only when it spans literal input to saved output.
- Keep cold, warm, preparation, and hot-step measurements separate.
- Admit speed only after numerical, visual-quality, memory, and replay gates.
- Never manufacture a gain with fewer steps, lower resolution, changed conditioning, or relaxed quality thresholds.