FP8 Attention on B300: Videos, Accuracy, Latency

3 July 2026 • 2 minute read

All results: Cosmos3-Nano on a single NVIDIA B300 (sm_103) in vLLM-Omni, official generation configs, same seed (42) within each task, idle-GPU verified. Baseline is cuDNN (the platform default). FP8 (fixed) + Hadamard is the recommended mode: FlashAttention-4 FP8 with the max_offset saturation fix (Dao-AILab/flash-attention#2694) and Hadamard-rotated Q/K/V. Mixed + Hadamard (fp8 QKᵀ, bf16 P·V) is the conservative alternative. Code: vllm-omni#4858 (bf16 backend), lishunyang12/flash-attention#1 (kernel work).

Accuracy (same-seed SSIM vs cuDNN baseline)

Task · configstock FP8mixed + HadamardFP8 (fixed) + HadamardFA4 bf16 (floor)
t2v · 33f · 35 steps0.8720.9500.9510.975
i2v · 33f · 35 steps0.7330.9630.9650.968
t2v · 189f · 35 steps0.970
v2v · 121f · 50 steps0.81150.8115

Attention-kernel relative error on real Cosmos3 activations (vs the bf16 kernel): stock FP8 0.237 → mixed + Hadamard 0.032 → FP8 (fixed) + Hadamard 0.044. The v2v row’s lower absolute SSIM is step-count trajectory divergence, not quantization: the bf16 control lands at the identical 0.8115.

Latency

cuDNNFA4 bf16mixed + HadamardFP8 (fixed) + Hadamard
kernel · 21.5k tokens5058 µs4901 µs (1.03x)4048 µs (1.25x)3402 µs (1.49x)
kernel · 86k tokens81.4 ms78.3 ms (1.04x)65.4 ms (1.24x)54.5 ms (1.49x)
e2e · 189f t2v, s/step2.45–2.552.36 (1.08x)~2.29 (~1.1x)2.00 (1.28x)
e2e · 33f, ms/step267260 (1.03x)252 (1.06x)

Steady-state per denoising step, first-step compile/JIT excluded, measured from unsmoothed elapsed timestamps and cross-checked against profiler-trace timelines. FP8’s advantage scales with sequence length (quantization is O(S), attention O(S²)); 189 frames ≈ 86k tokens is its home turf.

Videos

i2v · official config · same seed

cuDNN · baseline
stock FP8 · SSIM 0.733
mixed + Hadamard · 0.963
FP8 (fixed) + Hadamard · 0.965
Image-to-video, 33 frames, 35 steps, 1280x720. bf16 floor: 0.968.

t2v · official config · same seed

cuDNN · baseline
stock FP8 · SSIM 0.872
mixed + Hadamard · 0.950
FP8 (fixed) + Hadamard · 0.951
Text-to-video, 33 frames, 35 steps, 1280x720. bf16 floor: 0.975.

Long video and v2v

cuDNN · t2v 189f / 35 steps
quantized · SSIM 0.970
cuDNN · v2v 121f / 50 steps
quantized · SSIM 0.811 (= bf16 floor)
Left: long-video t2v at the full official config (~86k-token attention). Right: v2v continuation conditioned on the official reference clip.