int8 vs nvfp4 — 单卡 32GB 能跑多长的视频

viggle 5090 · cuda 后端+sage · steps=1 无 LoRA 全管线(含 VAE 解码)探测 · 帧数取 17k+5 网格(24fps)· 二分到 ≤51 帧精度。探测用 1 步只为找显存上限;实际生成时间按步数线性放大。

上限总表

分辨率量化最大帧数最长时长首个失败帧数受限原因上限时VRAM(MiB)上限时1步耗时(s)
480p(864×480)int8199483.1s2028显存上限32108588.3
480p(864×480)nvfp4134856.2s1399显存上限32104162.8
720p(1280×704)int892338.5s957显存上限32098514.7
720p(1280×704)nvfp458324.3s634显存上限32104129.8
1080p(1920×1088)int839616.5s430显存上限32098497.0
1080p(1920×1088)nvfp424310.1s294显存上限32106166.1

480p · int8 探测轨迹

帧数时长结果1步全管线耗时(s)峰值VRAM(MiB)
1245.2s14.031746
1928.0s16.830850
29412.2s26.331746
43017.9s40.031554
63426.4s62.832002
92338.5s99.132066
134856.2s166.024290
196081.7s536.732098
199483.1s588.332108
202884.5s❌ OOM39.632084
207986.6s❌ OOM39.732084
218190.9s❌ OOM41.332084
2402100.1s❌ OOM30.531330
2844118.5s❌ OOM2.86050

480p · nvfp4 探测轨迹

帧数时长结果1步全管线耗时(s)峰值VRAM(MiB)
1245.2s300.031784
1928.0s16.030856
29412.2s25.227528
43017.9s40.731912
63426.4s62.431560
92338.5s98.731816
134856.2s162.832104
139958.3s❌ OOM36.132106
143359.7s❌ OOM36.432106
150162.5s❌ OOM44.032106
165468.9s❌ OOM44.832106
196081.7s❌ OOM5.732104

720p · int8 探测轨迹

帧数时长结果1步全管线耗时(s)峰值VRAM(MiB)
1245.2s346.930242
1928.0s34.932066
29412.2s56.732066
43017.9s90.332066
63426.4s156.324418
92338.5s514.732098
95739.9s❌ OOM41.332084
99141.3s❌ OOM42.032084
104243.4s❌ OOM43.432084
114447.7s❌ OOM31.731324
134856.2s❌ OOM2.46050

720p · nvfp4 探测轨迹

帧数时长结果1步全管线耗时(s)峰值VRAM(MiB)
1245.2s304.629576
1928.0s31.131944
29412.2s52.931752
43017.9s86.831848
53222.2s165.332106
58324.3s129.832104
63426.4s❌ OOM9.232104

1080p · int8 探测轨迹

帧数时长结果1步全管线耗时(s)峰值VRAM(MiB)
1245.2s362.824844
1928.0s88.221026
29412.2s159.926050
36215.1s263.632052
39616.5s497.032098
43017.9s❌ OOM11.932098

1080p · nvfp4 探测轨迹

帧数时长结果1步全管线耗时(s)峰值VRAM(MiB)
1245.2s333.130528
1928.0s87.232104
24310.1s166.132106
29412.2s❌ OOM10.832104