How long can a de-rope go on your card

Measured 2026-08-18 on one machine (RTX PRO 6000 Blackwell fenced to smaller cards with a VRAM balloon, RAM fenced with cgroups), MiniMax-H3 W4A8 DiT, MAINodes H3 Streamed Blocks (exact), ComfyUI dynamic VRAM, --fast-disk, PyTorch flash attention. Hover a point. Frames are at 1376x768; other canvases scale by tokens.

VRAM: activations are card-independent; the card only sets the ceiling06121824303600k29490k702214k906276k1110339k1416432k1722525kframes at 1376x768 (packed tokens below)GiB16 GB24 GB32 GBtorch activation peak (in-forward, any card) 90 f / 28k tokens: torch peak +1.8 GiB (TE matrix A, 25 steps)torch activation peak (in-forward, any card) 90 f / 28k tokens: torch peak +1.8 GiB (TE matrix A, 25 steps)torch activation peak (in-forward, any card) 702 f / 214k tokens: torch peak +11.9 GiB (v014: bank, fast-disk, balloon 63)torch activation peak (in-forward, any card) 702 f / 214k tokens: torch peak +11.9 GiB (v012: planner+cap card)torch activation peak (in-forward, any card) 702 f / 214k tokens: torch peak +11.9 GiB (v011: normal mode, no fast-disk)torch activation peak (in-forward, any card) 702 f / 214k tokens: torch peak +11.9 GiB (line sweep, 2 steps)torch activation peak (in-forward, any card) 702 f / 214k tokens: torch peak +11.9 GiB (line sweep, 2 steps)torch activation peak (in-forward, any card) 702 f / 214k tokens: torch peak +11.9 GiB (line sweep, 2 steps)torch activation peak (in-forward, any card) 702 f / 214k tokens: torch peak +11.9 GiB (line sweep, 2 steps)torch activation peak (in-forward, any card) 906 f / 276k tokens: torch peak +14.9 GiB (line sweep, 2 steps)fit fit: 0.34 + 53 KB/tokenfit fit: 0.34 + 53 KB/tokenfit fit: 0.34 + 53 KB/tokenfit fit: 0.34 + 53 KB/tokenfit fit: 0.34 + 53 KB/tokenfit fit: 0.34 + 53 KB/tokenfit fit: 0.34 + 53 KB/token16 GB card, process VRAM peak 90 f: process VRAM peak 15.7 GiB on a 16 GB card (TE matrix A, 25 steps)16 GB card, process VRAM peak 702 f: process VRAM peak 15.5 GiB on a 16 GB card (line sweep, 2 steps)16 GB card, process VRAM peak 702 f: process VRAM peak 15.5 GiB on a 16 GB card (line sweep, 2 steps)24 GB card, process VRAM peak 90 f: process VRAM peak 23.0 GiB on a 24 GB card (TE matrix A, 25 steps)24 GB card, process VRAM peak 702 f: process VRAM peak 23.5 GiB on a 24 GB card (line sweep, 2 steps)24 GB card, process VRAM peak 702 f: process VRAM peak 23.4 GiB on a 24 GB card (line sweep, 2 steps)24 GB card, process VRAM peak 906 f: process VRAM peak 23.4 GiB on a 24 GB card (line sweep, 2 steps)31 GB card, process VRAM peak 702 f: process VRAM peak 25.2 GiB on a 31 GB card (v014: bank, fast-disk, balloon 63)32 GB card, process VRAM peak 702 f: process VRAM peak 33.9 GiB on a 32 GB card (v012: planner+cap card)32 GB card, process VRAM peak 702 f: process VRAM peak 34.0 GiB on a 32 GB card (v011: normal mode, no fast-disk)torch activation peak (in-forward, any card)fit16 GB card, process VRAM peak24 GB card, process VRAM peak31 GB card, process VRAM peak32 GB card, process VRAM peakfit: torch peak = 0.34 GiB + 53 KB per packed token (K/V 28.7 + x 10.8 + residents/chunks). Points not rendered are listed in the table.Host RAM (process RSS peak): add your own OS + ComfyUI headroom011223344556600k29490k702214k906276k1110339k1416432k1722525kframes at 1376x768 (packed tokens below)GB32 GB64 GBsingle T2VA pass, --fast-disk 90 f: RSS 14.3 GB, single pass, 16 GB card, 64 GB machine (TE matrix A, 25 steps)single T2VA pass, --fast-disk 90 f: RSS 9.4 GB, single pass, 24 GB card, 64 GB machine (TE matrix A, 25 steps)single T2VA pass, --fast-disk 702 f: RSS 27.4 GB, single pass, 16 GB card, 32 GB machine (line sweep, 2 steps)single T2VA pass, --fast-disk 702 f: RSS 27.4 GB, single pass, 16 GB card, 64 GB machine (line sweep, 2 steps)single T2VA pass, --fast-disk 702 f: RSS 25.8 GB, single pass, 24 GB card, 32 GB machine (line sweep, 2 steps)single T2VA pass, --fast-disk 702 f: RSS 25.8 GB, single pass, 24 GB card, 64 GB machine (line sweep, 2 steps)single T2VA pass, --fast-disk 906 f: RSS 29.8 GB, single pass, 24 GB card, 32 GB machine (line sweep, 2 steps)full de-rope pipeline, --fast-disk 702 f: RSS 40.2 GB, de-rope pipeline, 31 GB card (v014: bank, fast-disk, balloon 63)full de-rope pipeline, --fast-disk 702 f: RSS 36.4 GB, de-rope pipeline, 32 GB card (v012: planner+cap card)de-rope pipeline, no --fast-disk 702 f: RSS 59.4 GB WITHOUT --fast-disk (v011: normal mode, no fast-disk)single T2VA pass, --fast-diskfull de-rope pipeline, --fast-diskde-rope pipeline, no --fast-diskRSS with --fast-disk = ComfyUI base + CPU-side intermediates (a 702-frame fp32 IMAGE is 8.9 GB); without it, add every model's CPU copy (~33 GB here).

Points

framestokenscardmachinegraphtorch peakVRAM peakRSSnote
9028k16 GB64 GBsingle pass1.815.714.3TE matrix A, 25 steps
9028k24 GB64 GBsingle pass1.823.09.4TE matrix A, 25 steps
702214k16 GB32 GBsingle pass11.915.527.4line sweep, 2 steps
702214k16 GB64 GBsingle pass11.915.527.4line sweep, 2 steps
702214k24 GB32 GBsingle pass11.923.525.8line sweep, 2 steps
702214k24 GB64 GBsingle pass11.923.425.8line sweep, 2 steps
702214k31 GB64 GBde-rope pipeline11.925.240.2v014: bank, fast-disk, balloon 63
702214k32 GB64 GBde-rope pipeline11.933.936.4v012: planner+cap card
702214k32 GB96 GBde-rope pipeline (no fast-disk)11.934.059.4v011: normal mode, no fast-disk
906276k16 GB32 GBsingle pass---OOM
906276k16 GB64 GBsingle pass---OOM
906276k24 GB32 GBsingle pass14.923.429.8line sweep, 2 steps

Reading it: VRAM needed = resident weights (dynamic VRAM shrinks these to a few GB when it must) + the blue line + ~1 GB context; a card fits while that stays under it. RAM needed = the RSS series for your graph type + your OS and ComfyUI headroom, which you know and we do not.