rented a 96GB Blackwell to test LTX-2.5, 231s for 10 seconds of 1080p with audio
retained text
https://reddit.com/link/1wo66x5/video/adhgop21t9rh1/player Rented a Blackwell RTX PRO 6000 (96GB) for an evening to see what the LTX-2.5 template actually costs in wall clock. Stock video\_ltx2\_5\_t2v workflow, no custom nodes, nothing tuned. https://preview.redd.it/ekn39gf9u9rh1.png?width=1389&format=png&auto=webp&s=13961c806e9961a605f819c40cc715cf9e43c482 Box was a single RTX PRO 6000 Blackwell Server Edition, 96GB VRAM, 22 vCPU, 110GB system RAM, driver 580.95.05 on CUDA 13.0, torch 2.8.0+cu128. Two runs, same workflow https://preview.redd.it/snpbc60ou9rh1.png?width=1901&format=png&auto=webp&s=2de1f3a3fcf07846962f64ebf0a36efa62f842ff just duration and megapixels changed: * | run | length | resolution | pass 1 (8 steps) | pass 2 (3 steps) | total | * | 1 | 5s | 1280x736 | 1.17 s/it = 9s | 4.84 s/it = 14s | 69.4s | * | 2 | 10s | 1920x1088 | 5.68 s/it = 45s | 28.50 s/it = 85s | 231.5s | Run 1 includes the cold load of \~36GB of weights off disk. Run 2 cost me about 7 cents of rental time. VRAM and power on the same two runs: * | value * | idle, weights resident | 39.0 GB * | peak during 10s 1080p | 52.5 GB * | card total | 96 GB * | peak power | 530W of 600W * | peak util | 100% Both runs came out as a single mp4 with a real AAC track, 48kHz stereo. ffprobe shows two streams. Audio and video come out of the same latent, there's a ConcatAVLatent before sampling and a SeparateAVLatent after, so it isn't a soundtrack bolted on at the end. In the clip the rain and the train horn I asked for both actually landed. The part I didn't expect is the VRAM. Idle with the weights loaded is already 39GB, and ten seconds of 1080p with audio only added about 13GB on top. Length and resolution aren't what eat your card, the 21GB transformer plus the 15GB Gemma text encoder is. On a 32GB card you're fighting the weights, not the clip. Second thing: at 1080p, sampling was 130s of the 231s. The rest was VAE decode and mux. At low res decode is noise, at high res it's almost half the run. Scaling, for reference: pixel-frame load went up 4.4x between the two runs, pass 1 went up 4.9x, pass 2 went up 5.9x. The refine pass after the 2x latent upscale is the one that hurts. One gotcha that cost me a download: the template's CLIPLoader points at gemma4\_e2b\_it\_int8\_convrot.safetensors, which the notes say is only needed when prompt\_enhance is on. ComfyUI validates it statically anyway and the node stays red, so either pull the 4.9GB file or just Ctrl+B the node. Worth saying what this run does not tell you. I wrote the prompt with her back to camera, so she stays turned away the whole shot and I never got a face out of it. No idea how it handles close-up faces or hands, and that's usually where these models fall over. Anyone pushed this past 10 seconds? Genuinely curious where it breaks, because at this size it clearly isn't VRAM.
document fetched 2026-09-24 00:44 · canonical https://reddit.com/r/comfyui/comments/1wo66x5/rented_a_96gb_blackwell_to_test_ltx25_231s_for_10