NodeValid

Published video-model weights, licences and hardware notes

Open-Sora: 60.3GB on one card, 44.3GB on four

The figures are 60.3GB peak at 768px on one GPU and 44.3GB across four or more, measured on H100 or H800 cards. Same weights, same output size, two machine shapes. As of 2026-09-22.

Open-Sora in the hardware column, as publishedPublishing a single-card figure and a multi-card figure for one output size is a measurement of what sharding costs and saves, rather than a claim about it.The memory statement as published60.3GB peak at 768px onone GPU44.3GB across four or more, on H100 or H800256px and 768pxThe pair the 11B model is stated to supportnum_frames of 4k+1, under129Frames, as the constraint the code enforcesSharding measured, not asserted
Fig. 1 What four cards buy over one, in the vendor's own numbers at the same output size.
Open-Sora in the hardware column, with the columns that qualify it. Read from github.com/hpcaitech on 2026-09-22.
ColumnWhat the pages state
Hardware, in the vendor's own words60.3GB peak at 768px on one GPU
The detail published beside it44.3GB across four or more, on H100 or H800
Resolution, from the same pages256px and 768px
Length, from the same pagesnum_frames of 4k+1, under 129

Inclusion rule. The value this page is about, the detail the vendor attached to it, and the two columns that change how it should be read. A column the vendor left empty keeps its row and says so. Order. This column first, then its detail, then the two columns that qualify it.

1Sharding does not divide the requirement

Four cards do not bring 60.3GB down to 15GB. They bring it to 44.3GB each, because activations and buffers are replicated while only part of the load is split.

That is the number most worth having published. A reader who assumed linear scaling would budget for a quarter of the memory and discover the shortfall after buying the cards.

2Naming the cards changes the claim

H100 or H800 names the interconnect as much as the memory. Splitting a model across four cards is bandwidth-bound, and the same figures on four cards without a fast link between them would not reproduce.

The register records the cards because the multi-card figure is meaningless without them.

3The output is held constant, and stated

Both figures are at 768px, which is one of the two resolutions the 11B model is stated to support. The other is 256px, with no memory figure attached.

Frames are described as a constraint the code enforces, 4k+1 and under 129, rather than as part of the memory configuration. So the pair varies the machine and nothing else.

4Sources

Every figure above is the vendor's own, taken from github.com/hpcaitech on 2026-09-22 with nothing measured here. The rest of the release is on its family entry, the whole column is on hardware, and all memory figures together are on requirements. Nearby: A stopwatch, not a floor, Six gigabytes, named cards.