Which releases give gigabytes and name no card
Eight of the 15 families name a card or an architecture beside their memory figure. Four publish gigabytes with no hardware named at all, and three publish neither a figure nor a card. As of 2026-09-12.
| Item | What is published | Detail |
|---|---|---|
| A card model named | Six families | An RTX 4090 twice, an A100 80GB, RTX 30XX to 50XX, at least one H100, and H100 or H800 in counts of one, four and eight |
| An architecture named | One family | Ampere, Hopper and Blackwell, with 32.54 GB |
| A capacity but no model | One family | A single 80G GPU, behind two peak figures |
| Gigabytes and no hardware | Four families | From 4GB in FP16; 78.55 GB peak; 9.3G with offload; about 14.7GB at 540P |
| Neither figure nor card | Three families | No memory statement of any kind |
Inclusion rule. One row per way of identifying hardware. A row is kept where a single family uses that form, because the form is what decides whether a figure transfers. Order. From the most specific identification to none at all.
1A card with the same memory is not the same card
A figure of 24GB matches several generations of hardware, and only some of them support the numeric formats a modern release assumes. A card that satisfies the number and not the format does not run the model.
One repository here names card generations and the formats together, which closes that gap in one sentence. Another names three chip architectures instead, which covers cards that did not exist when the page was written.
2Why the four bare figures are still worth having
Three of the four are attached to a resolution, and two are attached to a frame count or a precision, so they are not unqualified. What they leave open is the hardware side rather than the workload side.
A reader can therefore use them to compare workloads and not to check their own machine. That is half of what the column is for, and it is worth distinguishing from the rows that offer neither half.
3The register quotes the form, it does not normalise it
Converting a named card into gigabytes would lose the formats and the bandwidth. Converting gigabytes into a card name would invent hardware nobody mentioned. Neither conversion happens here.
So the cells are uneven on purpose, and the unevenness is the answer to this question. A tidy column would have required the register to make up the missing halves.
- Hardware78.55 GB peak GPU memory at 768px768px204f, 77.64 GB at 544px992px204f and 72.48 GB at 544px992px136fstated per configuration at batch size one
- Hardware9.3G BF16 with cpu_offload, against 27.5G if CPU offload is not enabledsingle GPU memory usage on the same run
- HardwareAbout 14.7GB peak VRAM for a 540P video on the 1.3B model, and around 51.2GB on the 14Bthe same output size on two model sizes
- HardwareThis model requires 32.54 GB of GPU VRAM, on Ampere, Hopper or Blackwell hardwarea figure and the supported architectures
- HardwareNvidia GPU in RTX 30XX, 40XX, 50XX series that supports fp16 and bf16; the GTX 10XX/20XX are not testedcard generations and numeric formats
4Sources
Repository statements and hosted rates are linked from the model pages and the how read page, checked 2026-09-12. Hardware figures are set out on requirements. Nearby: Frames, no rate, Exact or rounded, Complete output.