Floorsweep read of the 8 GB card: gated versus defective partitions
More actions
Reported 8 GB
The 8 GB card's disabled memory partitions and compute clusters are software-gated, not all physically dead, but the gate that would re-enable them is fuse-locked. The reversible unlock ceiling of 64 GB and 74 SM therefore stands. This settles, for a set of eight cards, whether the 80 or 96 GB claim has a physical basis: about 80 GB of framebuffer is physically latent, but 96 GB is not reachable and even 80 GB would need an irreversible fuse burn the 8 GB card does not allow in software.
How it was read
The floorsweep fuses, FBPA_CFG1 and the software-override control registers were read with gpcprobe, the read-only kernel module by @matstatman that maps BAR0 while the NVIDIA driver stays bound: no writes, no driver unbind, no reboot. The cards were read with the volatile 64 GB unlock active, as applied by cmpunlocker (@amoghmunikote). Eight 8 GB cards (10de:20c2) were on the chassis; seven were floorsweep-read in this pass. Nothing was written to any card and no fuse was changed.
Result (aggregate, seven cards)
Every one of the seven cards shows the same shape.
| Unit | Active | Gated, not dead | Defective | Total |
|---|---|---|---|---|
| Framebuffer partition (FBP) | 8 | 2 | 2 | 12 |
| Compute cluster (GPC) | 5 | 2-3 | 0-1 | 8 |
| FBPA / FBIO / ROP | 16 | 4 | 4 | 24 |
- SM: 74 (37 active TPC) on every card, the known post-unlock ceiling. No card exceeds it.
FBPA_CFG1 = 0x02779000: the 64 GB volatile unlock is active.CTRL_OPT_FBP = 0,CTRL_OPT_FBPA = 0: no live or volatile override is in effect, and none is available. The status shadows equal the fused values, so the floorsweep above is what the hardware actually runs.
Source: @darkmaniac7, 2026-11-05 (lab notes)
Source: repo github.com/matstatman/gpcprobe by matstatman on GitHub
Source: repo github.com/amoghmunikote/cmpunlocker by amoghmunikote on GitHub
What it means
- The gated-not-dead partitions are real and consistent. About 2 framebuffer partitions and 2 to 3 compute clusters per card are software-disabled, not physically defective, and the pattern repeats across all seven cards.
- About 80 GB is physically latent. The 8 active FBPs are the 64 GB in use; the 2 gated FBPs are roughly 16 GB more, so about 80 GB of framebuffer physically exists on the die. The "80 GB" figure has a real physical basis.
- The enable path is non-volatile. Reaching the gated partitions needs the one-time-programmable software-override-disable fuse (
FUSE_DISABLE_SW_OVERRIDE_STATUS) cleared, plus the factory-burned RIR rows worked around.CTRL_OPT_FBP = 0confirms there is no volatile register to flip. The only path is burning a fuse: irreversible, with real brick risk, and outside any reversible unlock. See RIR fuse burn. - The reversible ceiling stays 64 GB / 74 SM. Any volatile or software method tops out here, and cmpunlocker already reaches it.
- 96 GB is not supported. 2 FBPs per card are genuinely defective (dead silicon), so even a fuse burn could reach at most about 80 GB (10 of 12 FBPs), never the full 12.
A gated-not-dead partition is not the same as recoverable memory on this card. Whether populated DRAM even sits behind the gated channels is a separate question that a volatile poke plus a dense unique-data write and read-back over all reported bytes would answer, and that was not part of this read-only pass. The honest result: the gated partitions exist and are consistent across the cards, but they are not software-openable on the 8 GB card, so no extra usable capacity was obtained.