Compatibility

Compatibility is determined by the PCI device IDs the module accepts, not by the badge on the cooler or the chip family alone. The module accepts GeForce cards built on GA102, GA103, GA104 and GA106 — and within that, a shorter list of cards has actually been tested and reported by users.

Data: gpu-unlock README "Tested cards" table and the module source (kernel/gpu_unlock/main.c), pinned at commit af28cbb · Checked 2026-10-25

Chips and the retail cards built on them

ChipGeForce retail cardsStatus in the project
GA102 RTX 3090 Ti · RTX 3090 · RTX 3080 Ti · RTX 3080 (12 GB / 10 GB) Tested Multiple models unlocked and reported.
GA103 One desktop device ID only (10de:2414); mobile GA103 parts are not in the list Accepted, no reports Device IDs are in the module; nobody has filed a result yet.
GA104 RTX 3070 Ti · RTX 3070 · RTX 3060 Ti Tested RTX 3070 unlocked and reported; 3070 Ti not yet tested with this module.
GA106 RTX 3060 12 GB (tested) · RTX 3060 8 GB (no reports yet) Tested RTX 3060 12 GB unlocked and reported.
GA107 RTX 3050 Not covered The chip is not in the module's accepted list.
Ada (AD10x) RTX 40 series Different project See the gpu-unlock-ada fork.

The module accepts every GeForce device ID listed in its source (kernel/gpu_unlock/main.c); the table above maps those chips to the cards you can actually buy.

Accepted is not tested

The source list is much wider than the tested list. In main.c at commit af28cbb, supported_ids[] contains 12 GA102 IDs, 29 GA104 IDs, one GA103 ID (0x2414) and nine GA106 IDs. The comments narrow that sharply: GA103 is marked “(untested)”, and GA106 is “validated only on RTX 3060 12GB 0x2504”. The list — drawn from pci.ids — also deliberately excludes audio functions and professional (GL) parts — which already run BF16 at full rate — and, for GA103 and GA106, excludes mobile parts.

So check the exact device ID, not just the name on the box. Two cards can share a retail name or chip family while carrying different device IDs; conversely, an accepted ID only proves the module will try that function, not that anyone has filed a result for that board. When in doubt, compare lspci -nn output with both lists: the accepted IDs in main.c and the reported cards in the README.

Card-by-card status — as documented upstream

Transcribed from the gpu-unlock README. "Result" and gains are as reported by the users who filed them; each report used its own benchmark, so gains are not directly comparable across rows — and the power limit matters a lot (the lowest RTX 3090 figure, +25%, came from a card capped at 225 W). The 3090 Ti's +80% in the table is its 450 W run; the lower +68% sometimes quoted for it comes from a report at unequal power limits (300 W vs 380 W) and is not a controlled unlock-only gain — it is broken down on the benchmarks page.

ChipCardPCI IDResultBF16 GEMM gain
GA102RTX 3090 Ti10de:2203Unlocked+80% (450 W)
GA102RTX 309010de:2204Unlocked+25–67%
GA102RTX 3080 Ti 20 GB (OEM, GA102-250)10de:2205Unlock completes BF16 stays at half rate—
GA102RTX 3080 10 GB10de:2206Unlocked+70–73%
GA102RTX 3080 20 GB (memory upgrade mod)10de:2206Unlocked+56–59%
GA104RTX 307010de:2484Unlocked+73–81%
GA104RTX 3070 Ti10de:2482Not tested with this module — the same unlock worked from a UEFI port on Windows—
GA106RTX 3060 12 GB10de:2504Unlocked~+80%
GA103any10de:2414No reports yet—

Source: github.com/Zagitta/gpu-unlock README, "Tested cards" (commit af28cbb) — each row links to its underlying user report there. For what these gains mean in context — power limits, workloads, and the numbers that did not move — see the benchmarks page.

The one known exception

The OEM RTX 3080 Ti 20 GB built on GA102-250 is the only tested card where the sequence completes but the BF16 rate stays at half. If your card is a retail 12 GB 3080 Ti it is a different variant — but note it has no report in the tested list yet either, so treat it as "chip covered, outcome unconfirmed" rather than guaranteed.

Setups reported working

Not sure which chip your card has?

Run lspci -nn | grep -i nvidia and compare the PCI ID against the table above, or use the quick-check on the home page. If your card works (or doesn't), the project wants your compatibility report either way.