Does this unlock Token2Any?
No. It does not change whether RTX 30 can enable Token2Any in games; the reports concern FP/BF16 GEMM compute throughput, not game frame rates.
GPU unlock RTX 30 / 4090: FP/BF16 community reports and limits
This does not unlock Token2Any.This concerns FP/BF16 instruction issue rates and matrix-operation (GEMM) throughput for workloads such as local AI inference/training. It does not change whether RTX 30 can enable Token2Any in games and is not a claim of improved game frame rates.
We have not reproduced this unlock procedure on our machines. Our game-character style conversion and Studio image pipeline are separate and are not measurement evidence for this unlock procedure.
Checked: . This page collects reference material; it does not deliver an implementation tutorial.
- Zagitta/gpu-unlock: Linux RTX 30 Series kernel module.
- elias-gaeros/gpu-unlock, rtx4090 branch: Linux RTX 4090 branch.
- impacifist/Windows-GPU-UNLOCK: Windows / UEFI paths for exact machine configurations.
What changes
According to the main Linux repository README, the module writes speed selectors before the NVIDIA driver binds to the card: SS0 0x88888888 and 4-bit SS1 0x8, then performs FLR. The README describes this as doubling the issue rate of FP/BF16 FMA into FP32; this is not our measurement. The module approach does not require patching again after every driver update.
RTX 30 Series: cards with community reports
The following dense GEMM community table is from Zagitta README. Community measurements reported by the repository, not measurements by this site; no numbers are filled in for unreported results.
| Card | Chip | Result | BF16 GEMM gain | Report number and attribution |
|---|---|---|---|---|
| RTX 3090 Ti | GA102 | Unlocked | +68–80% | Zagitta README · #1 Community measurements reported by the repository, not measurements by this site |
| RTX 3090 | GA102 | Unlocked | +25–67% | Zagitta README · #5、#9、#10、#17、#25、#27、#33 Community measurements reported by the repository, not measurements by this site |
| RTX 3080 Ti 20 GB (OEM) | GA102 | Unlock completed, but BF16 remains at half speed | — | Zagitta README · #28 Community measurements reported by the repository, not measurements by this site |
| RTX 3080 10 GB | GA102 | Unlocked | +70–73% | Zagitta README · #14 Community measurements reported by the repository, not measurements by this site |
| RTX 3080 20 GB (modified VRAM) | GA102 | Unlocked | +56–59% | Zagitta README · #10、#15 Community measurements reported by the repository, not measurements by this site |
| RTX 3070 | GA104 | Unlocked | +73–81% | Zagitta README · #10、#12 Community measurements reported by the repository, not measurements by this site |
| RTX 3070 Ti | GA104 | Not tested with this module (the same unlock succeeded with the Windows UEFI path) | — | Zagitta README · #32 Community measurements reported by the repository, not measurements by this site |
| RTX 3060 12 GB | GA106 | Unlocked | Approximately +80% | Zagitta README · #13、#31 Community measurements reported by the repository, not measurements by this site |
| Any GA103 model | GA103 | No reports yet | — | Zagitta README · No reports yet Community measurements reported by the repository, not measurements by this site |
Community measurements reported by the repository, not measurements by this site: the Zagitta README cites report #1 for an RTX 3090 Ti single-card baseline, median of five samples, with FP32 accumulation: FP16 87.46 → 118.39 TFLOPS; BF16 87.50 → 157.41 TFLOPS; TF32 42.58 → 82.61 TFLOPS. Source: main repository README, report #1。
RTX 4090: fork README comparison
Community measurements reported by the repository, not measurements by this site. Source: rtx4090 branch README; this comparison lists no issue report number, so none is invented here.
| Mode | Before | After | Report source and attribution |
|---|---|---|---|
| FP16 → FP32 accumulation | 165.6 TF/s | 331.6 TF/s | rtx4090 branch README; no report number listed Community measurements reported by the repository, not measurements by this site |
| BF16 → FP32 accumulation | 165.5 TF/s | 330.6 TF/s | rtx4090 branch README; no report number listed Community measurements reported by the repository, not measurements by this site |
The branch reports that INT8, dp4a and scalar FP32 remain approximately at baseline. The limitation is one card and one successful cold boot; independent cold-boot reproduction and application-level numerical validation are still pending. These are not our measurements or a promise of application-workload gains.
Three paths and prerequisites
The repository paths involve disabling Secure Boot and changing boot parameters. This page records conditions, not execution commands. Check each repository's documentation separately; one path's conditions do not replace another's.
- Main Linux 30 Series repository: covers only GA102/103/104/106. Architecture coverage does not mean every model has a successful report; GA103 has no reports yet.
- 4090 branch: integrates the published AD102 payload for RTX 4090 only (
10de:2684, AD102); other models, professional cards and data-center cards are not accepted. - Windows / UEFI: the repository lists three exact machine configurations, including headless and USB paths; a model name does not establish compatibility. v0.3.1 is a source prerelease without packaged firmware or private EFI. Do not treat Linux reports as verification of a Windows configuration.
Risks and rollback
Unofficial, newly appeared in the last few days, and involves low-level GPU operations; requires disabling Secure Boot and changing boot parameters; carries compatibility and stability risks. We do not recommend trying it directly on your primary machine. Keep a rollback path available before experimenting.
Our position
We do not provide, host, mirror or package firmware, payloads or EFI, and we do not promise results. We have not reproduced this unlock procedure on our machines.
Check the official Token2Any hardware scope；See real examples of our independent image style conversion. Neither path turns this page's compute-throughput reports into Token2Any game-compatibility conclusions.
FAQ
Will it work on my 3080 Ti?
A model name alone is not enough. Main repository report #28 says the RTX 3080 Ti 20 GB OEM unlock completed but BF16 remained at half speed; this cannot be generalized to other configurations.
Does it work on Windows?
A community UEFI path exists, but it requires an exact machine-configuration match; a model name does not establish compatibility. The Windows repository's v0.3.1 is a source prerelease without packaged firmware or private EFI.
Does it survive driver updates?
The Linux module executes before the NVIDIA driver binds to the card and does not require patching again after every driver update. Consult the corresponding repository documentation for the Windows UEFI path; the Linux conclusion cannot simply be applied to it.
Is it official?
No. This is a community approach. This site only collects reference material, does not represent NVIDIA and does not promise results.
Is it reversible?
We do not promise reversibility. Assess it using the corresponding repository documentation and retain a rollback path before trying it; this page provides no firmware, payload or EFI.