The Ampere module doesn't touch Ada cards — but a fork does. gpu-unlock-ada ports the same pre-driver speed-select override to Ada Lovelace, with a demonstrated result on the RTX 4090 and early, explicitly experimental support for two more cards.
Based on gpu-unlock-ada README v1.1.1 · Checked 2026-10-25 · Project on GitHub
Project-reported On the RTX 4090 (AD102, PCI ID 10de:2684), the fork reports approximately 2× FP16 and BF16 tensor throughput with FP32 accumulation. The override survived PCIe reset (FLR), stock driver initialisation and CUDA benchmarks; INT8, dp4a and scalar FP32 stayed at baseline.
| Mode (FP32 accumulate) | Before | After | Change |
|---|---|---|---|
| FP16 tensor | 165.6 TFLOPS | 331.6 TFLOPS | ~2× |
| BF16 tensor | 165.5 TFLOPS | 330.6 TFLOPS | ~2× |
A matched run by a card owner, linked from the fork’s own README, reproduced the transaction and found the same shape at a 450 W cap (FP16 160.4 → 320.4 TFLOPS, BF16 159.7 → 320.2 TFLOPS), with all numerical checks passing. (This site has not run any of these benchmarks — every figure on this page is the fork’s published result or a report its README links.) Read the small print, though: in that same matched testing, real cuBLAS GEMM throughput gained a much more modest 12–34% (FP16) and 16–20% (BF16), and was essentially unchanged at 4096³. The 2× figure is a register-level instruction issue rate, not an application speedup — the fork's own documentation says so.
| Card | Chip / PCI ID | Status | Evidence |
|---|---|---|---|
| RTX 4090 | AD102 · 10de:2684 |
Demonstrated | ~2× FP16/BF16 issue rate (165.6→331.6 / 165.5→330.6 TFLOPS); reproduced in an owner’s matched run linked from the README; real GEMM gains +12–34%. |
| RTX 4090 D | AD102 · 10de:2685 |
Experimental opt-in | Uses the AD102 path with the extra flag gpu_unlock.rtx4090d_experimental=1. |
| RTX 4070 Ti SUPER | AD103 · 10de:2705 |
Experimental — one card | FP16/BF16 issue rate 95.0 → 189.9 / 189.7 TFLOPS at 240 W; TF32 47.49 → 94.51 TFLOPS; outputs matched exactly; one FP8 GEMM showed no meaningful gain (+0.3%). Requires the experimental flag and an exact PCI address — targets=all never selects it. |
| RTX 4060 Ti 16 GB | AD106 · 10de:2805 |
Experimental — one card | One cold boot passed every gate and the rate readbacks flipped from half to full — but no same-card stock throughput run has been made yet, so there is no honest gain figure. |
| Other Ada cards | — | Not accepted | Other Ada IDs, professional RTX A-series and data-centre parts are not accepted by the module. Shared firmware across Ada chips does not establish support for untested cards. |
0x88888888 / SS1 0x8 selector writes before the stock driver binds.Calibrate your expectations
For Ada, the demonstrated, replicated result is "the tensor issue-rate ceiling doubles on a 4090". Application-level gains measured so far are far smaller (tens of percent in GEMM, near zero in one FP8 case), two of the three supported cards rest on single-card validation, and there is no gaming data. This is promising, fast-moving, early-stage work — interesting to follow and test, not a finished product.
The Ada repository also documents an experimental Windows UEFI reference port for one specific MSI RTX 4070 Ti (AD104) board, contributed by Latence.app: one successful manual boot with matched cuBLAS gains of +80.1% (FP16) and +76.4% (BF16). It is tied to a single captured motherboard profile, distributes no installer or binaries, and does not make AD104 a supported target of the Linux module. Interesting proof of concept — not a Windows release.
Primary source: github.com/JustAResearcher/gpu-unlock-ada (README v1.1.1 and its linked evidence reports), pinned at commit 4e7fefc (2026-10-24). For the Ampere original, see the install guide and benchmarks.