GPU Unlock 1.1.1 — TF32 and INT4 evidence

GPU Unlock 1.1.1 publishes a reproducible SM89 tensor issue-rate probe and matched stock/unlocked TF32 and unsigned INT4 results on one RTX 4070 Ti SUPER (10de:2705, NVIDIA 580.178.04, 240 W cap). The Ada unlock transaction is unchanged from 1.1.0; the module and DKMS version metadata advance to 1.1.1.

Register-resident mode	Stock	Unlocked
FP16 inputs, FP32 accumulate (control)	94.96 TF/s	188.97 TF/s
TF32 inputs, FP32 accumulate	47.49 TF/s	94.51 TF/s
Unsigned INT4 inputs, INT32 accumulate	754.81 TOP/s	754.68 TOP/s

The same cubin, card, driver, 240 W cap, and launch settings were used in both boots. Each mode checked 270,336 output cells with zero mismatches per state. TF32 instruction issue rate gained 1.99×; unsigned INT4 did not improve. These all-one, register-resident loops do not establish an application GEMM or inference speedup. INT8 and one previously tested FP8 GEMM were also unchanged.

Read the canary report, sanitized numeric output and hashes, and GPL benchmark source and reproduction steps. Installation and boot rollback are in the Quick Start.

The exact source-only archive ran 22 Linux host tests with 8 optional-input skips, and its kernel module compiled against the tested host's 5.15.0-177-generic headers as version 1.1.1. The owner miner stayed running through this CPU-only validation. The downloadable archive contains no NVIDIA firmware binary; pinned NVIDIA inputs are generated locally under NVIDIA's separate license.

Additional FP8 evidence — October 23, 2026

The earlier unchanged FP8 result was the selected staged-accumulation kernel. A later custom harness using NVIDIA CUTLASS 4.8.1's fast-accumulation GEMM measured 158.54 -> 309.08 dense TFLOP/s (1.949527x) for 8192x8192x8192 on the same RTX 4070 Ti SUPER, same CUBIN and same 240 W cap. The requested core ceiling was 2400 MHz; unlocked core boundary samples were 2265/2370 MHz versus 2400/2400 MHz stock, with normal 10251 MHz memory in both states. This is a same-power-cap measurement, not clock-normalized throughput.

All 19 matched kernel/shape cases passed full-output cuBLAS pedantic FP32 checks before and after timing. The predeclared inputs were independent deterministic +/-1 matrices, K <= 8192. The original staged profiler still gained only about 0.7%; the faster result uses the identical fast kernel in both hardware states. This does not establish arbitrary-input fast-accumulation accuracy, a 2x full-model gain, or Pearl mining performance.

Separately, both boots reran the original Ada FP8 numerical-profile suite: 35 suites, 3,227,648 FP32 output cells, zero bitwise differences between stock, unlocked and the archived Ada results. These rounding-sensitive tests preserve the observed Ada 16+16 model on the tested inputs; Ada still requires its own Pearl verifier/circuit profile.

Evidence assets:

• cutlass-fp8-stock-vs-unlocked.zip — SHA-256 b9468068a25cc208749eb0004fc86e15ac98c25c06517fc7c76d3a4f8ac6e861
• ada-fp8-profile-stock-unlocked-evidence.zip — SHA-256 5e9fa282298b996828e9554a47bc32c2da9ac65037fe50368e80f5bb95bac05c
• FP8 evidence checksums

The CUTLASS package includes source, binaries, disassembly, all timings, telemetry, correctness results and the documented atomic-scope compatibility patch. The profile package includes all 70 raw output files, source, validators and manifests; the input generator is linked.

GPU Unlock 1.1.0 — experimental Ada 2705

GPU Unlock 1.1.0 adds an experimental, exact-device Ada path for RTX 4070 Ti SUPER (10de:2705) alongside the RTX 4090 and existing Ampere paths.

On one 4070 Ti SUPER at a 240 W cap, FP16/FP32 tensor issue rate rose from 95.0 to 189.9 TF/s and BF16/FP32 from 95.0 to 189.7 TF/s. A matched, verified CUTLASS E4M3 FP8/FP32 GEMM measured 59.326 TF/s stock and 59.495 TF/s unlocked, effectively unchanged for that kernel. INT8 also remained unchanged. The Ada unlock succeeded on two full AC cold boots of this card; broader hardware support and long-duration stability remain unproven.

The downloadable tarball and repository contain GPL-2.0-only source, build recipes, and sanitized evidence. They contain no NVIDIA firmware binary. The build downloads SHA-pinned NVIDIA inputs and generates firmware locally under NVIDIA's separate license.

Start with the copy-paste Quick Start. See the canary report and raw benchmark output and hashes.

Boot-persistence update (October 23, 2026): On the tested 4070 Ti SUPER, the same Ada initramfs successfully reapplied the unlock after an ordinary warm reboot. A second warm reboot selected that entry from a persistent systemd-boot default, again returned full tensor-rate readbacks, and restored owner mining with an accepted share. This is one host and card; other systems may require a full power cut. See the updated boot instructions and selected boot evidence. The v1.1.0 source tarball and tag are unchanged.