Following the release of gpu-unlock, an unofficial tool for Linux, reports are mounting that AI workloads run faster on some GeForce RTX 30 series cards. In ComfyUI measurements on an RTX 3090 Ti, BF16 processing for the image generation model "MOSS-Image SFT" dropped from 71.98 seconds to 45.40 seconds, and BF16 processing for "MiniMax H3," which generates video and audio, fell from 253.27 seconds to 158.05 seconds. In both cases throughput rose by roughly 60%, which works out to a reduction of about 36.9% to 37.6% in the wait until generation finishes. Meanwhile, the "2x" figure cited by the developer refers to the issue rate of a specific math instruction, and some GPUs have been reported to show no BF16 performance improvement even after the setting change. The actual speedup varies with the numeric format, the workload, and the measurement conditions.
The "2x" refers to the issue rate of a specific math instruction
According to the developer's GitHub repository, gpu-unlock 1.0.0 targets the GA102, GA103, GA104, and GA106 GPU chips. It changes the configuration of the Tensor Cores, which handle matrix operations inside the GPU, so that instructions taking FP16 or BF16 inputs and accumulating the multiplication results in FP32 are issued at twice the rate. The "2x" here refers to the speed at which instructions are sent to the execution units. It does not mean that overall GPU compute performance or application speed doubles across the board.
It is also necessary to distinguish between the format of the numbers fed into a computation and the precision used when accumulating the multiplication results. Section 1.4.1.5 of NVIDIA's "Ampere Tuning Guide v11.6.1" also describes the two separately, stating explicitly that BF16 operations use only FP32 for accumulation. However, that document describes the specifications of the Ampere architecture and does not mean that NVIDIA has certified how gpu-unlock works.
The current implementation of gpu-unlock is provided as a Linux kernel module. It rewrites a register that selects the GPU's operating speed, then resets the device through PCIe Function Level Reset (FLR). This sequence must run before the unmodified NVIDIA driver takes control of the GPU. As a result, manually loading the module afterward on a GPU the driver is already using has no effect.
Matrix-operation speedups differ between FP16 and BF16
In Issue #1, posted to the developer's GitHub on October 18, 2026, applepines reported benchmark results using an RTX 3090 Ti. The test conditions were a 450W power limit, an 8192×8192×8192 dense matrix multiplication, and FP32 accumulation, comparing the median of five runs for each condition. The results varied widely by numeric format, with BF16 and TF32 recording especially large improvements.
| Numeric format | Before unlock | After unlock | Performance gain |
|---|---|---|---|
| FP16 | 87.46 TFLOPS | 118.39 TFLOPS | +35.4% |
| BF16 | 87.50 TFLOPS | 157.41 TFLOPS | +79.9% |
| TF32 | 42.58 TFLOPS | 82.61 TFLOPS | +94.0% |
The performance gain is calculated as (after ÷ before − 1) × 100 and rounded to one decimal place. Even on the same GPU, the gains differ greatly: 35.4% for FP16, 79.9% for BF16, and 94.0% for TF32. Note that TF32 is one of the formats measured in this benchmark and should be distinguished from the FP16/BF16 instruction issue rate the developer describes.
In addition, this initial measurement was made in an environment different from the currently published kernel module. The test used Ubuntu kernel 7.0.0-38-generic, a modified NVIDIA driver 610.43.02, and a custom initramfs hook. The poster also says that other graphics cards and integration into packages have not been tested. The results in this table therefore cannot be treated as the performance of the current gpu-unlock combined with an unmodified NVIDIA driver.
In ComfyUI, wait times for image and video generation fell by about 37%
Beyond matrix-operation benchmarks, there are also reports testing the speedup of whole generation pipelines using AI models. Reddit user under-_-null published a ComfyUI comparison on an RTX 3090 Ti, showing that the degree of speedup varies greatly by model. The test system was a Dell Precision 5820 with an EVGA FTW3 RTX 3090 Ti, with the GPU running at stock clocks and the power limit set to 450W both before and after the unlock. The OS was Ubuntu, the NVIDIA driver was 615.71.09, and ComfyUI was 0.42.1. The batch size was 1, and PyTorch SDPA was used for items where nothing else was specified.
Warm-up runs were excluded: 15 still images were generated for each condition, and 3 videos. The first 3 images and the first video were treated as warm-up, and the median was taken from the remaining measurements. Processing time includes text encoding, sampling, decoding, and saving, as well as any time spent on CPU offloading or data transfers. The time to load the model initially was excluded from measurement.
Comparing BF16 processing for MOSS-Image SFT and MiniMax H3, which showed especially large improvements, gives the following.
| Model / numeric format | Before unlock | After unlock | Throughput increase | Time reduction |
|---|---|---|---|---|
| MOSS-Image SFT BF16 | 71.98 s | 45.40 s | +58.6% | ~36.9% |
| MiniMax H3 BF16 (video + audio) | 253.27 s | 158.05 s | +60.2% | ~37.6% |
It is worth noting that a roughly 60% increase in throughput is not the same as a 60% reduction in processing time. The throughput increase is calculated as (time before ÷ time after − 1) × 100, and the time reduction as (1 − time after ÷ time before) × 100. In other words, the throughput increase uses the post-unlock time as its baseline, while the time reduction uses the pre-unlock time. In these results, MOSS BF16 throughput rose 58.6% while the wait fell by about 36.9%. Likewise for H3 BF16, a 60.2% throughput increase corresponds to a wait-time reduction of about 37.6%.
The figures in the table are from a single user comparing before and after on the same GPU without changing each model's settings. MOSS and H3 are different models, and the data are not meant for comparing their absolute speed or output quality.
Other models show even wider differences in speedup. MOSS INT8 barely changed, going from 27.31 seconds to 27.23 seconds, with a throughput increase of only 0.3%. LTX-2.5 BF16 went from 70.81 seconds to 60.63 seconds, a 16.8% increase in throughput. Meanwhile, an FP8 checkpoint of FLUX.1 dev recorded a 54.6% improvement, but that workload used BF16 math internally. In other words, even if a model file is stored in FP8 or INT8, that does not determine the numeric format actually used in computation or the degree of speedup.
These results show that workloads relying heavily on BF16 math can expect large improvements, but not every AI model speeds up in the same way. Furthermore, there are real-hardware reports of performance not improving even though the GPU's settings could be changed, so carrying a supported chip alone does not guarantee a benefit.
A case where BF16 performance did not improve despite a successful setting change
Issue #28, posted on October 24, 2026, reports that no BF16 improvement could be confirmed on an RTX 3080 Ti 20GB with a GA102-250. This GPU is an OEM or engineering product and differs from the GA102-225 in the 12GB RTX 3080 Ti sold at retail. The poster confirmed that the speed-select register value "SS0=0x88888888" was retained after running gpu-unlock, but the expected performance gain did not materialize.
In measurements with the GPU clock fixed at 1200MHz and power held to about 200W, FP16 reached 50.00 TFLOPS at 203W and BF16 reached 50.22 TFLOPS at 199W. The BF16-to-FP16 performance ratio was 1.004, meaning the compute performance remained nearly identical. However, this is not a direct before-and-after comparison; it is a measurement showing that BF16 compute performance did not reach the expected level even after the setting change. The poster suggests that BF16 compute speed may be limited by fuse settings inside the GPU, but the cause has not been determined.
GPU specifications are not the only factor affecting the improvement. Where an application spends its time also matters. In user reports compiled by the developer, prefill, which processes input to an LLM in bulk, saw throughput increases of 14–36%, while token generation, which tends to be memory-bandwidth-bound, and INT8 workloads showed almost no change. Even for dense matrix multiplication on the RTX 3090, gains ranged widely from 25% to 67%, with the lowest, 25%, measured at a 225W power limit. However, because benchmark conditions differ between reports, these figures cannot be used to rank GPUs by performance.
Some verification of output consistency has also been done. According to the ComfyUI tester, in 150 pairs of still images compared before and after the unlock, all decoded RGB pixels matched, and 6 audio pairs also produced identical results. H3's video frames did not match perfectly, but because differences also appeared when generation was repeated within the same post-unlock session, gpu-unlock cannot be singled out as the cause. In addition, the temperature measurements did not control for starting GPU temperature or room temperature, and neither VRAM junction temperature nor GPU hotspot temperature was recorded.
These results alone therefore cannot be said to guarantee anything about the effect on output quality in general, lower GPU temperatures, or long-term operational stability. There are no real-hardware reports for GPUs with the GA103, and the RTX 3070 Ti has not been tested with the current Linux module. Nor can the effect on games, DLSS, speedups on newer GPU generations, or power savings be judged from these measurements.
What this round of reports makes clear is that certain AI workloads can see substantial speedups even on existing GeForce RTX 30 series cards, while the effect varies greatly with the GPU's specifications, the numeric format, and the workload. Going forward, it will be important to compare execution times before and after the setting change using the same GPU, power conditions, and workload. If measurements that distinguish BF16 from INT8, and LLM input processing from token generation, accumulate, it should become clearer which uses gpu-unlock is actually effective for.