← sources

I tested the RTX 30-series GPU Unlock: up to 59% higher throughput in ComfyUI (3090 Ti)

source: reddit (reddit:StableDiffusion)published 2026-10-23 20:02ingested 2026-10-24 02:38

extract

Everybody's been glazing [Zagitta’s GPU Unlock](https://github.com/Zagitta/gpu-unlock), so I figured I’d test it on my 3090 Ti. Overall, I’m pretty impressed. Can't complain about free performance. **Results (median time per image, or per clip for H3)** |Model / configuration|Stock|Unlocked|Throughput gain| |:-|:-|:-|:-| |MOSS-Image SFT BF16|71.98 s|45.40 s|**+58.6%**| |MOSS-Image SFT INT8|27.31 s|27.23 s|\+0.3%| |Qwen-Image 2.1 Turbo BF16|8.51 s|5.41 s|**+57.2%**| |Z-Image Turbo BF16|6.79 s|4.35 s|**+56.1%**| |FLUX.1 dev FP8 checkpoint\*|31.96 s|20.66 s|**+54.6%**| |Z-Image Turbo BF16 + SageAttention|6.47 s|4.38 s|**+47.7%**| |Qwen-Image 2.1 Turbo INT8|3.92 s|3.75 s|\+4.6%| |Qwen-Image 2.1 Turbo INT8 + SageAttention|3.92 s|3.76 s|\+4.3%| |Krea 2 Turbo BF16|15.65 s|10.70 s|**+46.3%**| |Krea 2 Turbo INT8|6.50 s|5.98 s|**+8.7%**| |MiniMax H3 BF16 (video + audio)|253.27 s|158.05 s|**+60.2%**| |MiniMax H3 INT8 ConvRot (video + audio)|155.07 s|118.21 s|**+31.2%**| All 150 matched stock/unlocked still-image pairs had identical decoded RGB pixels. See the H3 caveat below. *\*The FLUX FP8 checkpoint used BF16 computation in this pipeline.* These percentages describe throughput gains (e.g. MOSS BF16’s 58.6% higher throughput translates to approximately 36.9% less time per image). **Setup and testing** * **Hardware:** RTX 3090 Ti, stock clocks, factory 450W power ceiling. * **Software:** Ubuntu, NVIDIA 615.71.09, ComfyUI 0.42.1. PyTorch SDPA unless SageAttention is specified. * **Batch size:** 1. * **Resolution / steps:** * Z-Image Turbo, Qwen-Image 2.1 Turbo and Krea 2 Turbo: 1024×1024, 8 steps. * FLUX.1 dev: 1152×896, 30 steps. * MOSS-Image SFT: 1024×1024, 50 steps; Euler/simple, CFG 3, shift 5, layer-22 SLG. * MiniMax H3: 864×480, 124 frames at 24 fps (\~5.17 seconds), with audio, 20 steps. * **Method:** 15 measured images per configuration per condition, after three excluded warmups. H3 used three measured clips after one excluded warmup. Same prompts, seeds, model files and workflows across stock/unlocked runs. * **Timing:** Median complete request time, including text encoding, sampling, decoding and saving. Initial model loading excluded; CPU offloading and transfers included where used. Unfortunately, I only have a 3090 Ti to test. I’d expect a similar pattern on other supported 30-series cards, although the percentage gains will depend on the card and workflow. If anyone tests this on other cards or workflows the results would be appreciated. **Edit:** Added Krea 2 Turbo and MiniMax H3 results. H3’s video frames were not pixel-identical between stock and unlocked runs, but repeated renders within the same unlocked boot also differed, so those differences can’t be attributed solely to the unlock. All six matched audio pairs were identical.

retained text

Everybody's been glazing [Zagitta’s GPU Unlock](https://github.com/Zagitta/gpu-unlock), so I figured I’d test it on my 3090 Ti. Overall, I’m pretty impressed. Can't complain about free performance.

**Results (median time per image, or per clip for H3 and LTX-2.5)**

|Model / configuration|Stock|Unlocked|Throughput gain|
|:-|:-|:-|:-|
|MOSS-Image SFT BF16|71.98 s|45.40 s|**+58.6%**|
|MOSS-Image SFT INT8|27.31 s|27.23 s|\+0.3%|
|Qwen-Image 2.1 Turbo BF16|8.51 s|5.41 s|**+57.2%**|
|Z-Image Turbo BF16|6.79 s|4.35 s|**+56.1%**|
|FLUX.1 dev FP8 checkpoint\*|31.96 s|20.66 s|**+54.6%**|
|Z-Image Turbo BF16 + SageAttention|6.47 s|4.38 s|**+47.7%**|
|Qwen-Image 2.1 Turbo INT8|3.92 s|3.75 s|\+4.6%|
|Qwen-Image 2.1 Turbo INT8 + SageAttention|3.92 s|3.76 s|\+4.3%|
|Krea 2 Turbo BF16|15.65 s|10.70 s|**+46.3%**|
|Krea 2 Turbo INT8|6.50 s|5.98 s|**+8.7%**|
|MiniMax H3 BF16 (video + audio)|253.27 s|158.05 s|**+60.2%**|
|MiniMax H3 INT8 ConvRot (video + audio)|155.07 s|118.21 s|**+31.2%**|
|LTX-2.5 distilled BF16 (video + audio)|70.81 s|60.63 s|**+16.8%**|
|LTX-2.5 distilled INT8 ConvRot (video + audio)|48.02 s|43.02 s|**+11.6%**|

All 150 matched stock/unlocked still-image pairs had identical decoded RGB pixels. See the H3 caveat below.

*\*The FLUX FP8 checkpoint used BF16 computation in this pipeline.*

These percentages describe throughput gains (e.g. MOSS BF16’s 58.6% higher throughput translates to approximately 36.9% less time per image).

**GPU core temperatures - averages during measured renders**

|Model|Stock|Unlocked|
|:-|:-|:-|
|MOSS-Image SFT BF16|77.9°C|78.7°C|
|Qwen-Image 2.1 Turbo BF16|74.2°C|72.7°C|
|Z-Image Turbo BF16|73.7°C|66.6°C|
|FLUX.1 dev|75.9°C|77.6°C|
|Krea 2 Turbo BF16|74.6°C|73.3°C|
|MiniMax H3 BF16|79.2°C|78.5°C|
|MiniMax H3 INT8 ConvRot|77.4°C|78.4°C|
|LTX-2.5 BF16|70.9°C|69.5°C|
|LTX-2.5 INT8 ConvRot|70.6°C|71.5°C|

Same EVGA FTW3 RTX 3090 Ti inside a Dell Precision 5820, with a 450W ceiling in both conditions. Warmups excluded. Highest sampled core temperatures: **81°C stock / 80°C unlocked**. GPU fans adjusted automatically, peaking at **80% / 79%**.

**Caveat:** Starting temperatures and ambient conditions weren’t controlled, and unlocked renders finished sooner. These results don’t establish that unlocking makes the card cooler. I did not record VRAM junction and hotspot temperatures.

**Setup and testing**

* **Hardware:** EVGA FTW3 RTX 3090 Ti, stock clocks, factory 450W power ceiling.
* **Software:** Ubuntu, NVIDIA 615.71.09, ComfyUI 0.42.1. PyTorch SDPA unless SageAttention is specified.
* **Batch size:** 1.
* **Resolution / steps:**
   * Z-Image Turbo, Qwen-Image 2.1 Turbo and Krea 2 Turbo: 1024×1024, 8 steps.
   * FLUX.1 dev: 1152×896, 30 steps.
   * MOSS-Image SFT: 1024×1024, 50 steps; Euler/simple, CFG 3, shift 5, layer-22 SLG.
   * MiniMax H3: 864×480, 124 frames at 24 fps (\~5.17 seconds), with audio, 20 steps.
   * LTX-2.5: 1280×704, 121 frames at 24 fps (\~5.04 seconds), with audio, 8+3 sampling steps and native 2× latent upscaling.
* **Method:** 15 measured images per configuration per condition, after three excluded warmups. H3 and LTX-2.5 used three measured clips after one excluded warmup. Same prompts, seeds, model files and workflows across stock/unlocked runs.
* **Timing:** Median complete request time, including text encoding, sampling, decoding and saving. Initial model loading excluded; CPU offloading and transfers included where used.

Unfortunately, I only have a 3090 Ti to test. I’d expect a similar pattern on other supported 30-series cards, although the percentage gains will depend on the card and workflow.

If anyone tests this on other cards or workflows the results would be appreciated.

**Edit:** Added Krea 2 Turbo and MiniMax H3 results. H3’s video frames were not pixel-identical between stock and unlocked runs, but repeated renders within the same unlocked boot also differed, so those differences can’t be attributed solely to the unlock. All six matched audio pairs were identical.

**Edit 2:** Added LTX-2.5 results.

**Edit 3:** Added recorded core temps during runs. 

document fetched 2026-10-24 10:34 · canonical https://reddit.com/r/StableDiffusion/comments/1xdoh8k/i_tested_the_rtx_30series_gpu_unlock_up_to_59