GPU Benchmark! 4070 4080 4090 3080 3090

[metamountain]

Could we Do a benchmark for GPU´s ?
Need to get new hardware .
To keep it simple we could use the default load worklow on 1024*1024 with sdxl 1.0.
Seed 1
I know that ram comes in play when workflows get more complex , but for the start simple ist best.
Please post the second run. first run ist a little bit corupted by loading data.

theory pudget:

toms hardware 1111:

[metamountain]

2070 rtx mobile
100%|█████████████████████████████████████████████████████████████████████████████████████████████████| 20/20 [00:19<00:00, 1.05it/s]
Requested to load AutoencoderKL
Loading 1 new model
Prompt executed in 24.22 seconds

[epaulekas]

Nvidia RTX 4090 FE

100%|██████████████████████████████████████████████████████████████████████████████████| 20/20 [00:03<00:00, 6.17it/s]
Prompt executed in 3.61 seconds

(edit: used default resolution first go around)

[BBCreativo]

NVIDIA RTX5060TI 16GB

[Robson1970]

CPU vs GPU 1024x1024 2 steps

Intel 13900KS 24,25 sec.
Nvidia 4090 0,42 sec.

https://files.catbox.moe/f4wb53.mp4

CPU.vs.GPU.test_2.mp4

[metamountain]

Not quite a benchmark, you Gorilla.

[metamountain]

RTX 3090
100%|████████████████████████████████████████████████████████████████████████████████████| 20/20 [00:05<00:00, 3.61it/s]
Prompt executed in 6.16 seconds

[BBCreativo]

NVIDIA RTX 5060TI 16GB

[JorgeR81]

Just for laughs ...
GTX 1070
100%|██████████████████████████████████████████████████████████████████████████████████| 20/20 [01:03<00:00, 3.16s/it]
Prompt executed in 72.44 seconds

[BrechtCorbeel]

My 1070 is my favorite, was my longtime friend still in good use, pretty good for what it was, 3060 is 2 faster with only half as many additional cores, but with the 4090 I have now, uncomparable.

[TheMadHatt3r]

RTX 3070 OC Drivers:~528.xx

100%|██████████████████████████████████████████████████████████████████████████████████| 20/20 [00:08<00:00, 2.26it/s]
Requested to load AutoencoderKL
Loading 1 new model
Prompt executed in 10.89 seconds

[BBCreativo]

@BBCreativo impressive speed. Which settings for running comfyui - sage attention, --fast, ...? Thanks in advance

I don´t now, where to change settings for "Sage attention" ?!

Now, the settings are correct, as in the posted Picture above...

[jkrauss82]

Thanks for the clarification. The settings I mentioned are CLI args so you need to install sage attention and then start comfyui with --sage-attention --fast (--fast enables some fp8 optimizations for ada lovelace and blackwell GPUs).

Thinking about it would actually be better if we would know the startup logs per benchmark, too or better the CLI args + the startup log. Even different versions of pytorch can have impact on the performance, especially for newer cards.

[Rogala]

@BBCreativo
Yours is sd_xl_base_1.0.safetensors, while the example refers to v1-5-pruned-emaonly-fp16.safetensors.

I have the same video card, so I was confused.

Total VRAM 16311 MB, total RAM 32599 MB
pytorch version: 2.7.1+cu128
xformers version: 0.0.32+0f0bb9d9.d20250720
Enabled fp16 accumulation.
Set vram state to: NORMAL_VRAM
Device: cuda:0 NVIDIA GeForce RTX 5060 Ti : cudaMallocAsync
Using sage attention
Python version: 3.12.10 (tags/v3.12.10:0cc8128, Apr  8 2025, 12:21:36) [MSC v.1943 64 bit (AMD64)]
ComfyUI version: 0.3.45
ComfyUI frontend version: 1.23.4

Here is my test with sd_xl_base_1.0.safetensors
--fast +xformers --use-sage-attention

Requested to load SDXL
loaded completely 12039.872177886962 4897.0483474731445 True
100%|██████████████████████████| 20/20 [00:05<00:00,  3.56it/s]
Requested to load AutoencoderKL
loaded completely 3603.9759979248047 159.55708122253418 True
Prompt executed in 16.44 seconds
got prompt
100%|██████████████████████████| 20/20 [00:05<00:00,  3.71it/s]
Prompt executed in 6.15 seconds

and this is from v1-5-pruned-emaonly-fp16

--fast +xformers --use-sage-attention

loaded completely 13364.83067779541 1639.406135559082 True
100%|██████████████████████████| 20/20 [00:04<00:00,  4.28it/s]
Requested to load AutoencoderKL
loaded completely 8186.616992950439 159.55708122253418 True
Prompt executed in 7.07 seconds
100%|██████████████████████████| 20/20 [00:04<00:00,  4.40it/s]
Prompt executed in 5.31 seconds

But I compile xformers, Flash Attention, and SageAttention for myself.

[BBCreativo]

@BBCreativo Yours is sd_xl_base_1.0.safetensors, while the example refers to v1-5-pruned-emaonly-fp16.safetensors.

I have the same video card, so I was confused.

Total VRAM 16311 MB, total RAM 32599 MB
pytorch version: 2.7.1+cu128
xformers version: 0.0.32+0f0bb9d9.d20250720
Enabled fp16 accumulation.
Set vram state to: NORMAL_VRAM
Device: cuda:0 NVIDIA GeForce RTX 5060 Ti : cudaMallocAsync
Using sage attention
Python version: 3.12.10 (tags/v3.12.10:0cc8128, Apr  8 2025, 12:21:36) [MSC v.1943 64 bit (AMD64)]
ComfyUI version: 0.3.45
ComfyUI frontend version: 1.23.4

Here is my test with sd_xl_base_1.0.safetensors --fast +xformers --use-sage-attention

Requested to load SDXL
loaded completely 12039.872177886962 4897.0483474731445 True
100%|██████████████████████████| 20/20 [00:05<00:00,  3.56it/s]
Requested to load AutoencoderKL
loaded completely 3603.9759979248047 159.55708122253418 True
Prompt executed in 16.44 seconds
got prompt
100%|██████████████████████████| 20/20 [00:05<00:00,  3.71it/s]
Prompt executed in 6.15 seconds

and this is from v1-5-pruned-emaonly-fp16

--fast +xformers --use-sage-attention

loaded completely 13364.83067779541 1639.406135559082 True
100%|██████████████████████████| 20/20 [00:04<00:00,  4.28it/s]
Requested to load AutoencoderKL
loaded completely 8186.616992950439 159.55708122253418 True
Prompt executed in 7.07 seconds
100%|██████████████████████████| 20/20 [00:04<00:00,  4.40it/s]
Prompt executed in 5.31 seconds

But I compile xformers, Flash Attention, and SageAttention for myself.

I´ve runned it, with the settings from the picture, on top of this thread?

[Rogala]

I thought that the first time you tested with a small v1-5-pruned-emaonly-fp16 and 512x512, which is why you got such high speed. I was wrong, I apologize.

[YulienPohl]

My poor old GPU 😢
GTX 1050 Ti
100%|██████████████████████████████████████████████████████████████████████████████████| 20/20 [02:51<00:00, 8.55s/it]
Prompt executed in 202.83 seconds

[ltdrdata]

RTX 4080
(Ubuntu 22.04, pytorch 2.2.0)

got prompt
100%|██████████| 20/20 [00:05<00:00,  3.53it/s]
Prompt executed in 6.46 seconds

[TheMadHatt3r]

Would love to get some 3060 / 4060 data to better complete the regression... Then I'll post proposed for every 3000/4000 card.
Some older 2000 could be helpful to understand how the smaller VRAM scales as well.

[UnusMundusWave]

RTX 3090 driver:551.61

RTX 3090 driver:531.18

[ecpknymt]

RTX 4060ti 16GB

[TheMadHatt3r]

RTX 4070 12Gb

got prompt
100%|██████████████████████████████████████████████████████████████████████████████████| 20/20 [00:06<00:00, 3.21it/s]
Prompt executed in 7.13 seconds

[askmyteapot]

WSL2 (6.1.x kernel)
Pytorch 2.3.0 from source
Xformers recent build (0.25)

3060Ti

Requested to load SDXL
Loading 1 new model
100%|███████████████████████████████████████████████████████████████████████████████████| 20/20 [00:09<00:00,  2.05it/s]
Requested to load AutoencoderKL
Loading 1 new model
Prompt executed in 13.29 seconds

Tesla P40

got prompt
100%|███████████████████████████████████████████████████████████████████████████████████| 20/20 [00:45<00:00,  2.28s/it]
Requested to load AutoencoderKL
Loading 1 new model
Prompt executed in 49.26 seconds

No xformers

3060ti

Requested to load SDXL
Loading 1 new model
100%|███████████████████████████████████████████████████████████████████████████████████| 20/20 [00:09<00:00,  2.11it/s]
Requested to load AutoencoderKL
Loading 1 new model
Prompt executed in 13.00 seconds

Tesla P40

got prompt
100%|███████████████████████████████████████████████████████████████████████████████████| 20/20 [00:45<00:00,  2.27s/it]
Requested to load AutoencoderKL
Loading 1 new model
Prompt executed in 48.97 seconds

[stepahin]

4090 FE, Ryzen 7900X, Samsung 980 2TB M.2
Win 11, ComfyUI cu121

got prompt
100%|██████████████████████████████████████████████████████████████████████████████████| 20/20 [00:02<00:00, 7.61it/s]
Prompt executed in 3.08 seconds

[kilobait3]

не менее 9,1 должно быть. скиньте дрова моложе 2024г. совсем другая карта будет. 4090 будет как изначально с 94TF на fp16 и 32.

[stepahin]

Моложе 2024г это как? Этот пост итак из апреля 2024. Ну и тут ни у кого не вижу на 4090 чтобы было 9 it/s

[RysDawid]

RTX 3070 Laptop GPU
Asus ZenBook Duo

Loading 1 new model 100%|██████████████████████████████████████████████████████████████████████████████████| 20/20 [00:11<00:00, 1.73it/s] Requested to load AutoencoderKL Loading 1 new model Prompt executed in 15.44 seconds

[CeciliaXCIX]

Hardware:

• AMD Ryzen 5 7600X
• AMD Radeon RX 7900 XTX / XFX MERC 310 AMD Radeon RX 7900 XTX Black
• 2x 16 GB DDR5 5600 MT/s
• 3840x2160 / 240 Hz

Software:

• Win11 24H2
• HIP SDK 6.2.4
• ZLUDA v3.9.2
• ComfyUI ZLUDA

Results seed 1-3:

[FeepingCreature]

Well darn, okay then. Thanks for trying!

[FeepingCreature]

@Andyholm #2970 (comment) guy over here says 5it/s on wsl2 before compile. Might be worth trying?

[CeciliaXCIX]

Hey thanks for the update. It works and I got around 4.2it/s with flash attention and no compile, but everything else felt so sluggish so I'm just going to stay with zluda for now until I get another drive I can use for linux

[Hakim3i]

Hey thanks for the update. It works and I got around 4.2it/s with flash attention and no compile, but everything else felt so sluggish so I'm just going to stay with zluda for now until I get another drive I can use for linux

If it is under wsl increase your windows swap file it will help am using 48GB of swap.

[CeciliaXCIX]

Hey thanks for the update. It works and I got around 4.2it/s with flash attention and no compile, but everything else felt so sluggish so I'm just going to stay with zluda for now until I get another drive I can use for linux

If it is under wsl increase your windows swap file it will help am using 48GB of swap.

Already at 48GB automatically :)

[PluckTheDragon]

My new PC has Windows 11, a Nvidia GeForce RTX 3060 Ti with a AMD Ryzen 5 3600 6-Core Processor.

Is it okay or meh to generates AI arts with a big picutre size?

[Reznar]

R7 5700X3D
64 RAM
RX 7600XT driver 25.3.1
Windows 11 LTSC
ComfyUI version: 0.3.29 Zluda

got prompt
100%|██████████████████████████████████████████████████████████████████████████████████| 20/20 [00:17<00:00, 1.13it/s]
Prompt executed in 20.52 seconds

[l33tx0]

3080TI

[Crystools INFO] Crystools version: 1.22.1
[Crystools INFO] CPU: AMD Ryzen 9 5900X 12-Core Processor - Arch: AMD64 - OS: Windows 11
[Crystools INFO] Pynvml (Nvidia) initialized.
[Crystools INFO] GPU/s:
[Crystools INFO] 0) NVIDIA GeForce RTX 3080 Ti
[Crystools INFO] NVIDIA Driver: 576.02
Total VRAM 12288 MB
pytorch version: 2.6.0+cu126
Set vram state to: NORMAL_VRAM
Device: cuda:0 NVIDIA GeForce RTX 3080 Ti : native

1024x1024 SDXL BASE

100%|███████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 20/20 [00:04<00:00, 4.01it/s]
Prompt executed in 5.60 second

[PatsySE]

Intel B580 Steel Legend OC 12 GB

Total VRAM 11874 MB, total RAM 32679 MB
pytorch version: 2.5.1+cxx11.abi
Set vram state to: HIGH_VRAM
Device: xpu
Using pytorch attention
Python version: 3.10.6 | packaged by conda-forge | (main, Oct 24 2022, 16:02:16) [MSC v.1916 64 bit (AMD64)]
ComfyUI version: 0.3.33
ComfyUI frontend version: 1.18.9
[Crystools INFO] Crystools version: 1.22.1
[Crystools INFO] CPU: AMD Ryzen 5 5600X 6-Core Processor - OS: Windows 11

When you use Pytorch(as well as I could manage to get it working at all) you have to live with some warnings now and then

Standard operation SDXL BASE 1.0

2025-05-12 17:01:33,326 - _logger.py - IPEX - INFO - Currently split master weight for xpu only support sgd
2025-05-12 17:01:33,347 - _logger.py - IPEX - INFO - Conv BatchNorm folding failed during the optimize process.
2025-05-12 17:01:33,366 - _logger.py - IPEX - INFO - Linear BatchNorm folding failed during the optimize process.
2025-05-12 17:01:33,368 - _logger.py - IPEX - WARNING - [NotSupported]failed to apply concat_linear on unet, please report bugs
100%|██████████████████████████████████████████████████████████████████████████████████| 20/20 [00:05<00:00, 3.96it/s]

Slightly overclocked GPU
100%|██████████████████████████████████████████████████████████████████████████████████| 20/20 [00:04<00:00, 4.08it/s]
2025-05-12 17:03:40,597 - _logger.py - IPEX - INFO - Currently split master weight for xpu only support sgd
2025-05-12 17:03:40,608 - _logger.py - IPEX - INFO - Conv BatchNorm folding failed during the optimize process.
2025-05-12 17:03:40,618 - _logger.py - IPEX - INFO - Linear BatchNorm folding failed during the optimize process.
2025-05-12 17:03:40,618 - _logger.py - IPEX - WARNING - [NotSupported]failed to apply concat_linear on unet, please report bugs

For 12gb card that didnt cost me NVIDIA money I am reasonable happy.

[PatsySE]

Thanks for your reply and details! I use python 3.12 and intel extension 2.7.10+xpu before. I might reset the environment and test it later on.
I had so much problem with 3.12 and 2.7.10 just to make it work so I try to stick as long as it works.
I tried out to eliminate my errors as well now this morning:

(ailibuv) C:\Users\mrced\ComfyUI>python main.py ^
More? --disable-xformers ^
More? --disable-ipex-optimize ^
More? --fp16-unet ^
More? --bf16-vae ^
More? --bf16-text-enc
Total VRAM 11874 MB, total RAM 32679 MB
pytorch version: 2.5.1+cxx11.abi
Set vram state to: NORMAL_VRAM
Device: xpu
Using pytorch attention
Python version: 3.10.6 | packaged by conda-forge | (main, Oct 24 2022, 16:02:16) [MSC v.1916 64 bit (AMD64)]
ComfyUI version: 0.3.43
ComfyUI frontend version: 1.23.4
VAE load device: xpu:0, offload device: cpu, dtype: torch.bfloat16
Requested to load SDXLClipModel
loaded completely 9.5367431640625e+25 1560.802734375 True
CLIP/text encoder model load device: xpu:0, offload device: cpu, current: xpu:0, dtype: torch.bfloat16
Requested to load SDXL
loaded completely 8886.17548828125 4897.0483474731445 True

[PatsySE]

Above was with OC on the card, without it the decoding takes double the time. iterations is almost the same but look at this image:

[chundot]

I have dual intel gpu(ACER A770 16GB, B580 TRI 12GB) installed, so I use ONEAPI_DEVICE_SELECTOR to change device.
My motherboard is X670E PG Lightning (support PCIe 5.0 x16 + PCIe 4.0 x4).
running with: python main.py --disable-ipex-optimize --bf16-unet.

---

pytorch version: 2.7.0+xpu
Python version: 3.12.11 | packaged by Anaconda, Inc. | (main, Jun 5 2025, 12:58:53) [MSC v.1929 64 bit (AMD64)]
ComfyUI version: 0.3.43
ComfyUI frontend version: 1.23.4

---

9700X+B580 TRI(PCIe 4.0 x4)
Using pytorch attention in VAE
VAE load device: xpu:0, offload device: cpu, dtype: torch.bfloat16
CLIP/text encoder model load device: xpu:0, offload device: cpu, current: cpu, dtype: torch.float16
Requested to load SDXLClipModel
loaded completely 10525.03828125 1560.802734375 True
Requested to load SDXL
loaded completely 8951.23408203125 4897.0483474731445 True
100%|█████████████| 20/20 [00:09<00:00, 2.13it/s]
Requested to load AutoencoderKL
loaded completely 2161.5671875000003 159.55708122253418 True
Prompt executed in 17.73 seconds
100%|█████████████| 20/20 [00:07<00:00, 2.51it/s]
Prompt executed in 8.73 seconds
100%|█████████████| 20/20 [00:07<00:00, 2.54it/s]
Prompt executed in 8.69 seconds

---

9700X+ACER A770(PCIe 4.0 x16)
Using pytorch attention in VAE
VAE load device: xpu:0, offload device: cpu, dtype: torch.bfloat16
CLIP/text encoder model load device: xpu:0, offload device: cpu, current: cpu, dtype: torch.float16
Requested to load SDXLClipModel
loaded completely 14511.67890625 1560.802734375 True
Requested to load SDXL
loaded completely 12937.87470703125 4897.0483474731445 True
100%|█████████████| 20/20 [00:14<00:00, 1.36it/s]
Requested to load AutoencoderKL
loaded completely 4434.55859375 159.55708122253418 True
Prompt executed in 23.44 seconds
100%|█████████████| 20/20 [00:13<00:00, 1.53it/s]
Prompt executed in 14.28 seconds
100%|█████████████| 20/20 [00:13<00:00, 1.52it/s]
Prompt executed in 14.31 seconds

---

Don't know why, but my A770 is just not even close to B580.
And what's more, A770 consume more power (~165W) than B580 (~100W).
Waiting for B580 24GB or B770 to replace my A770.

[PatsySE]

i installed a conda env with latest pytorch. went from 4.18 it s to 3.60 it s using python main.py --disable-xformers --disable-ipex-optimize --bf16-unet --bf16-vae --bf16-text-enc start args

[simonlui]

After the changes I made to implement --async-offload and turning on non-blocking memory transfers for the XPU backend, keeping in mind I run Fedora Linux with a Ryzen 9 5950x and an Acer Bifrost A770 16GB with a desktop environment, I get the following with async offload added with --bf16-unet --bf16-vae --bf16-text-enc --async-offload --gpu-only --mmap-torch-files --reserve-vram 1 with Pytorch Nightly 2.9.0.dev20250721+xpu

Using pytorch attention in VAE
VAE load device: xpu:0, offload device: xpu:0, dtype: torch.bfloat16
Requested to load SDXLClipModel
loaded completely 9.5367431640625e+25 1560.802734375 True
CLIP/text encoder model load device: xpu:0, offload device: xpu:0, current: xpu:0, dtype: torch.bfloat16
loaded diffusion model directly to GPU
Requested to load SDXL
loaded completely 9.5367431640625e+25 4897.0483474731445 True
100%|█████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 20/20 [00:16<00:00,  1.20it/s]
Requested to load AutoencoderKL
loaded completely 9.5367431640625e+25 159.55708122253418 True
Prompt executed in 23.54 seconds
got prompt
100%|█████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 20/20 [00:13<00:00,  1.48it/s]
Prompt executed in 14.39 seconds

Before, the fastest speed I saw using the right environments and driver versions and etc. was around 1.31it/s so this is a good ~12-13% improvement. This is only marginally slower than the results above for the A770 which I think is running headless without the desktop environment overhead. Would like to see if anyone wants to test with the changes now in.

[Rogala]

I updated all elements, settings and drivers, it became faster

Checkpoint files will always be loaded safely.
Total VRAM 16311 MB, total RAM 32599 MB
pytorch version: 2.7.0+cu128
Set vram state to: NORMAL_VRAM
Device: cuda:0 NVIDIA GeForce RTX 5060 Ti : cudaMallocAsync
Using sage attention
Python version: 3.12.9 (tags/v3.12.9:fdb8142, Feb  4 2025, 15:27:58) [MSC v.1942 64 bit (AMD64)]
ComfyUI version: 0.3.34
ComfyUI frontend version: 1.19.9
[Crystools INFO] Crystools version: 1.22.1
[Crystools INFO] CPU: 12th Gen Intel(R) Core(TM) i3-12100F - Arch: AMD64 - OS: Windows 11
[Crystools INFO] Pynvml (Nvidia) initialized.
[Crystools INFO] GPU/s:
[Crystools INFO] 0) NVIDIA GeForce RTX 5060 Ti
[Crystools INFO] NVIDIA Driver: 576.40
[ComfyUI-Easy-Use] server: v1.2.9 Loaded
ComfyUI-GGUF: Partial torch compile only, consider updating pytorch
### Loading: ComfyUI-Manager (V3.31.13)
[ComfyUI-Manager] network_mode: public
### ComfyUI Version: v0.3.34-6-g2156ce94 | Released on '2025-05-12'

python.exe -s ComfyUI\main.py --windows-standalone-build --use-sage-attention
loaded completely 12035.872093963622 4897.0483474731445 True
100%|██████████████████████████████████████████████████████████████████████████████████████| 20/20 [00:06<00:00,  2.86it/s]
Requested to load AutoencoderKL
loaded completely 3601.975914001465 159.55708122253418 True
Prompt executed in 12.14 seconds
got prompt
100%|██████████████████████████████████████████████████████████████████████████████████████| 20/20 [00:07<00:00,  2.81it/s]
Prompt executed in 7.82 seconds

python.exe -s ComfyUI\main.py --windows-standalone-build --fast fp16_accumulation --use-sage-attention
loaded completely 12039.872177886962 4897.0483474731445 True
100%|██████████████████████████████████████████████████████████████████████████████████████| 20/20 [00:05<00:00,  3.67it/s]
Requested to load AutoencoderKL
loaded completely 3603.975914001465 159.55708122253418 True
Prompt executed in 10.99 seconds
got prompt
100%|██████████████████████████████████████████████████████████████████████████████████████| 20/20 [00:05<00:00,  3.73it/s]
Prompt executed in 6.02 seconds

[Clusters]

Can this be that the B580 is faster than the RTX 5060 TI 16GB?

[Rogala]

Nvidia is better than all other cards, they simply work with all AI applications, maybe in some cases B580 is better, but I doubt it's worth it

[Alekkk1111]

Let me know, a month has passed. Maybe new drivers have been released and the speed has changed?

[CeciliaXCIX]

Hardware:

• AMD Ryzen 7 9800x3D
• RTX 5090 - ASUS GeForce RTX 5090 ROG Astral LC OC
  
• 2x32 GB DDR5 6000 MT/s
• 3840x2160 / 240 Hz

Software:

• OS: Win11 24H2
• Python version: 3.12.0
• Pytorch version: 2.8.0.dev20250510+cu128
• NVIDIA Driver: 576.40
• Using sage attention
• NORMAL_VRAM

Seed 0 & 1

got prompt
100%|██████████████████████████████████████████████████████████████████████████████████| 20/20 [00:02<00:00,  7.73it/s]
Prompt executed in 8.93 seconds
got prompt
100%|██████████████████████████████████████████████████████████████████████████████████| 20/20 [00:02<00:00,  9.59it/s]
Prompt executed in 2.44 seconds

Average speed of 16 random seeds: 10.50 it/s | 2.26 seconds

100%|██████████████████████████████████████████████████████████████████████████████████| 20/20 [00:02<00:00,  9.46it/s]
Prompt executed in 2.47 seconds
100%|██████████████████████████████████████████████████████████████████████████████████| 20/20 [00:02<00:00,  9.54it/s]
Prompt executed in 2.47 seconds
100%|██████████████████████████████████████████████████████████████████████████████████| 20/20 [00:01<00:00, 10.01it/s]
Prompt executed in 2.37 seconds
100%|██████████████████████████████████████████████████████████████████████████████████| 20/20 [00:02<00:00,  9.52it/s]
Prompt executed in 2.48 seconds
100%|██████████████████████████████████████████████████████████████████████████████████| 20/20 [00:02<00:00,  9.70it/s]
Prompt executed in 2.43 seconds
100%|██████████████████████████████████████████████████████████████████████████████████| 20/20 [00:01<00:00, 11.22it/s]
Prompt executed in 2.11 seconds
100%|██████████████████████████████████████████████████████████████████████████████████| 20/20 [00:01<00:00, 11.11it/s]
Prompt executed in 2.15 seconds
100%|██████████████████████████████████████████████████████████████████████████████████| 20/20 [00:02<00:00,  9.95it/s]
Prompt executed in 2.39 seconds
100%|██████████████████████████████████████████████████████████████████████████████████| 20/20 [00:01<00:00, 10.86it/s]
Prompt executed in 2.19 seconds
100%|██████████████████████████████████████████████████████████████████████████████████| 20/20 [00:01<00:00, 10.97it/s]
Prompt executed in 2.15 seconds
100%|██████████████████████████████████████████████████████████████████████████████████| 20/20 [00:01<00:00, 11.21it/s]
Prompt executed in 2.12 seconds
100%|██████████████████████████████████████████████████████████████████████████████████| 20/20 [00:01<00:00, 11.20it/s]
Prompt executed in 2.12 seconds
100%|██████████████████████████████████████████████████████████████████████████████████| 20/20 [00:01<00:00, 11.31it/s]
Prompt executed in 2.11 seconds
100%|██████████████████████████████████████████████████████████████████████████████████| 20/20 [00:01<00:00, 11.17it/s]
Prompt executed in 2.15 seconds
100%|██████████████████████████████████████████████████████████████████████████████████| 20/20 [00:01<00:00, 10.10it/s]
Prompt executed in 2.31 seconds
100%|██████████████████████████████████████████████████████████████████████████████████| 20/20 [00:01<00:00, 10.68it/s]
Prompt executed in 2.21 seconds

[Alex3c]

GTX 1080 + Intel 4790K
100%|██████████████████████████████████████████████████████████████████████████████████| 20/20 [00:49<00:00, 2.50s/it]
Prompt executed in 56.21 seconds

[FraudR]

GTX 970 + i7-8700k
100%|██████████| 20/20 [02:53<00:00, 8.67s/it]
Requested to load AutoencoderKL
0 models unloaded.
loaded partially 128.0 127.99989318847656 0
Prompt executed in 179.18 seconds

[FraudR]

RTX 4060 + i7-8700k
100%|██████████| 20/20 [00:11<00:00, 1.72it/s]
Prompt executed in 15.92 seconds

[gonztok]

6700 XT user here, just wanted to share my numbers.

  ::  ------------------------ ZLUDA -----------------------  ::
  ::  Triton core imported successfully
  ::  Running Triton kernel test...
  ::  Triton kernel test passed successfully
  ::  Triton initialized successfully
  ::  Patching ONNX Runtime for ZLUDA — disabling CUDA EP.
  ::  Using ZLUDA with device: AMD Radeon RX 6700 XT [ZLUDA]
  ::  Applying core ZLUDA patches...
  ::  Initializing Triton optimizations
  ::  Configuring Triton device properties...
  ::  Triton device properties configured
  ::  Flash attention components found
  ::  AMD flash attention enabled successfully
  ::  Configuring PyTorch backends...
  ::  Disabled CUDA flash attention
  ::  Enabled math attention fallback
  ::  ZLUDA initialization complete
  ::  ------------------------ ZLUDA -----------------------  ::

Total VRAM 12272 MB, total RAM 65191 MB
pytorch version: 2.7.0+cu118
Set vram state to: NORMAL_VRAM
Device: cuda:0 AMD Radeon RX 6700 XT [ZLUDA] : native
Using sub quadratic optimization for attention, if you have memory or speed issues try using: --use-split-cross-attention
Python version: 3.11.9 (tags/v3.11.9:de54cf5, Apr  2 2024, 10:12:12) [MSC v.1938 64 bit (AMD64)]
ComfyUI version: 0.3.44
ComfyUI frontend version: 1.23.4

SDXL Gen:

100%|██████████████████████████████████████████████████████████████████████████████████| 20/20 [00:18<00:00,  1.10it/s]
Prompt executed in 23.51 seconds

[BTekV4]

6800XT, 5800X3D, Bazzite Linux, kernel Linux 6.13.7-107, ROCm

[START] Security scan
[DONE] Security scan

Prestartup times for custom nodes:
0.0 seconds: /var/home/maurizio/AppImages/Data/Packages/ComfyUI/custom_nodes/rgthree-comfy
1.3 seconds: /var/home/maurizio/AppImages/Data/Packages/ComfyUI/custom_nodes/ComfyUI-Manager

Checkpoint files will always be loaded safely.
Total VRAM 16368 MB, total RAM 64185 MB
pytorch version: 2.7.1+rocm6.3
AMD arch: gfx1030
ROCm version: (6, 3)
Set vram state to: NORMAL_VRAM
Device: cuda:0 AMD Radeon Graphics : native
Using sub quadratic optimization for attention, if you have memory or speed issues try using: --use-split-cross-attention
Python version: 3.10.18 (main, Jun 3 2025, 18:23:41) [GCC 11.4.0]
ComfyUI version: 0.3.44
ComfyUI frontend version: 1.23.4
[Prompt Server] web root: /var/home/maurizio/AppImages/Data/Packages/ComfyUI/venv/lib/python3.10/site-packages/comfyui_frontend_package/static

Loading: ComfyUI-Manager (V3.33.8)

[ComfyUI-Manager] network_mode: public

ComfyUI Revision: 3645 [c5de495] *DETACHED | Released on '2025-07-08'

Starting server

To see the GUI go to: http://0.0.0.0:8188
To see the GUI go to: http://[::]:8188

got prompt
got prompt
model weight dtype torch.float16, manual cast: None
model_type EPS
Using split attention in VAE
Using split attention in VAE
VAE load device: cuda:0, offload device: cpu, dtype: torch.float32
CLIP/text encoder model load device: cuda:0, offload device: cpu, current: cpu, dtype: torch.float16
Requested to load SDXLClipModel
loaded completely 15102.8 1560.802734375 True
Requested to load SDXL
loaded completely 12132.11580078125 4897.0483474731445 True
100%|███████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 20/20 [00:12<00:00, 1.64it/s]
Requested to load AutoencoderKL
loaded completely 3791.5863281250004 319.11416244506836 True
Prompt executed in 16.48 seconds
100%|███████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 20/20 [00:12<00:00, 1.67it/s]
Prompt executed in 13.22 seconds
got prompt
0%| | 0/20 [00:00<?, ?it/s]got prompt
100%|███████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 20/20 [00:11<00:00, 1.75it/s]
Prompt executed in 12.62 seconds
100%|███████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 20/20 [00:11<00:00, 1.74it/s]
Prompt executed in 12.66 seconds
got prompt
0%| | 0/20 [00:00<?, ?it/s]got prompt
100%|███████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 20/20 [00:11<00:00, 1.75it/s]
Prompt executed in 12.59 seconds
100%|███████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 20/20 [00:11<00:00, 1.74it/s]
Prompt executed in 12.65 seconds

[BTekV4]

1080 TI, 3700X, TrueNAS Scale, kernel 6.12.15

[START] Security scan
[DONE] Security scan

ComfyUI-Manager: installing dependencies done.

** ComfyUI startup time: 2025-07-10 16:11:44.737
** Platform: Linux
** Python version: 3.10.12 (main, Feb 4 2025, 14:57:36) [GCC 11.4.0]
** Python executable: /comfy/mnt/venv/bin/python3
** ComfyUI Path: /comfy/mnt/ComfyUI
** ComfyUI Base Folder Path: /comfy/mnt/ComfyUI
** User directory: /basedir/user
** ComfyUI-Manager config path: /basedir/user/default/ComfyUI-Manager/config.ini
** Log path: /basedir/user/comfyui.log

Checkpoint files will always be loaded safely.
Total VRAM 11163 MB, total RAM 31994 MB
pytorch version: 2.7.1+cu126
Set vram state to: NORMAL_VRAM
Device: cuda:0 NVIDIA GeForce GTX 1080 Ti : cudaMallocAsync
Using pytorch attention
Python version: 3.10.12 (main, Feb 4 2025, 14:57:36) [GCC 11.4.0]
ComfyUI version: 0.3.43
ComfyUI frontend version: 1.23.4
[Prompt Server] web root: /comfy/mnt/venv/lib/python3.10/site-packages/comfyui_frontend_package/static

Loading: ComfyUI-Manager (V3.33.3)

[ComfyUI-Manager] network_mode: public

ComfyUI Revision: 3614 [e18f53c] *DETACHED | Released on '2025-06-27'

Starting server

To see the GUI go to: http://0.0.0.0:8188

got prompt
model weight dtype torch.float16, manual cast: torch.float32
model_type EPS
Using pytorch attention in VAE
Using pytorch attention in VAE
VAE load device: cuda:0, offload device: cpu, dtype: torch.float32
Requested to load SDXLClipModel
loaded completely 9.5367431640625e+25 1560.802734375 True
CLIP/text encoder model load device: cpu, offload device: cpu, current: cpu, dtype: torch.float16
loaded diffusion model directly to GPU
Requested to load SDXL
loaded completely 9.5367431640625e+25 4897.0483474731445 True
100%|██████████| 20/20 [00:16<00:00, 1.21it/s]
Requested to load AutoencoderKL
loaded completely 3110.153739929199 319.11416244506836 True
Prompt executed in 33.08 seconds
got prompt
100%|██████████| 20/20 [00:16<00:00, 1.23it/s]
Prompt executed in 16.69 seconds
got prompt
100%|██████████| 20/20 [00:16<00:00, 1.23it/s]
Prompt executed in 16.69 seconds
got prompt
100%|██████████| 20/20 [00:16<00:00, 1.23it/s]
Prompt executed in 16.70 seconds
got prompt
100%|██████████| 20/20 [00:16<00:00, 1.23it/s]
Prompt executed in 16.70 seconds

[chundot]

start with: python main.py --disable-ipex-optimize
X670E PG Lightning + 9700X + 96GB DDR5 5200MHz

---

B580 TRI(PCIe 4.0 x4)
Using pytorch attention in VAE
VAE load device: xpu:0, offload device: cpu, dtype: torch.bfloat16
CLIP/text encoder model load device: xpu:0, offload device: cpu, current: cpu, dtype: torch.float16
Requested to load SDXLClipModel
loaded completely 10525.03828125 1560.802734375 True
Requested to load SDXL
loaded completely 8951.23408203125 4897.0483474731445 True
100%|████| 20/20 [00:08<00:00, 2.27it/s]
Requested to load AutoencoderKL
loaded completely 2161.5671875000003 159.55708122253418 True
Prompt executed in 15.83 seconds
100%|████| 20/20 [00:06<00:00, 3.05it/s]
Prompt executed in 7.05 seconds
100%|████| 20/20 [00:06<00:00, 3.05it/s]
Prompt executed in 7.00 seconds

---

ACER A770(PCIe 4.0 x16)
Normally, 1.09s/it, Prompt executed in 21.76 seconds.
With --bf16-unet, 1.58it/s, Prompt executed in 13.15 seconds

---

Today I also tried comfyui_openvino. All the offload is on cpu.
python main.py --cpu --use-pytorch-cross-attention
It's slow, maybe it's because my cpu is 9700x.
GPU.0(B580)
100%|████| 20/20 [00:54<00:00, 2.75s/it]
Prompt executed in 69.87 seconds.
100%|████| 20/20 [00:06<00:00, 3.25it/s]
Prompt executed in 17.46 seconds
100%|████| 20/20 [00:06<00:00, 3.17it/s]
Prompt executed in 17.53 seconds
GPU.1(A770)
100%|████| 20/20 [00:56<00:00, 2.81s/it]
Prompt executed in 68.12 seconds
100%|████| 20/20 [00:07<00:00, 2.72it/s]
Prompt executed in 18.99 seconds
100%|████| 20/20 [00:07<00:00, 2.73it/s]
Prompt executed in 19.12 seconds

It takes forever to compile, and if you change model or selected device without enable cache, you have to compile it again.

[kilobait3]

очень хороший результат!

[Not-Daijoubu]

Recently built my first PC, major upgrade from an old Dell G7 laptop that was constantly thermal throttling.

Intel 245k
PNY RTX 5060 Ti 16GB OC
PCIe 4.0x8 due to riser cable

Total VRAM 16311 MB, total RAM 32209 MB
pytorch version: 2.7.1+cu128
Set vram state to: NORMAL_VRAM
Device: cuda:0 NVIDIA GeForce RTX 5060 Ti : cudaMallocAsync
Using pytorch attention
Python version: 3.12.10 (tags/v3.12.10:0cc8128, Apr  8 2025, 12:21:36) [MSC v.1943 64 bit (AMD64)]
ComfyUI version: 0.3.44
ComfyUI frontend version: 1.23.4

NON-OC

100%|██████████████████████████████████████████████████████████████████████████████████| 20/20 [00:07<00:00,  2.60it/s]
Prompt executed in 8.41 seconds

OC using MSI Afterburner, +330core/+2600mem (real core speed is ~3150-3200MHz)

100%|██████████████████████████████████████████████████████████████████████████████████| 20/20 [00:06<00:00,  2.86it/s]
Prompt executed in 7.70 seconds

SageAttention 2++, NON-OC

100%|██████████████████████████████████████████████████████████████████████████████████| 20/20 [00:06<00:00,  2.89it/s]
Prompt executed in 7.58 seconds

SageAttention 2++, OC

100%|██████████████████████████████████████████████████████████████████████████████████| 20/20 [00:06<00:00,  3.19it/s]
Prompt executed in 6.90 seconds

SageAttention 2++, Fast FP16 Accumulation, Non-OC

100%|██████████████████████████████████████████████████████████████████████████████████| 20/20 [00:05<00:00,  3.67it/s]
Prompt executed in 6.12 seconds

SageAttention 2++, Fast FP16 Accumulation, OC

100%|██████████████████████████████████████████████████████████████████████████████████| 20/20 [00:05<00:00,  3.96it/s]
Prompt executed in 5.65 seconds

Nice little card. The image difference using both Sage and Fast FP16 is unnoticeable with this SDXL Benchmark, but with my standard NoobAI + Euler A workflow, the difference can be pretty significant though without noticeable quality loss.

Seems Intel may have the better AI card in terms of raw speed (per older posts). Can't wait to see what the B60 and Celestial can do.