3090 PCI-e gen3 vs. gen4 performance?

[space-nuko]

Hi so I have a 3090 but with xformers and default settings am only getting ~2it/s depending on the sampler. According to other threads this seems low so I checked and realized others are getting things like 16it/s and I wondered if it has to do with my specs or smth

The one thing that stands out is my motherboard only supports PCI-e gen3, I've read online that for games it doesn't actually affect performance much

Big question for me is does it affect SD performance that much

I'd like to know before I commit to spending a few hundred dollars and ripping out my motherboard

Or at least some data on other peoples specs and their its/s so I can compare baselines

[kidol]

Not sure, but check whether your current motherboard has slots with different lanes (1x, 2x, 4x, 8x, 16x). Your card should be put in a 16x slot for full bandwidth.

[space-nuko]

Just checked and my card is indeed in a 16x slot

[cmp-nct]

It also needs to be routed on the mainboard.
Though the performance you describe would be more like a 1x connection. Hard to get it that badly set up.

[nProtect]

I'm using ASUS PRIME-X470 Pro (PCI-E 3.0) with Nvidia 3090
Image generation take about 10-12it/s (7-8it/s without xformers), may slower if apply too many promps or hypernetwork.

[B34STW4RS]

you should be getting around 17.5 -18.5, if your negative prompt is long, pad the positive prompt to match to regain full speed or use ddim while you build the prompt, as ddim seems to not suffer from this limitation.

[space-nuko]

I tried using this prompt purely for testing

a pair of dice, high quality, photorealistic, sharp focus, depth of field
Steps: 20, Sampler: Euler a, CFG scale: 7, Seed: 809010432, Size: 512x512, Model hash: 925997e9, Model: animefull-final-pruned, Batch size: 4, Batch pos: 0

With xformers enabled batch size of 4 gives ~5it/s, batch size of 1 gives ~11it/s

But most of the time I use way more complex prompts with hypernetworks and large negative prompts

I guess I would like to know how complex the prompts you're using that result in 18.5, if they're the final result you're aiming for or not

[B34STW4RS]

Here you go, this one was giving me a 17.05-17.67 it/s range.

masterpiece, realistic, (((perfect))), photorealistic, ultrarealism, (official art), 1girl, bobcut, blue eyes, brunette, serious look on face, (sleeveless red dress), (white thighhighs), (hightops), hands on handlebars, ((riding on motorcyle)), speeding, looking at viewer, highway, scenery, dusk, raining, puddles, bokeh, intricate, highly detailed, motion lines
Negative prompt: nsfw, lowres, ((mangled)), inaccurate, mutated, mutant, tumor, growth, cancer, guro, injury, violence, (deformed), (bad anatomy), (disembodied limbs), bad hands, text, error, missing fingers, extra digit, fewer digits, cropped, worst quality, low quality, normal quality, jpeg artifacts, signature, watermark, username, blurry
Steps: 28, Sampler: Euler, CFG scale: 11, Seed: 1747816934, Size: 512x512, Model hash: 925997e9, Clip skip: 2, ENSD: 31337

Time taken: 31.00s

Torch active/reserved: 3692/4816 MiB, Sys VRAM: 6066/24576 MiB (24.68%)

Also no, I learned of this balancing trick later on, and had since abandoned getting this image perfect and animated it as it was which ended up less than ideal. After rebuilding the prompt for this post, it came much closer to what I was looking for in the first place.

[space-nuko]

Most I got was ~14it/s with batch size 1 and no hypernetworks

With hypernetworks I get ~12it/s and with batch size 4 it's ~4it/s

I guess this is normal? Thanks for sharing

[tmode-1960]

Hello ... without xformers and all settings on default I get ~2.5it/s on a Titan X Pascal, a really old card. Your problem seems not to be PCIe-3

[eroc123]

meanwhile me with stable diffusion running on a core i3 4th gen without gpu at 140s/it :D

[cmp-nct]

CPU makes no sense, you get a free online GPU from Google Colabs after all which should get you about 4it/sec for free.
Also any GPU for 100 bucks will do an acceptable job ;)

[cooperdk]

CPU makes no sense, you get a free online GPU from Google Colabs after all which should get you about 4it/sec for free.
Also any GPU for 100 bucks will do an acceptable job ;)

But you need to use colab.
Not everyone wants Google to know what they're doing.

[cmp-nct]

Hi so I have a 3090 but with xformers and default settings am only getting ~2it/s depending on the sampler. According to other threads this seems low so I checked and realized others are getting things like 16it/s and I wondered if it has to do with my specs or smth

The one thing that stands out is my motherboard only supports PCI-e gen3, I've read online that for games it doesn't actually affect performance much

Big question for me is does it affect SD performance that much

I'd like to know before I commit to spending a few hundred dollars and ripping out my motherboard

Or at least some data on other peoples specs and their its/s so I can compare baselines

Based on what you wrote it's more guessing than knowing.

1. A 3090 that delivers only 2it/sec is possibly a fake, I've seen 3090ies that were powered by cheap internals. Get GPU-Z and look at the details, you'll maybe directly spot something.

2. I'd go back to roots, use the vanilla repo, fresh installation, no xformers. See what happens.

3. Run GPU-Z with all sensor values exposed, look at power draw, frequencies of GPU and memory, TDP and temperatures.
   You will see something is OFF and with a bit dedication you'll see the issue.

[Krader13]

I have a really old rig and just upgraded my old GTX 1060 with a RTX 3090 Ti. Here are my system specs with performance measures of stable diffusion:

PCIe 2.0
Old Dual Xeon X5470 3.33Ghz CPU motherboard
32GB DDR2 Ram
3 15K RPM SAS Drives in RAID
RTX 3090 Ti

With xformers enabled, I'm getting ~6 it/s on default settings. Better than my gtx 1060 which was getting 1.23 it/s. I ran the example above and got 5.97 it/s.

I'm on the lookout to see if I can get this to perform better, but I'm happy with it not taking as long as my old gpu.

[USBhost]

And I thought my system was old... Dual e5-2690v2 128gb ddr3.

[nProtect]

Sorry for digging,

I have discovery something, Disable Hardware-accelerated GPU scheduling will improve both generating / training performace on RTX 3090.

Image generating from 11it/s > 16-17 it/s when disabled.

[USBhost]

this really helped on my VM. went from around 5it/s to 7-8it/s. 3080 TI here

[amirhjalali]

Just wanted to say this boosted my 3090 from 13-14 to 16-17.

Thanks!

[Krader13]

I changed this and my 3090 Ti went from ~6it/s to 9.36it/s! Nice buddy! My system specs are a few comments up, so this is on a really old machine.