Is VRAM more important than performance on GPU? GTX4060TI 16GB vs GTX4070 12GB

[gamelife322]

I'm considering upgrading the computer graphics card, which one is recommended, GTX4060TI 16GB or GTX4070 12GB?

I don't mind the price difference, although the 4070 has higher performance, but the 4060TI has 16GB. Is the extra 4GB advantage critical for StableDiffusion?

My current motherboard interface is PCIe 3.0. Although it can also support PCIe 4.0 GPU, I don’t know if it will greatly reduce the AI ​​computing power. Do I need to replace the motherboard too?

Can you guys recommend me? Thanks

[suonnon]

If you have 3060ti 8GB like me, you would pick 4060ti 16GB immediately lol

[gamelife322]

Would you consider RTX4070? According to the GPU performance evaluation report, the difference in performance between the 3060ti and 4060ti is not that great, however the increase from 8GB to 16GB is very attractive.

Due to the performance of my graphic card GTX1650 Super, I can't even let others know that I am playing AI art XD

This is the highest limit of my current graphic card, and it is already on the verge of crashing.

[rungvang]

Its running on my 1060 3gb

[gamelife322]

Wow! Really? Why is my GTX1650 Super 4GB AI art images so different from yours? Every time I run stable diffusion, only about 2GB of VRAM can be used, and the other half is used by the system.

[rungvang]

Yes, i'm serious. 1060 3GB using --medram --xformers. The above image was using chilloutmix model. The rendering time was 2-3min/image. And for upscaling I have to use external program since 3gb VRAM is nearly impossible to upscale using SD.

[pw405]

Usually I would say the PCI Gen 3 or Gen 4 doesn't matter much in this case - but I would definitely run in PCI Gen 4 mode. The 4060Ti runs on a cut down PCI Gen 4 X 8 link (instead of X16).

If your motherboard was set to PCI Gen 3, you'd only be running on PCI Gen 3 X 8 which very likely may affect performance. This is pretty common for these ultra small graphics chips. AMD did the same thing with the RX 6500XT last generation.

[gamelife322]

Asus ROG Strix B360-H Gaming, I bought it in 2020. I don’t know if I need to change it. It feels like a big project, like reassembling the entire computer.

[pw405]

Oh yes, Intel was a bit late to the PCI Gen 4 party. That board only supports PCI Gen 3 edit - running a GPU that only supports X8 lanes on a PCI Gen 3 board would very likely cause performance issues because you'll be in PCI Gen 3 X 8. A GPU that can run X 16 PCI lanes would do much better. Not entirely sure how much it would hurt Stable Diffusion performance, but you can find additional info and testing about this specific problem here: https://www.youtube.com/watch?v=uU5jYCgnT7s

But, ultimately you are right, a motherboard swap is a lot of work and you'd need to trash your CPU and upgrade that as well. Any motherboard compatible with PCI Gen 4 will not work with your current CPU.

Granted, the only other way to get 16GB or higher of VRAM and X16 PCI interface is to buy a 4080 or 4090 which is likely more expensive than upgrading CPU & motherboard.

[pw405]

(to clarify: a card that has X16 lanes: 4070, 4070Ti, 4080, or 4090 would not suffer as much performance degradation as the 4060Ti/4060Ti 16GB running on X8 PCI lanes. Pretty sure Nvidia gimps the high VRAM/affordable cards on purpose to force you to consider a higher model, which they starve of VRAM until you get to $1,000+ range).

[gamelife322]

I hope everything could be plug and play. I am most worried about formatting the computer, and I am afraid that replacing the motherboard will lead to such consequences. I'd rather spend more money to build a brand new computer than format my current PC.
It feels like my budget keeps going up XD

[yamfun]

Here is a benchmark from a Jp guy.
https://chimolog.co/bto-gpu-stable-diffusion-specs/#16002151024SDXL_10

Scroll down a bit for a benchmark graph with the text SDXL. It shows that the 4060 ti 16gb will be faster than a 4070 ti when you gen a very big image.

Consider that there will be future version after SDXL, which probably need even more vram, it seems wise to get a card with more vram.

[gamelife322]

I might consider getting a 4060ti 16GB, since handling large images is my top priority, 512 pixels is too small by today's standards.

The last time I built my PC was three years ago, the performance is still good can meet the needs of daily use. I'm thinking whether to replace the motherboard in order to fully support the new graphic card performance. The LGA socket of the motherboard is bound to the processor's generation. If buy a new one (LGA1700 socket motherboard), I must replace the Intel processor too. The budget is almost equivalent to buying another computer.

[minienglish1]

TL, DR: Buy the card with the most ram. PCI-e gen 3 8x is 100% fine.

I have many gpus and tested them with stable diffusion, both in webui and training: gt 1010, tesla p40 (basically a 24gb 1080), 2060 12gb, 3060 12gb, 2 * 3090, & a 4090.

Once you get to the 20XX gen (because 10XX doesn't support fp16) and up, gpu vram beats everything else. You always need more vram, you will never have enough vram. VRAM will limit what you can and cannot do, gpu speed doesn't matter if you don't have enough vram.
The difference between what my 12gb & my 24gb cards can do is night and day, regardless of speed.

If you can get a used 3090 at a similar price, it's a better deal. Replace the thermal paste on the gpu, themal pads on the ram, and add a small heatsink to the backplate over the rear vram, and undervolt/overclock the hell out of it. Buy any decent name brand base model (don't buy the fancy models) and you're golden. Both of my 3090s were bought used (gigabyte - white eagle & gigabyte aorus - The cheaper white eagle out performs the "fancy" aorus after both are undervolted/overclocked). I've run them non-stop for weeks at a time doing fine-tuning, no problems at all. But you 100% must replace the paste, pads, and add the heatsink - 3090s are notorious for high temps due to idiots at the factory.

Don't worry about pci-e gen. I've done fine-tuning on my rtx4090 on pci-e gen 3 8x: absolutely 0% drop in speed vs gen 3 16x. The models may take slightly (as in a couple seconds max) longer to load, but that's it. Once the model is loaded, you're good to go until you change models.

[gamelife322]

Thanks for your reply, you opened my new world. I saw the price of RTX 3090 24GB on a second-hand platform, which price is almost same as brand new RTX4070 12GB. The price is very attractive, but I have no trading experience in this area, and I am researching whether it is safe and reliable.

[minienglish1]

Regarding reliablity and safety, I can only speak from personal experience. 90+% of my computer parts are used. I have many computers between home and my business.
Never once had a core desktop computer part fail (keyboards and mice break all the time, but they're really cheap).

I buy them from a used computer part reseller online (I'm from the USA, but I live in China, so I buy from Taobao. I would assume in the USA/Europe there are reputable used computer parts sellers. Maybe on Amazon?)

The parts I buy are tested as working before being sold, and have a short term exchange/return policy (7-30days) and 3-12months repair policy. I'm assuming the parts are from gaming PCs, old servers, or mining rigs. After I buy them, I test then as working, then take it apart, clean it, and replace themal compound/pads. The 3090s were pretty dirty, but they cleaned up nicely.

Haven't broken one yet during cleaning process. Just find a youtube guide, and take your time. Afterward, check that it works, slowly go through some benchmarks, and if everything is good, start overclock/undervolt. As long as you're patient, diligent, and careful, it's really difficult to permantly damage a GPU.

Wish I could be of more help. I'm sure there are websites/forums discussing where to buy used pc parts from in your country. My only suggestion is to buy from a reputable reseller company, not a random individual. This is because 1) If there is a problem, they have customer service and an inventory of parts to provide exchange/repair 2) They're a business, they have a brand name to protect, and want return customers so they can make a long-term profit.

Good luck. Oh, and make sure your power supply can handle a 3090, they use a good bit of electricity.

[gamelife322]

I'm negotiating with two people, they all offer RTX3090 24GB, different brands Zotac VS Galax. They both seem reliable, have their own online stores, and one of them just opened a retail store half a year ago.

- ZOTAC GAMING GeForce RTX3090 Trinity 24GB - used 2 years, this is office GPU, trade-in from architect firm. cost about USD $645
- GALAX RTX3090 SG (1-Click OC Feature) 24GB - used 2 years, this is used as game unit. cost about USD $602

I'm considering buying a 750W PSU, is that too much? Or 700W is enough. Will Copper Modded help in cooling?
Youtube link: https://www.youtube.com/watch?v=f8f6ZHCPVpw

[minienglish1]

Not familiar with either brand, but you easily able to find reviews/comparsions both cards online. Either should be fine.

I'd say copper shims are overkill, and dangerous. If they slip out of place, can short circuit the card. Might be useful for extreme overclocking, but you're probably more interested just having a reliable card at good price. Just get decent thermal paste and pads, no need anything fancy.

Its the rear memory that gets hot on the card. Just find an old heatsink, ideally one with a fan or add a fan on it. Put a large thin thermal pad on the heat sink, then strap it to back of the card with some zip ties. Done.

Keep it simple. Make sure it fits in your case. And don't waste money stuff you don't need.

Regarding psu, it depends on how good the psu is and what else is in your system. I was using a 650w psu originally, but my 3090 was undervolted and my system was very simple. I always suggest buy the biggest/best psu can. Spend less on the 3090 & mods, and buy a better psu that you can still use later next time you upgrade.

Attached is a pic of the heatsink on the back of one of my 3090s. Mine is fancy, because it only cost 5USD here in China. Any old used flat bottom heatsink will work fine, spend that money on a better psu instead.

[Staszek-org]

TL, DR: Buy the card with the most ram. PCI-e gen 3 8x is 100% fine.

I have many gpus and tested them with stable diffusion, both in webui and training: gt 1010, tesla p40 (basically a 24gb 1080), 2060 12gb, 3060 12gb, 2 * 3090, & a 4090.

Once you get to the 20XX gen (because 10XX doesn't support fp16) and up, gpu vram beats everything else. You always need more vram, you will never have enough vram. VRAM will limit what you can and cannot do, gpu speed doesn't matter if you don't have enough vram.
The difference between what my 12gb & my 24gb cards can do is night and day, regardless of speed.

If you can get a used 3090 at a similar price, it's a better deal. Replace the thermal paste on the gpu, themal pads on the ram, and add a small heatsink to the backplate over the rear vram, and undervolt/overclock the hell out of it. Buy any decent name brand base model (don't buy the fancy models) and you're golden. Both of my 3090s were bought used (gigabyte - white eagle & gigabyte aorus - The cheaper white eagle out performs the "fancy" aorus after both are undervolted/overclocked). I've run them non-stop for weeks at a time doing fine-tuning, no problems at all. But you 100% must replace the paste, pads, and add the heatsink - 3090s are notorious for high temps due to idiots at the factory.

Don't worry about pci-e gen. I've done fine-tuning on my rtx4090 on pci-e gen 3 8x: absolutely 0% drop in speed vs gen 3 16x. The models may take slightly (as in a couple seconds max) longer to load, but that's it. Once the model is loaded, you're good to go until you change models.

What about Tesla P40 comparing to rtx2060, 3060 and 3090?

[minienglish1]

a P40 is basically a rtx1080 witg 24gb ram. Which also means no fp16 speed boost. 1/10 speed of rtx3090 if I remember correctly from my prior testing.

[cisnez]

TL, DR: Buy the card with the most ram. PCI-e gen 3 8x is 100% fine.

I have many gpus and tested them with stable diffusion, both in webui and training: gt 1010, tesla p40 (basically a 24gb 1080), 2060 12gb, 3060 12gb, 2 * 3090, & a 4090.

Once you get to the 20XX gen (because 10XX doesn't support fp16) and up, gpu vram beats everything else. You always need more vram, you will never have enough vram. VRAM will limit what you can and cannot do, gpu speed doesn't matter if you don't have enough vram. The difference between what my 12gb & my 24gb cards can do is night and day, regardless of speed.

Tesla P40 with 24GB sounds great for upscale.

My RTX 3070 with 8GB is not enough memory. Very fast generating fp16 after 8 seconds and fp32 in about 30 seconds (2048 × 1152 pixels).

Curious how long the Tesla P40 with 24GB could do a 4x upscale.

[minienglish1]

It's not. The P40 (and all RTX10xx models) don't have fp16 fully integrated into the chipset. So it only runs at fp32. That's effectively 1/2 memory and 1/2 speed versus a chip with fp16 support. Real world performance is about /10 speed of a rtx3090 based on the last benchmark I did months ago.

[lingxuan630]

If you don't mind the price, RTX 4090 is the best choice from my experience.

[rebixpl]

what about 3090 24 gb second hand? i have currently 2060 6gb and it can't run sdxl because of vram

[minienglish1]

used 3090 24GB is fine, used 2060/3060 12GB are also ok. Best value on the market is a modded 2080TI 22GB. Modded 12GB/16GB cards are also an option.