Nvidia RTX 50xx

[Aslanna]

RTX 5070 - $549
RTX 5070 Ti - $749
RTX 5080 - $1000
RTX 5090 - $2000

According to Nvidia the 5070 gives "4090 l(levels of) performance" but supposedly people are saying that figure comes from comparing DLSS 3 numbers (2x) with DLSS 4 (4x). So in actuality it's marketing smoke and mirrors as far as raw performance goes when comparing the two.

I am also also not sure why Nvidia is so stingy with memory when these cards are priced as high as they are. The 5090 is fine at 32 GB but the 5070 should have 16 GB instead of 12, the 5070 Ti is fine at 16 GB, and the 5080 should have been 24 GB instead of 16.

It will be interesting to see what sort of supply these cards will have. And if the 'Trump tariffs' will impact them. If interested I'd suggest trying to get one as soon as they are available assuming they're not only available from scalpers. I'd say though if you have a 40 series and are happy with it there's no reason to upgrade. Unless you're rich.

Now as for older generations I'm not sure how the comparisons work out. However I think I'll just stick with my 3080 Ti for now. It does what I need it to do. Plus I do still feel a bit burned, even if it was a self-burn, by buying it just before the shortages ended and the price on it dropped $500 or something dumb. The 5080 does have 4 GB more RAM than I have now and only 10 watts more power usage so perhaps it may be a decent upgrade. Although with EVGA no longer making graphics cards I'd have to decide on a new brand to go with.

[Winnow]

I'm sure the 5070 will be a nice card but it won't be a 4090 simply because of the VRAM. It may be a great gamer card though.

when DDR7 comes out you'll have 3GB on a chip instead of 2GB so the next "Supers" will probably be 5090 48GB and 5080 24GB in a year or whenever.

Nice to see the 5090 is a 2-slot card. they at least were able to do that much.

I'll probably buy a 5090 at some point because I can't get enough AI. I've spent hours every day screwing around with it since march of 2023. Images, Videos, and LLMs. The 4090 is great and I don't feel pressured to get the 5090 with it but I want the best pro/consumer card I can get. I have a high end 750W PSU but don't think I'd risk it with the 1000W recommended PSU for the 575W from the 5090. I'll see how much willpower I have and if I can hold out for a 48GB VRAM card.

The Nvidia 128GB of Unified ram $3,000 mini DIGITS "super computer" is interesting. If you were only into LLMs that would be great. This is at least a tiny bit of good news as LLM enthusiasts might go for it instead of 5090s...although for images and videos it will still be the 5090 since the DIGITS unified ram would be 5-10X slower.

I spent hours yesterday and today trying to get a freaking linux image to run in Windows WSL.WSL2/Docker etc. It's the only way to train Hunyuan video LORAs right now. I failed miserably. Got close a few times but I'm just not good enough with linux to troubleshoot stuff. I'll post on another thread regarding that not to hijack this one.

[Funkmasterr]

Yeah I'm definitely fine with my 4090, likely for another 2-3 years. At this point, I've still been able to play any game I've tried at max settings at 90+ FPS, but often 120 (I cap at this due to my TV). I also do a little video editing for Drone/Mountain bike videos on my YouTube channel and using Resolve, I think the longest it's taken me to render a video is still under a minute (granted I haven't made any videos longer than maybe 5 minutes yet). For all those reasons it would be silly to spend the money, even if I did have it.

[Aslanna]

The 4090 was $1600 and now the 5090 is $2000. Scared to see how much the 6090s will be.

I have a 800W power supply so I'm not even sure I could run a 5080 since I think the recommendation for that is 850W. It would probably be fine but other than FOMO there's no need for me to upgrade. I don't even play games on my PC so I think my money would be better spent on upgrading my 6 year old PC.

when DDR7 comes out you'll have 3GB on a chip instead of 2GB so the next "Supers" will probably be 5090 48GB and 5080 24GB in a year or whenever.

The 50's are using GDDR7. The 4090 didn't get a Super or a TI so I doubt the 5090 will. The 4060 had a Ti, the 4070 had a Super, Ti, and Ti Super and the 4080 got a Super. When the 4080 got a Super the memory amount stayed the same so I think the same with happen with any 5080 revision.

I spent hours yesterday and today trying to get a freaking linux image to run in Windows WSL.WSL2/Docker etc. It's the only way to train Hunyuan video LORAs right now. I failed miserably. Got close a few times but I'm just not good enough with linux to troubleshoot stuff

I set that up last year but then never ended up using it. At least as far as getting Ubuntu installed. I didn't make it to messing around with Docker. I doubt I even know the root password any longer. I probably should have written that down.

[Winnow]

I set that up last year but then never ended up using it. At least as far as getting Ubuntu installed. I didn't make it to messing around with Docker. I doubt I even know the root password any longer. I probably should have written that down.

Setting up WSl/WSL2 is super easy as is installing linux images. probably only takes a few seconds to delete the default image and create a new one if forgot password.

The problem for me is the hunyuan training stuff, diffusion pipeline etc is a bitch to setup. didn't even work with a premade linux image.

Hopefully there will be easier windows training for hunyuan videos. It's becoming popular and the image to video will be out soon which will really make it take off. I don't mind figuring out the python. I know the basics of linux but really don't want to work in that OS.

[Winnow]

This video is hilarious covering all the third party BS marketing to try and get you to buy their brand. I have an MSI so might stick with that brand if/when I get one but not sure. MSI is including circles with their 80XX cards!

[Aslanna]

With the cooling improvements to the FE cards I think I'd probably go with one of those. The third-party boards usually come with a bit of a premium on the cost. And I don't care about all the RGB and shit most of them have.

[Winnow]

With the cooling improvements to the FE cards I think I'd probably go with one of those. The third-party boards usually come with a bit of a premium on the cost. And I don't care about all the RGB and shit most of them have.

The reference card would be fine. It's more a matter of finding them in stock. I under volt anyway and never see inside my case unless I'm working on it.

Just a random note. AI advancements are still coming at an incredibly fast pace. Video/voice/LLMs. I'm having a blast playing around with it even with the frustrations of Python. Against my will I've been forced to become decent at workflows etc in ComfyUI. I'd hate to have to teach that in a class though.

These 2 reddits are pretty good for checking in on progress etc.

https://www.reddit.com/r/StableDiffusion/

https://www.reddit.com/r/LocalLLaMA/

[Winnow]

Kinda lengthy but it certainly looks like the FE is designed well.

[Aslanna]

According to a posted spreadsheet my local Micro Center had 4 5090s and 80 5080s. Sucks for the people who lined up waiting if that's an accurate count. I thought the whole reason NVidia stopped making 4090s a few months ago was to build up stock of the new cards. I guess not! I suppose that was just to get old stock out of the retail channel prior to the release of new models.

Also, based on reviews, the 5080 is a bad deal so if interested in that maybe hold out and hope they release a 5080 Super or 5080Ti. Maybe by that time you'll be able to actually buy one.

[Winnow]

According to a posted spreadsheet my local Micro Center had 4 5090s and 80 5080s. Sucks for the people who lined up waiting if that's an accurate count. I thought the whole reason NVidia stopped making 4090s a few months ago was to build up stock of the new cards. I guess not! I suppose that was just to get old stock out of the retail channel prior to the release of new models.

Also, based on reviews, the 5080 is a bad deal so if interested in that maybe hold out and hope they release a 5080 Super or 5080Ti. Maybe by that time you'll be able to actually buy one.

Yeah the Gamers Nexus (same people from last video posted) said something like 233 5090s were available on launch day in the US. Paper launch.

What a mess for anyone interested in local AI. If you don't already have a 4090, those will be expensive and with Chinese New Year, not many 5090s will be made over the next month and they have to be shipped from China after that.

Only good news (for me) is that a lot can be done on 24GB VRAM since that's all consumers have. Things are tailored for that amount.

At least for gamers you can get the 5080 easier but sounds like reviewers aren't happy with it. I'd wait for the mid series release with 24GB most likely or for 5090's to actually be available if ever.

[Winnow]

Another thing that is putting even more pressure on the 5090 is that the latest LLMs like Deepseek R1 and the new 24B Mistral Small 3 are amazingly good (close in quality to the top line OpenAI.Claude/etc) and quantized versions of those fit on 24GB so more small business and consumers can run those locally as long as you have at least 24GB VRAM.

It's crazy how much demand and use cases there are for VRAM and it's not going away. Nvidia first had the crypto miners and now AI is ridiculously good and getting better not monthly or weekly, but daily there is some awesome new advancement in AI LLM, Images, video, audio etc.

Nvidia will sellout of as many 5090's as they can make and they don't even need to make them since they are making way more money off thier pro GPU sales that they also can't make enough of.

[Aslanna]

From all accounts it sounds like the 50 series "launch" has been a shit show.

I remember a time when, if you wanted a graphics card, you could go out and buy one. It would be the specs listed and it wouldn't potentially burn your house down due to a poorly designed power connector.

Then along came crypto mining and fucked everything up. Things haven't been the same since.

[Aslanna]

The 4090 didn't get a Super or a TI so I doubt the 5090 will. The 4060 had a Ti, the 4070 had a Super, Ti, and Ti Super and the 4080 got a Super. When the 4080 got a Super the memory amount stayed the same so I think the same with happen with any 5080 revision.

Well I am right so far about no 5090 Super but I was (potentially) wrong in one regard:

GeForce RTX 5080 SUPER rumored to feature 24GB memory, RTX 5070 SUPER with 18GB config

Not that it matters since whatever MSRP ends up being you won't be able to find them for that. Especially in the US with the current tariff situation. But if I had to guess I'd say 24GB 5080 Super for $1250 and a 18GB 5070 Super at $850.

[Winnow]

Not that it matters since whatever MSRP ends up being you won't be able to find them for that. Especially in the US with the current tariff situation. But if I had to guess I'd say 24GB 5080 Super for $1250 and a 18GB 5070 Super at $850.

$1250 for 24GB VRAM sounds too good. I would immediately be thinking if I could buy 2 of them for 48GB VRAM. $2500 would be an awesome price for 48 GB...even marked up another $500 per card to $3500 people would be throwing punches to get their hands on them.

VRAM is more important than the rest of the card. That's why the 3090 has hung in there so long.

I will say Nvidia made the 4090 well. I've tortured mine almost non stop since Jan 2024 and it's held up. Training loras at max MHz each for 3-4 hours straight. 100's of hours just training LORAs. then every image, video, LLM, audio Ai inference taps out at max MHz, Videos taking 2.5-10 mins each at max Mhz, 1000+ of them and 100K-200K images generated. It's a good product. Overpriced due to demand but good.

[Spang]

Not that it matters since whatever MSRP ends up being you won't be able to find them for that.

Maybe the children will have two graphics cards instead of 30.

[Aslanna]

I received an email from Micro Center today. They have the Asus 5080 in stock for $1800 which is $800 more than MSRP.

They also have the Gigabyte 5090 for $2920. $920 more than MSRP.

The GPU situation these days is truly dire. Especially with companies still releasing 8GB cards.

Overpriced due to demand but good.

It's overpriced due to demand due to NVidia not making as many consumer cards since they are using all their resources towards AI bullshit for corporations for higher profit. Don't let NVidia off the hook for abandoning (or fucking quite harshly) all the people who bought their shit previously and made them the company they are today.

I wish there was a better alternative to NVidia in the higher-end GPU market because right now there really isn't.

[Winnow]

I received an email from Micro Center today. They have the Asus 5080 in stock for $1800 which is $800 more than MSRP.

They also have the Gigabyte 5090 for $2920. $920 more than MSRP.

The GPU situation these days is truly dire. Especially with companies still releasing 8GB cards.

Overpriced due to demand but good.

It's overpriced due to demand due to NVidia not making as many consumer cards since they are using all their resources towards AI bullshit for corporations for higher profit. Don't let NVidia off the hook for abandoning (or fucking quite harshly) all the people who bought their shit previously and made them the company they are today.

I wish there was a better alternative to NVidia in the higher-end GPU market because right now there really isn't.

Don't get the 5080 no matter what. It's 100% about VRAM. Anything less than 24 GB VRAM is a waste if you have even a hint of wanting to do any AI stuff, LLMs, Image, Video, etc.


5090 eventually is worth it at some price point. The drivers are starting to be worked out. They should be ok in a few months. I paid $2100 for my 4090. ($1,400 retail)



Another VEO 3 collection of videos all themed about AI saying they aren't prompts. It really is amazing that you can creat these voice/sound sync'd video with just a few lines of a text prompt. Spang probably still doesn't realize he's a prompt!

[Aslanna]

I have no desire to purchase either one at those prices. There's no way in hell I'd spend $3k on a graphics card.

And since I'm also never paying more than $2k either (other than maybe a 5090 at MSRP) I guess I'm stuck with my 3080 Ti forever. And that's only 12GB so the 16GB on the 5080 would actually be an upgrade.

From what I can tell even the 5070 Ti outperforms the 3080 Ti so maybe come the 6xxx generation I'll go lower than a 6080. I hope they stop being stingy with the VRAM on those but that probably won't be the case. Assuming they even bother to still be in the consumer market at that time.

[Winnow]

And since I'm also never paying more than $2k either (other than maybe a 5090 at MSRP) I guess I'm stuck with my 3080 Ti forever. And that's only 12GB so the 16GB on the 5080 would actually be an upgrade.

It would probably be best to stick with your 3080 than buy anything less than 24GB. 16 over 12 doesn't really do a whole lot for AI.

At 24GB I still have to watch my VRAM constantly.

It comes down to "fun factor" . You can do stuff with less VRAM (12GB will work to try things out with gguf/quant models) but if it takes a long time to generate a nice image and like 20+ mins for a video, it's not something you can play around with on a whim.

For most of the things I do, my VRAM runs between 18-22GB. 23+ gets a little iffy since I'm running other things , web browsers etc along with it and fluctuations with less than 1 GB ram can crash the system.

Image generation becomes a little less fun/experimental if it takes more than 60 seconds to generate. With the latest speed boosts, i've gotten AI videos down to 1.5 mins to 3 mins for 8 second videos that are at a resolution good enough but not high res.

If you bought a 5080 I'd say trade it immediately to Funk for his 4090 that he probably doesn't take advantage of.

The thing with the latest AI stuff WAN2.1, HiDream, etc is that it already is too large for 24GB cards (and usually 32GB) and they have to FP8 it to get it to fit (which is fine because no quality loss).

At some point the crazy VEO 3 all-in-one video/audio/lip sync I linked examples of above will be available locally and that's going to push the absolute limits of whatever VRAM you have but at the same time, it's pretty damn amazing.

I've made probably close to 4,000 AI videos now. Images are probably 300,000 but really no idea since hey are all over my drives.

90 seconds to generate a video at around 18-20GB VRAM usage is a "fun factor" level where you can do it and still watch YouTube or a video etc. There's a great new looping workflow where you can loop any video smoothly which to me personally is important due to 8-10 second videos being short so seamless looping is nice. I've also gotten pretty good at extending videos and stitching them together. I've made some 1-2 minute videos.

Also, Image to Video is where it's at. With WAN2.1 (and especially the newer WAN VACE model, you can make awesome videos from any still image. At least until the VEO 3 tech level stuff is locally available.

I'd say get a 4090 but those are $2,700 lowest with most over $3K (makes the $2,100 I paid for mine seem pretty good ~16 months ago) They are almost the same price as the 5090's which are around $3,300 on Amazon.

Anyway. My advice: Consider the "fun factor". 12-16GB you can still play around with some things with the best being out of reach or crazy long generation time...so you may as well stick with 12GB VRAM. Also remember for LLMs, you need to fit the entire model into VRAM. Anything in normal RAM is going to make the fun factor drop off a cliff as you watch text scroll across your screen at 1982 300 Baud modem rates.



I close with a short video someone put together already using VEO 3. It may not be "Tom Cruise" level but it's better than Steven Seagal!