Support NeoGAF

[YeulEmeralda]

I'm wondering if a 5070/80 will be enough to last the PS6 generation. Planning on building a new PC next year.

 

[Danknugz]

You guys definitely need a 5090 to play all those indie games lmao. Wake me up when we get actual demanding visual beasts on PC. Until then, 7080 is a more realistic wait for buy.

4090 is shit tier when it comes to ai training though.

 

[Danknugz]

I've been one of the premium retards who bought a 4090 for the price Nivida were asking for.

What can I say? I am a massive idiot.

i got mine at launch (generic gigabyte) for 1600 or so and never regretted it

also sold my 3090 for like 1100 or something (got 9 something after ebay took its cut)

 

[Kataploom]

I'm more concerned about the power port they're using these days, literally the only thing keeping me from thinking on getting an Nvidia card until it's solved, but I'm whiling to give it a go if reports are good.

I want to confirm if the "fire" reaction from Celcius had some double sense lmao

 

[Hudo]

4090 is shit tier when it comes to ai training though.

Eh. It's actually one of the reasons why I bought one. It's not that bad. You can do surprisingly a lot with 24 gigs of RAM.

But for anything related to LLMs (even with LoRA etc.) you probably need something with 64 gigs or more. Amazon AWS, Google Colab, etc. Or you use a smaller LLM, like Mistral 7B or so. And you can often get by with just Retrieval Augmented Generation instead of fine-tuning it for real.

 

[sendit]

5080 should once again be the sweet spot in terms of performance and power consumption. In for one when it shows up.

Hopefully this means they’ve got 5xxx series laptop GPUs for next year’s gaming laptop refreshes too

Lol. Okay.

 

[Ozriel]

Lol. Okay.

 

[Danknugz]

Eh. It's actually one of the reasons why I bought one. It's not that bad. You can do surprisingly a lot with 24 gigs of RAM.

But for anything related to LLMs (even with LoRA etc.) you probably need something with 64 gigs or more. Amazon AWS, Google Colab, etc. Or you use a smaller LLM, like Mistral 7B or so. And you can often get by with just Retrieval Augmented Generation instead of fine-tuning it for real.

not sure what kind of training you're doing but i've trained my own lora's all the training i've done has taken 4+ hours at least, if not 8. of course, it could be the way i set it up but based on the it/sec it only makes sense the 5090 would handled these operations significantly faster. shit tier for me is anything that makes you wait for hours.

i've also trained my own checkpoints which took even longer

 

[Kiya's Guard]

They probably need to improve the cable

 

[Hudo]

not sure what kind of training you're doing but i've trained my own lora's all the training i've done has taken 4+ hours at least, if not 8. of course, it could be the way i set it up but based on the it/sec it only makes sense the 5090 would handled these operations significantly faster. shit tier for me is anything that makes you wait for hours.

i've also trained my own checkpoints which took even longer

Training time is not really my concern, rather if I can fit my model on the card during training (I am already using stuff like gradient accumulation etc.). So I guess we have just different aspects that are important to us.

 

[Danknugz]

Training time is not really my concern, rather if I can fit my model on the card during training (I am already using stuff like gradient accumulation etc.). So I guess we have just different aspects that are important to us.

ah, thanks, i had to look up gradient accumulations. i'm using kohya_ss and wonder if that's an option there, and if it would speed up my training given the right parameters.

 

[Hudo]

ah, thanks, i had to look up gradient accumulations. i'm using kohya_ss and wonder if that's an option there, and if it would speed up my training given the right parameters.

Ah, I am not familiar with Kohya_ss. At first glance, it seems like a training library/framework for diffusion models? Interesting. I am using vanilla diffusion models as pre-trainer for my encoder-decoder networks. But I had so many specific things that I opted to write my own stuff (using Pytorch, obviously).

...which was a fucking bitch to debug.

 

[elbourreau]

OK, let's just hope the 5080 is actually affordable and a worthy upgrade from the 3080.

 

[Danknugz]

Ah, I am not familiar with Kohya_ss. At first glance, it seems like a training library/framework for diffusion models? Interesting. I am using vanilla diffusion models as pre-trainer for my encoder-decoder networks. But I had so many specific things that I opted to write my own stuff (using Pytorch, obviously).

...which was a fucking bitch to debug.

yeah i don't have the time to get that low level i am basically just kiddie scripting my way through it with a very rough understanding.

 

[GustavoLT]

I'm wondering if a 5070/80 will be enough to last the PS6 generation. Planning on building a new PC next year.

4080/4090 are already PS6 proof

 

[StereoVsn]

yeah i don't have the time to get that low level i am basically just kiddie scripting my way through it with a very rough understanding.

Training taking hours doesn’t make a card “shit tier”. You would have to pay significantly more to get much better results. A 48GB version of basically 4090 tuned for AI will cost at least 3x and it’s only going up from there.

For the price a 4090 isn’t bad. Also as long as the model does into memory you can have parallel cards to decrease the time required, again depends on what you are doing of course.

 

[StereoVsn]

4080/4090 are already PS6 proof

Let’s not get crazy. PS6 will launch in 2028 and the gen will go for 7-8 years if it’s business as usual.

4090 would be a 14 year old card by end of that gen. Even a 5090 would be a 12 year old card by then. I don’t believe that will possibly be enough. Either way you would need to upgrade somewhere in the PS6 generation, sooner with a 4090.

 

[OverHeat]

4080/4090 are already PS6 proof

Thats crazy talk…

 

[FalconPunch]

4080/4090 are already PS6 proof

Nah. If the ps5 pro is supposedly 4070 level, the 4090 is not safe at all. It should either get matched or beaten by the ps6.

 

[Gaiff]

Let’s not get crazy. PS6 will launch in 2028 and the gen will go for 7-8 years if it’s business as usual.

4090 would be a 14 year old card by end of that gen. Even a 5090 would be a 12 year old card by then. I don’t believe that will possibly be enough. Either way you would need to upgrade somewhere in the PS6 generation, sooner with a 4090.

Your math is a bit off.

PS6 is likely to launch in late 2027/early 2028. The 4090 would be just a bit over 5 years old by then. Add another 7 years and it'd be just a bit over 12 years old.

The 5090 is presumably launching in late 2024/early 2025. It would be just about 3 years old by the time the PS6 launches and about 10 years old at the end.

We're talking 10 and 12 years old, not 12 and 14. Not a big difference but still. The GTX 1080 is turning 8 years old this month. I believe it can last the rest of gen as long as ray tracing remains optional. A 5090 could feasibly get someone through the PS6 generation and will very likely be close to a PS6 in raw performance but with an older feature set.

 

[StereoVsn]

Your math is a bit off.

PS6 is likely to launch in late 2027/early 2028. The 4090 would be just a bit over 5 years old by then. Add another 7 years and it'd be just a bit over 12 years old.

The 5090 is presumably launching in late 2024/early 2025. It would be just about 3 years old by the time the PS6 launches and about 10 years old at the end.

We're talking 10 and 12 years old, not 12 and 14. Not a big difference but still. The GTX 1080 is turning 8 years old this month. I believe it can last the rest of gen as long as ray tracing remains optional. A 5090 could feasibly get someone through the PS6 generation.

You are missing a year here and there. Let’s say 5090 would launch at end 2024. PS5 launches toward end of 2028 (probably Nov as usual). That’s a 4 year card.

PS5 launched in 2020. PS6 would launch in 2028. That’s an 8 year interval. So again, a 12 year card by the time PS7 launches (who knows what would happen by then honestly). 4090 will be a 2 year old card so at the end of gen it will be 14.

Sony might cut next gen to 7 years or may launch in early 2028, who knows, but if things go as with past gen’s, the above is likely outcome. IMO a better bet is to get a 5080, save up some money and get a 6080 halfway through. That is if a 90 series card midway replacement is out of budget.

In any case, a 1080 wouldn’t last the rest of this gen well in either case. Some games are already coming out with RT being used as the main option.

 

[Gaiff]

You are missing a year here and there. Let’s say 5090 would launch at end 2024. PS5 launches toward end of 2028 (probably Nov as usual). That’s a 4 year card.

PS6 is unlikely to launch at the end of 2028. End of 2027 is more plausible.

PS3 - Nov 2006
PS4 - Nov 2013
PS5 - Nov 2020

Pretty much 7 years every time. PS6, if the trend continues, will launch at the end of 2027, not 2028.

PS5 launched in 2020. PS6 would launch in 2028. That’s an 8 year interval. So again, a 12 year card. 4090 will be a 2 year old card so at the end of gen it will be 14.

Nope, it would be 7 years. PS5 launched in Nov 2020, less than 2 months before 2021. The 4090 launched in Oct 2022. If the PS6 launches in Nov 2027, the 4090 will be 5 years and 1 month old. Basically 5 years.

Sony might cut next gen to 7 years or may launch in early 2028, who knows, but if things go as with past gen’s, the above is likely outcome. IMO a better bet is to get a 5080, save up some money and get a 6080 halfway through. That is if a 90 series card midway replacement is out of budget.

What do you mean, "cut to 7 years"? It's been 7 years since the PS3 days.

In any case, a 1080 wouldn’t last the rest of this gen well in either case. Some games are already coming out with RT being used as the main option.

And they still work without ray tracing. RT will invariably be optional in console games because they don't have the horsepower to run it properly most of the time. There's also the fact that the Series S needs to be supported until the end of the generation and a 1080 will destroy that console. A 1080 will last to the end of the generation. It won't run game great or anything, but if you're willing to lower some settings and forego ray tracing, it'll last you until 2028.

I have little doubt that a 5090 will be able to carry one through the entire PS6 generation. It just won't be a great experience by the end but it'll probably be able to play every game.

 

[StereoVsn]

I guess we will see what happens with 5000 series cards, but IMO, a much better experience would be to get say a 5080 and halfway through the gen replace it.

 

[Gaiff]

I guess we will see what happens with 5000 series cards, but IMO, a much better experience would be to get say a 5080 and halfway through the gen replace it.

Obviously. I wouldn't recommend anyone riding any GPU for a decade but hey, if you wanna do it, you can. Just don't expect great results.

 

[//DEVIL//]

I’ll pass on the 5080 as I have a 4090.

I’ll wait for the 5090 and see the performance difference.

Even if there is a big gap between the 4090 and the 5080, the worm in my ass will want to have the 5090.

With that being said. If there is no exclusive features like AI enhancements in games ( something like the DLSS toggle in games ) and just pure extra resta performance, there is a slight chance I might skip the 5000 as there are no games that require the performance of 4090 at 4K 144 ( the 4080 is more than enough to be honest ).

So I am at the just wait and see stage at the moment

 

[RJMacready73]

£2000 to go from 4k/120fps to 4k/240fps...

 

[00_Zer0]

Keep in mind that Kopite7kimi might be seeing a cut down die and thinking that it’s the 5080… while Nvidia will release that die and named it 5090 at $2000

Radeon 8000 series better be a miracle worker with the ray tracing or just gonna wait for a 9000 series or buy a 4080 Super used at the middle or end of the 5000 series lifecycle. Screw paying $1000 or more for a GPU. I would only consider it if I was building a brand new rig.

 

[Danknugz]

Training taking hours doesn’t make a card “shit tier”. You would have to pay significantly more to get much better results. A 48GB version of basically 4090 tuned for AI will cost at least 3x and it’s only going up from there.

For the price a 4090 isn’t bad. Also as long as the model does into memory you can have parallel cards to decrease the time required, again depends on what you are doing of course.

i was exaggerating but let's be honest who wants to wait around for hours unless they have a dedicated machine or are renting a cloud GPU. it's true that vram is really more important for reducing time although as far as I understand the tensor cores are involved in operations getting the data back and forth from the vram, and generally newer generations have more performant tensor cores, i'm pretty sure the 5090 will be faster than a 4090 even if it is still only 24GB.

 

[StereoVsn]

i was exaggerating but let's be honest who wants to wait around for hours unless they have a dedicated machine or are renting a cloud GPU. it's true that vram is really more important for reducing time although as far as I understand the tensor cores are involved in operations getting the data back and forth from the vram, and generally newer generations have more performant tensor cores, i'm pretty sure the 5090 will be faster than a 4090 even if it is still only 24GB.

It will be faster, sure, but nobody will know till specs are released.

And really, inference is more of an issue time wise as you need a big enough farm to answer requests in seconds. Well, for actual service providers vs individual uses of course.

Training taking hours is ok in that scenario. But yeah, if this is your main PC, that could be an issue.

 

[Danknugz]

It will be faster, sure, but nobody will know till specs are released.

And really, inference is more of an issue time wise as you need a big enough farm to answer requests in seconds. Well, for actual service providers vs individual uses of course.

Training taking hours is ok in that scenario. But yeah, if this is your main PC, that could be an issue.

even with inference if i want to create a simple animatediff at 30fps that's a lot of frames to generate and queuing that up could potentially be just as long as training a model, correct me if i'm wrong but this is more reliant on tensor cores because you aren't using as much vram for each frame generation, and in this case a later generation tensor core such as those presumably that will be found in the 5090 will improve these kinds of processes as well. i wonder, would having twice as much vram even matter for that case?

 

[StereoVsn]

even with inference if i want to create a simple animatediff at 30fps that's a lot of frames to generate and queuing that up could potentially be just as long as training a model, correct me if i'm wrong but this is more reliant on tensor cores because you aren't using as much vram for each frame generation, and in this case a later generation tensor core such as those presumably that will be found in the 5090 will improve these kinds of processes as well. i wonder, would having twice as much vram even matter for that case?

This is all true, but what I am saying is that for the last 1.5 years or so, 4090 was a very good card for AI compared to its price point. You would need to spend like $5-6K on something remotely better. That’s not shit tier. Training just takes time.

But yeah, a 5090 should give a nice uptick (still hoping for more VRAM as well). Best thing to do is run a separate training box so you don’t have your main PC occupied all the time.

It is price though .

 

[MrRenegade]

The sooner it comes, the better. It will drive down the price of previous generation of cards. You people go and sell your 4090 ASAP.

 

[Come On Tars]

1080 ti still mostly good for my needs. I keep saying aite next gen ill upgrade. The 5 series might be it.

 

[digital_ghost]

Anyone in need of a kidney? Will trade for 5090.

 

[winjer]

NVIDIA Testing GeForce RTX 50 Series "Blackwell" GPU Designs Ranging from 250 W to 600 W

According to Benchlife.info insiders, NVIDIA is supposedly in the phase of testing designs with various Total Graphics Power (TGP), running from 250 Watts to 600 Watts, for its upcoming GeForce RTX 50 series Blackwell graphics cards. The company is testing designs ranging from 250 W aimed at...

www.techpowerup.com

According to Benchlife.info insiders, NVIDIA is supposedly in the phase of testing designs with various Total Graphics Power (TGP), running from 250 Watts to 600 Watts, for its upcoming GeForce RTX 50 series Blackwell graphics cards. The company is testing designs ranging from 250 W aimed at mainstream users and a more powerful 600 W configuration tailored for enthusiast-level performance. The 250 W cooling system is expected to prioritize compactness and power efficiency, making it an appealing choice for gamers seeking a balance between capability and energy conservation. This design could prove particularly attractive for those building small form-factor rigs or AIBs looking to offer smaller cooler sizes. On the other end of the spectrum, the 600 W cooling solution is the highest TGP of the stack, which is possibly only made for testing purposes. Other SKUs with different power configurations come in between.

We witnessed NVIDIA testing a 900-watt version of the Ada Lovelace AD102 GPU SKU, which never saw the light of day, so we should take this testing phase with a grain of salt. Often, the engineering silicon is the first batch made for the enablement of software and firmware, while the final silicon is much more efficient and more optimized to use less power and align with regular TGP structures. The current highest-end SKU, the GeForce RTX 4090, uses 450-watt TGP. So, take this phase with some reservations as we wait for more information to come out.