Support NeoGAF

[SantaC]

I am mostly gaming on PC these days and been doing so for the last 10-15 years; and steam gaming has never been as good as it is today. However the PC upgrading landscape is just sad. Upgrading to a new graphics card has always excited me, but right now it is fucking dead. It used to be 2 year cycles at worst for new cards, but since Nvidia has zero highend competition from intel and amd they decided to throw in the towel and we are at best getting 3 year cycles now. Usually at computex there are talks about new graphic card generation, but nothing at all this year. It's so boring. I know that a 5090 is still very good, but with zero competition from AMD they can charge whatever for it. Where is RDNA5? Havent heard a lick about it.

I know AI is all that matters these days but i dont understand why they cant co-exist with consumer products.

 

[Deft Beck]

I know AI is all that matters these days but i dont understand why they cant co-exist with consumer products.

It doesn't make nearly as much money to sell to consumers. That's why many component manufacturers shifted gears and stopped selling RAM to consumers.

 

[Mownoc]

Considering the next consoles are using RDNA5 then you'd expect the GPU's to launch at some point in 2027 too. Regardless of if the consoles get delayed or not.

 

[StereoVsn]

Considering the next consoles are using RDNA5 then you'd expect the GPU's to launch at some point in 2027 too. Regardless of if the consoles get delayed or not.

I think Nvidia has been rumored to have delayed 6000 series release into early 2028.

No idea about AMD so maybe late 2027?

 

[winjer]

Because AI is a lot more profitable.
So everything else is an afterthought.

 

[pudel]

I know AI is all that matters these days but i dont understand why they cant co-exist with consumer products.

AI just eating up all the available supply right now. They can co-exist...but supply needs to be there.

 

[Zathalus]

A number of factors:

- Giant leaps in processing power are getting harder.
- Related to the above, process node shrinking is getting harder, vastly more expensive, and with increasing diminishing returns.
- Increasing power draw has been a tried and tested way of boosting performance, but the top end cards are getting a feasible limit here as well.
- VRAM demand has outstripped supply for now.
- AMD only competes in price in the low to mid range. When it comes to feature set and IP they haven't been able to trade blows with Nvidia since early GCN. Maybe RDNA5 will change that, but I've been hearing maybe X will change that for a while now.

 

[Gen.Grievous]

There is a limited supply of materials. They can sell that limited supply to business for 25 times the profit compared to selling it to consumers.

 

[Cumulus]

It used to be 2 year cycles at worst for new cards, but since Nvidia has zero highend competition from intel and amd they decided to throw in the towel and we are at best getting 3 year cycles now.

You realize that Moores Law doesn`t work anymore? Even if NVIDIA wanted to continue the old hardware cycles, the upgrades would be laughable if you stuck to certain die size limits....And even those miniscule upgrades would still cost an arm and a leg.
This is one of those rare occasions where Nvidia isn`t to blame. That´s just the state of the technology.

 

[Hyet]

Money, greed, the dismantling of cheap international trade

 

[HeisenbergFX4]

 

[Liopleurodon]

Good.

It's better to save money, than to waste it on useless crap like high-end GPUs.

 

[Topher]

It's all about the cost of VRAM

 

[Hudo]

The problem actually is that nothing is happening.

 

[Bojji]

It's all about the cost of VRAM

Now it's the vram, earlier if was AI taking away TSMC production lines, earlier than that - mining.

GPU are fucked one way or the other for like 6 years now (at least).

 

[Topher]

Now it's the vram, earlier if was AI taking away TSMC production lines, earlier than that - mining.

GPU are fucked one way or the other for like 6 years now (at least).

and before that it was crypto mining

 

[Smiggs]

Dark days are ahead, I'm afraid. If you didn't buy one last year like everyone was screaming to you about, you're gonna be paying a lot more for your next GPU. Until the madness of these data center build outs either slow down or stop entirely, there will be no respite for any of us.

 

[Dr.Morris79]

Fuck knows. I'm just hoping my 4070 doesn't shit the bed

 

[kruis]

Because AI is a lot more profitable.
So everything else is an afterthought.

It's not just "a lot more profitable", but "vastly more profitable". It's so ridiculously profitable that the people working at Samsung Electronics' memory chip division will get enormous bonuses thanks to Samsung's profit sharing contact agreement. Average bonus size per employee: $400,000. (Samsung Memory Chip division made $260 billion in profits.)

Samsung memory chip staff in line for £310,000 bonuses after AI profit-sharing deal

Agreement averts strike and shows latest impact of AI boom as two more chipmakers join $1tn club

www.theguardian.com

 

[analog_future]

Why sell to consumers for narrow margins when they can sell to corporations for exponentially higher profits?



AI has fucked the average consumer.

 

[Roman Empire]

Next step companies should be prohibited of selling to end user consumers. Let's go back to the stone age and let the graphic cards overlords rule us!

FUCK NVIDIA

 

[Zacfoldor]

So AI is cool and you can run it from your computer if you have a videocard with enough vram.

But what about when everyone in India is using Claude to code at once?

A hyperscaler is a building that contains these chips and power and allows them to do two things. 1. Serve more people, and 2. Charge less per unit of AI(compute) they provide.

If you noticed that some companies recently went live with Claudecode and saw that they used up their yearly budget in a few months. Meaning that AI compute is going to cost these companies more than anticipated. The answer to this is to build more hyperscalers which will reduce the cost of these services and also make them available to everyone. The hyperscalers need videocards and ram to function and they are being built out right now, which is why things are the way they are. New tech like turboquant is coming online that also significantly reduces AI ram usage and will also help in reducing these prices. We need to reach a point where we have less chips than power to power them before we stop really because we need way more hyperscalers than we have now honestly. To give everyone worldwide an AI agent, we need a lot of compute.

 

[FingerBang]

So AI is cool and you can run it from your computer if you have a videocard with enough vram.

But what about when everyone in India is using Claude to code at once?

A hyperscaler is a building that contains these chips and power and allows them to do two things. 1. Serve more people, and 2. Charge less per unit of AI(compute) they provide.

If you noticed that some companies recently went live with Claudecode and saw that they used up their yearly budget in a few months. Meaning that AI compute is going to cost these companies more than anticipated. The answer to this is to build more hyperscalers which will reduce the cost of these services and also make them available to everyone. The hyperscalers need videocards and ram to function and they are being built out right now, which is why things are the way they are. New tech like turboquant is coming online that also significantly reduces AI ram usage and will also help in reducing these prices. We need to reach a point where we have less chips than power to power them before we stop really because we need way more hyperscalers than we have now honestly. To give everyone worldwide an AI agent, we need a lot of compute.

I agree, let's nuke India!

 

[Durin]

There simply isn't the capacity to keep up with the demand around AI, and it's more lucrative to build out the infrastructure to sell AI servers that can be used for consumers + businesses...than just using that memory for consumer-only GPUs, which largely are going to be used just for gaming or content creation by most people.

Eventually a company will fill in the gap producing memory to make money off the consumer market, the AI bubble is going to deflate, or ideally both. Just sucks waiting to see what happens if you didn't buy last year.

 

[A.Romero]

Computer hardware is fucked thanks to the AI landscape but at the same time I find it very exciting what they are managing to do with deep learning like DLSS and Framegen.

That tech is advancing pretty fast and it's yielding pretty cool results. I understand skepticism but after seeing it at work and with the blind tests I'm a believer. At some point in the future hardware will go back to normal (it has to) and algorithms will still be there so they will run on much more powerful hardware as well.