Potential graphics card shortage incoming... interesting idea. A.I. demand taking up TSMC space.

[Space Lynx]

jayz2cents just posted this, it felt a bit overdramatic to me at first, but then when he showed the picture of like 20 7900 XTX's stacked behind this new AI farm... got me thinking... and as he mentions in the video, profit margins on a AI silicon from TSMC are like 60 grand versus the paltry 500 bucks for a gamer gpu, same silicon, same time required at TSMC... why in the world would Nvidia make any gamer gpu's at all when they can allocate all TSMC time to AI and massive profit gains... if you watch the whole video, Jayz has a point here. I actually wasn't convinced until the end of the video though. AI farms are going to be the new thing for small business owners...

glad I built my rig when I did anyway, good prices and can semi-retire for a good five years as i watch all this play out. I still think jayz is overlooking that fact that a massive influx of gaming gpu's has hit all markets recently, and most people only play at 1080p still, i think the markets will still be fine honestly, but I recommend you watch the video and make up your own mind.

 

[freeagent]

New fabs will be online soonish, that will help take up the slack.

 

[Space Lynx]

New fabs will be online soonish, that will help take up the slack.

true, the Arizona plant was delayed recently due to shortage of skilled workers, I expect this will be the case for a lot of these new fabs.

I think the real saving grace is that so many gamers still game at 1080p, so what's already out there is plenty for most people. so this transition phase to AI while new fabs are being built is actually going to go more smoothly than what Jayz2cents realizes I think.

that being said, I am still glad I built my rig recently while the prices were good, now I don't have to worry about any of it.

 

[cmguigamf]

New fabs will be online soonish, that will help take up the slack.

But won't most of that be taken over for AI silicon as well?

 

[Space Lynx]

But won't most of that be taken over for AI silicon as well?

this is also true, AI silicon is extremely in demand, the business report that jayz2cents references in the video says they do not expect the shortage for AI demand to ease anytime soon.

I am genuinely surprised Nvidia has not moved 100% TSMC production to AI yet, the profits are insane.

 

[freeagent]

But won't most of that be taken over for AI silicon as well?

I thought about that right after I said posted lol..

Maybe Intel will save us

 

[Space Lynx]

I thought about that right after I said posted lol..

Maybe Intel will save us

Intel does seem to be in last place in the Ai game so maybe they will focus on ARC yeah

 

[AsRock]

Who did not know this were not paying attention, for one of many examples if nVidia are not relaying on selling video cards they can mark up as much as they like and AMD will follow along with Intel if they get their shit together.

 

[joemama]

profit margins on a AI silicon from TSMC are like 60 grand versus the paltry 500 bucks for a gamer gpu

I highly doubt these numbers, aside from that, I would think the market demand isn't that big.
As you see in the AI farm, people are using high end gaming GPUs for AI and it already works. Why would some ditch a cheaper GPU that can do a lot of general tasks and buy overly expensive cards that can only be used on AI?

 

[Space Lynx]

As you see in the AI farm, people are using high end gaming GPUs for AI and it already works. Why would some ditch a cheaper GPU that can do a lot of general tasks and buy overly expensive cards that can only be used on AI?

I don't, which is why I said before in this thread that I think the lower end market being the majority of gamers is probably going to save the day during this transition.

 

[caroline!]

Clickbait. A GAN benefits from actual CUs in a distributed parallel config, not gaming graphics cards, the processing power you can get out of these is a fraction of what you can get out of an actual CU.

 

[Space Lynx]

Clickbait. A GAN benefits from actual CUs in a distributed parallel config, not gaming graphics cards, the processing power you can get out of these is a fraction of what you can get out of an actual CU.

jayz2cents says this in his video as well, but AI doesn't have to be super hardcore for a small business use case. this isn't about big corporations buying gpu's, this is about small timers wanting to do AI stuff, some just as a hobby. I was skeptical as well until I saw the 20+ boxes of XTX about to be used in a AI farm.

 

[phanbuey]

jayz2cents says this in his video as well, but AI doesn't have to be super hardcore for a small business use case. this isn't about big corporations buying gpu's, this is about small timers wanting to do AI stuff, some just as a hobby. I was skeptical as well until I saw the 20+ boxes of XTX about to be used in a AI farm.

I honestly don't see the economic payback there... it would seem to my noobish experience with AI that it would be much cheaper and more efficient for someone to rent AWS or AZURE rigs for AI than try to build out a gfx farm....

Also isn't the value of AI in the training and the application? What is the application of a small ai farm?

 

[Space Lynx]

I honestly don't see the economic payback there... it would seem to my noobish experience with AI that it would be much cheaper and more efficient for someone to rent AWS or AZURE rigs for AI than try to build out a gfx farm....

Also isn't the value of AI in the training and the application? What is the application of a small ai farm?

don't ask me, I don't know about that stuff. the guy who bought the several XTX's for the AI farm is the same guy who reverse hacked the playstation 3, the one who jailbroke it, based on the screenshot jayz2cents showed in the video, I searched the name from the video.

so my guess is, nefarious reasons is your answer.

 

[caroline!]

jayz2cents says this in his video as well, but AI doesn't have to be super hardcore for a small business use case. this isn't about big corporations buying gpu's, this is about small timers wanting to do AI stuff, some just as a hobby. I was skeptical as well until I saw the 20+ boxes of XTX about to be used in a AI farm.

Right. AI is the new "blockchain technology". Hopefully the fad will die soon and only the people actually interested in ML for non-nefarious purposes will be able to buy cards, etc.

Pricks like jay 2 cents are the ones causing the shortage by hoarding cards just for the sake of spending their millions in a new fancy toy, and boasting about it in video so others like him do it as well.
Same happened with crypto that ended up being killed by normies wanting X, Y and Z from it, and corporations making it boring shitty traditional financial system full of stupid prerequisites, regs, KYC and nonsense, version 2.
Predictive models were fun until Google, Microsoft and other corps started implementing moronic, dumbified and limited versions of GPT because muh politics bs, until it wasn't interesting anymore.

 

[Space Lynx]

that's a good meme, lol

 

[Solaris17]

I am genuinely surprised Nvidia has not moved 100% TSMC production to AI yet, the profits are insane.

Not really. The margin is insane, we bought several million $$ worth of GPUs for AI but thats an investment that lasts a long time. As new GPUs come out we will not purchase that much. Instead we will offer a different product tier on the newest and most customers except those that want to help offset the cost will stay on the older clusters totally happy that there calculations are only a few hours slower.

Most big AI investments are very large companies or divisions of them that need tremendous compute time. Most of the AI startups you see arent that big. They will crunch for a few hours maybe $100k worth of resources and maybe MAYBE do that every few months; if not longer.

None of the GPU manufacturers can afford to solely pump out new AI silicon because the "AI market" isnt as big as you think. There are only a few big companies that can foot the bill for the gear to create the compute. They arent swapping it every 2 years like gamers. Its a massive (contractual) 1 time purchase for the most part.

 

[Space Lynx]

Not really. The margin is insane, we bought several million $$ worth of GPUs for AI but thats an investment that lasts a long time. As new GPUs come out we will not purchase that much. Instead we will offer a different product tier on the newest and most customers except those that want to help offset the cost will stay on the older clusters totally happy that there calculations are only a few hours slower.

Most big AI investments are very large companies or divisions of them that need tremendous compute time. Most of the AI startups you see arent that big. They will crunch for a few hours maybe $100k worth of resources and maybe MAYBE do that every few months; if not longer.

None of the GPU manufacturers can afford to solely pump out new AI silicon because the "AI market" isnt as big as you think. There are only a few big companies that can foot the bill for the gear to create the compute. They arent swapping it every 2 years like gamers. Its a massive (contractual) 1 time purchase for the most part.

this is true, I never thought of that way, very very few players that can foot the bill for any of those 70k hopper gpu's or w.e they called.

edit: especially now that the Chinese government is banned from buying any. (at least that's what all this trade war stuff is about I think, might be wrong)

 

[oxrufiioxo]

Already have a 4090 dgaf.

 

[Darmok N Jalad]

Not all fabs are on the latest node, so what would likely happen is that we peasant consumers will just have our GPUs assembled on more primitive nodes, at least until the AI craze subsides. Just think how much energy this stuff will consume. It might make mining look like a strand of Christmas lights.

 

[phanbuey]

Not all fabs are on the latest node, so what would likely happen is that we peasant consumers will just have our GPUs assembled on more primitive nodes, at least until the AI craze subsides. Just think how much energy this stuff will consume. It might make mining look like a strand of Christmas lights.

Wasn't there some POC superconductor at room temperature that just came out?

The current fab hotness is tomorrow's primitive node so if that's true we will just end up staying on current node for 10 years... it will be the decade of intel quadcores -- but on the GPU side. @oxrufiioxo will still be talking about that 4090 and how his fans are burning out in like 5 years.

 

[R-T-B]

Wasn't there some POC superconductor at room temperature that just came out?

Not proven or replicated yet, so probably not.

 

[Space Lynx]

Wasn't there some POC superconductor at room temperature that just came out?

The current fab hotness is tomorrow's primitive node so if that's true we will just end up staying on current node for 10 years... it will be the decade of intel quadcores -- but on the GPU side. @oxrufiioxo will still be talking about that 4090 and how his fans are burning out in like 5 years.

you should watch this

 

[oxrufiioxo]

Wasn't there some POC superconductor at room temperature that just came out?

The current fab hotness is tomorrow's primitive node so if that's true we will just end up staying on current node for 10 years... it will be the decade of intel quadcores -- but on the GPU side. @oxrufiioxo will still be talking about that 4090 and how his fans are burning out in like 5 years.

I sure hope I don't have it in 5 years....

If it gets that bad I guess a PS6 will have to hold me over....

 

[Space Lynx]

I sure hope I don't have it in 5 years....

If it gets that bad I guess a PS6 will have to hold me over....

honestly I love my steam deck so much, combined with so much backlog, I can see my current rig lasting me ten years if i dont let greed get to my head. and also i am 100% buying steam deck 2 oled, so that will hold me over even longer.