Nvidias next gen GPUs are cooked
Hope Intel and AMD make reasonably priced alternatives for gamers, so that when the AI bubble collapses Nvidia goes bankrupt.
>rack redesignsAren't racks pretty standardized? yikes dawg
>>103220624Its Fermi all over again. All its missing is wood screws and 1.7% yields to complete the prophecy.
>>103220624Pretty fucking annoying. I'm ready to upgrade my 3070 and they're looking like they're going to fuck up for two generations in a row which is a first. 1 series->3 series->5 series was the supposed to be the smart way forward with jeets and chinks beta testing the cards between.
>>103220745>wood screwsQRD? this sounds funny
>>103220800Nvidia showed a mockup card that used some literal hardware store fasteners instead of actual flush mount fasteners that you could find anywhere on earth for pennies a piece.
>>103220815kek
Blackwell more like Blackened
>>103220624we had a perfectly adequate word already, it's "done".
this was obvious, nvidia is in pre ryzen scenario where all they do is make things bigger and draw more power but they have the cushion of tsmc doing the node shrinks for themthey've been carrying nvidia all this timei just hope amd will be there to capitalise with their chiplet research and experience
>>103220871pre ryzen intel scenario*
>>103220624There's already a thread for this.>>103220242
>>103220725No lolGoogle has their weird IKEA racks (no relation to IKEA furniture), little over 2x wider than a normal rack with machines running on 48v DC.Facebook is a champion for OCPOther big tech companies are probably also doing something weird.
>>103220624We already have a thread for this >>103220242
>>103220624There must be so many asians working 100 hour weeks on this issue
>>103220624nice source URL
>>103220624perfect timing for Apple's entry to the AI cloud game. thanks leatherjacket man, you got complacent and you'll get buttfucked like ayymd and intelaviv.
>>103220624Diversity is strength
>this garbage is "sold out" for the next 12 months
>>103222797>>103220624https://www.tomshardware.com/pc-components/gpus/nvidias-data-center-blackwell-gpus-reportedly-overheat-require-rack-redesigns-and-cause-delays-for-customers
>>103220624soon data centers will be water/oil submerged
>>103224201nvidia scammed every company super hard on these gpus
>>103224285What about shit submerged (Microsoft) ?
>>103224320Nvidia stock chart is a bubble. Finally bursting.
>>103220871>nvidia is in pre ryzen intel scenarioI thought the only problem Intel had before Ryzen was stagnation because they didn't need to take risks.>i just hope amd will be there to capitalise with their chiplet research and experienceI wish, but I'm not holding my breath.
>>103224581it is obviously a bubble but it hasn't popped yet. not until ai companies cant turn a profit.
Well, their 4000 series draws a lot of power on paper but still runs pretty cool, even stuff like the 4090 barely cracks 70°C, they have a really overkill cooler design.I don't see how they could go from being flush in one gen to cooking in the other, this must be a server problem, some 1U server running 100% all the time to train some bullshit AI model, having said that everything else will be shit for consumers, lack of VRAM plus crazy prices
>>103220624Dennard scaling sends its regards.>>103224626>not until ai companies cant turn a profitMore like until the shareholders realize that they can't turn a profit.
>>103224639ai is memory intensive and these are using next gen hbm. they run super hot compared to gddr6x or 7.
>>103223004Apple will eat NVIDIA alive if they can being their efficiency to the GPU/NPU space.
>>103224688Apple's efficiency was entirely dependent on early TSMC lithography access. They don't have some magic architecture.
>>103220624Laughs in 79xtx
>>103224702My m2 pro laptop (5nm) gpu at 25W runs qwen mlx 7b at 40 toks/sA 4090 (4nm) at 400W runs the gguf ollama version at 120 toks/sSo on an older node Apple's silicon is 5x more efficient/W for AI.And MLX isn't even one year old as a framework.>>103224688That's why they released mlx late last year. There's so much marketshare to eat from nvidia it's ridiculous, even if China gets EUV (which they will)
>>103224960efficiency is easy to get. down lock and under volt your gpu and presto it outperforms and is more efficient.nobody who gets a dedicated gpu for AI actually cares about efficiency they care about raw performance
>>103224688>>103223004Nice business opportunity if you can buy several racks of Mac minis and start selling time on them.
>>103225008you don't get 5x more efficient by downclocking.Apple is 3 generations ahead in gpu perf/watt compared to nvidia. it's software support they lack (for now)
>>103225017exolabs does distributed llms on apple silicon
>>103225188yea you do. you get 50% less power consumption at a loss of 8% performance.
>>103225188>in gpu perf/wattand 3 generations behind in actual performance
>>103225412show me a 4090 capped at 200w doing 100 tokens/s on a 7b 4 bit model
>>103225637you have a problem with what I said prove me wrong faggot.
>>103225901>prove me wrongthat's not how it works, troon
>>103220624sold my BAKA shares the moment i saw thisnvidia stock is gonna be nuked if blackwell is shit
>>103225997yes it does, I give you the argument you give me counter argument. it is on you to prove my statement wrong not on me to prove you right.
>>103226066no, troon. you (I assume) made the claim that Apple's efficiency was due to the node.I provided proofs that not only that's not true, but Apple is superior on an older node. (proofs exist online for both setups)To which you replied that>it's easy to obtain efficiency, just undervoltat this point it's up to you to prove that undervolting a 4090 at least matches the efficiency of the 5nm m2, filthy retard.
>>103220624I was worried about this then saw Dell went all in on Blackwell still. The companies literally do not give a shit. They will redesign their entire centers for these chips. Nvidia's moneyprinting machine won't stop
>>103225637Now compare the raster performance of the M2 to that of the 4090.
>>103226593>no, troon. you (I assume) made the claim that Apple's efficiency was due to the node.It is, when you take into account overall performance. Cherry-picking muh AI doesn't prove your point. The M2 also doesn't have AVX and is dogshit at software video conversion.
