i realize i'm completely ignorant of the ecosystem but isn't the bottleneck what TSMC is able to produce? Does this just cut Nvidia out as a middleman taking profits?
There are lots of bottlenecks. TSMCs factory output is one, yes. Nvidia takes a big margin as the chip designer too. The physical building of datacenters is also slow. The electronic infrastructure for datacenters is in and out of stock all the time too. Electricity supply for them is questionable too.
Just wondering about the big picture logic. I would assume their chips would be inferior to Nvidias... so if 1/2 as good but removing 80% NVDA margins that would work out well.. assuming they could get them produced, but would guess NVDA monopolizes a lot of that volume- but maybe TSMC also prioritizes diversifying their customer base...
I would assume their chips would be inferior to Nvidias
Do we know that or is it speculation? Nvidias chips are GPUs - they're made for graphics and just happened to be useful for AI due to their massively parallel design. These AI chips are ASICs, they're tailor made for AI training/inference. If OpenAI managed to make ASICs that perform worse than GPUs that would be spectacular.
nope, NVDA have been in the general-purpose custom-garbage space-heater business for a while now.
GPUs are for the hobbyist, TPUs[1] are for enterprise
according to that link, apparently TPUs are specifically Google bricks, while NVIDIA makes "Tensor Cores"; honestly, if it burns like a brick, and it cracks like a brick, is it not just a brick!? ↩
Nvidia does not make TPUs. Nvidia makes GPUs. Nvidias architecture does have cuda cores but the general architecture isn't an ASIC architecture. You're simply misinformed.
their chips would be inferior to Nvidias... so if 1/2 as good but removing 80% NVDA margins
it's a ludicrously multidimensional design space
"1/2 as good" makes the remainder of the napkin twice as dumb, and it was already just a napkin to begin with.
even just deciding to sell chips that customers solder onto their own boards, versus vertical integration, is a qualitative decision that makes the comparison so difficult as to cause my kind of cynical response, rather than following up directly where you left off.
big picture napkin? you need at least an awareness of deliberate ink, versus unwanted stains.
fine-grain picasso-find napkin? there is the red ink of the professor, the black ink of the salesman, and the impeccableblinding whiteness of an infinitelylargenapkin... before anyone has begun eating, the waiter brings to the table un carafe du l'eau, huil d'olive, et un screw-cap bottle of bleach.
---
yes those are three separate URL anchors; casual web surfers click an infinite, caution evergreen students visit large, foolish data hoarders download napkin.
i realize i'm completely ignorant of the ecosystem but isn't the bottleneck what TSMC is able to produce? Does this just cut Nvidia out as a middleman taking profits?
There are lots of bottlenecks. TSMCs factory output is one, yes. Nvidia takes a big margin as the chip designer too. The physical building of datacenters is also slow. The electronic infrastructure for datacenters is in and out of stock all the time too. Electricity supply for them is questionable too.
Just wondering about the big picture logic. I would assume their chips would be inferior to Nvidias... so if 1/2 as good but removing 80% NVDA margins that would work out well.. assuming they could get them produced, but would guess NVDA monopolizes a lot of that volume- but maybe TSMC also prioritizes diversifying their customer base...
Do we know that or is it speculation? Nvidias chips are GPUs - they're made for graphics and just happened to be useful for AI due to their massively parallel design. These AI chips are ASICs, they're tailor made for AI training/inference. If OpenAI managed to make ASICs that perform worse than GPUs that would be spectacular.
pure ignorant speculation lol.
i just assume Nvidia would be better at this since this is their space... CUDA moat etc (although not sure how much of a moat that remains).
I'm unsure but we'll see for sure in a few months
nope, NVDA have been in the general-purpose custom-garbage space-heater business for a while now.
GPUs are for the hobbyist, TPUs[1] are for enterprise
according to that link, apparently TPUs are specifically Google bricks, while NVIDIA makes "Tensor Cores"; honestly, if it burns like a brick, and it cracks like a brick, is it not just a brick!? ↩
Nvidia does not make TPUs. Nvidia makes GPUs. Nvidias architecture does have cuda cores but the general architecture isn't an ASIC architecture. You're simply misinformed.
it's a ludicrously multidimensional design space
"1/2 as good" makes the remainder of the napkin twice as dumb, and it was already just a napkin to begin with.
even just deciding to sell chips that customers solder onto their own boards, versus vertical integration, is a qualitative decision that makes the comparison so difficult as to cause my kind of cynical response, rather than following up directly where you left off.
i'm admittedly a retard trying to understand the big picture napkin math view lol
big picture napkin? you need at least an awareness of deliberate ink, versus unwanted stains.
fine-grain picasso-find napkin? there is the red ink of the professor, the black ink of the salesman, and the
impeccableblinding whiteness of an infinitely large napkin... before anyone has begun eating, the waiter brings to the table un carafe du l'eau, huil d'olive, et un screw-cap bottle of bleach.---
yes those are three separate URL anchors; casual web surfers click
an infinite, caution evergreen students visitlarge, foolish data hoarders downloadnapkin.Jalapeño on a stick?
https://twiiit.com/OpenAI/status/2069770172802773292
deleted by author
never trust a search engine, that's a "dead" link [to a parked domain]
this one might work better...