[akazantsev]

> Lol I feel like no one has any attention span here. Tech shit is expensive in the beginning when it's new. It gets cheaper with time.

The funniest comment here. Have you seen the prices of the technical shit for the past two years? Dang, GPUs are not getting any cheaper, but more expensive with each year.

[SwellJoe]

That's an artificially inflated market. OpenAI and xAI bought everything for like two years into the future, partly to inflate the AI bubble, partly to lock-in a monopoly on the kinds of compute you need for AI, and partly to scale up actual operations. They can't realistically keep buying all the RAM in the world forever, the money has to run out eventually (though the market can remain irrational for quite a long time and can keep giving OpenAI and Apartheid Clyde money well past the point of reason).

[dwaltrip]

It’s a massive supply crunch. More production will come online.

[VorpalWay]

Probably not, at least for DRAM. The demand has historically been very variable, and building production capacity takes multiple years. Also, spare capacity is really expensive. Thus the memory manufacturers don't want to expand, betting on it being yet another temporary bubble.

Also, DRAM fabs are not really usable to make compute (CPU, GPU, etc in this context) silicon. The production lines and tech have diverged some decades ago. So unlike TSMC which can relatively easily retool for another customer, no such luck for the big DRAM manufacturers.

[mannanj]

How did Apple figure this out? Isn't the solution to this to sort of evolve to a unified memory architecture, which wins in speed and cost anyway?

[VorpalWay]

I haven't looked at apple specifically, but generally the approach taken with HBM on GPUs etc is to make multiple chiplets and connect them together very close and with very high bandwidth. The chiplets are made with different processes, optimised for each specific use case.

There are a few different variation, such as having a carrier chip below (as an interconnect), having a small PCB that everything is mounted to, or even vertically stacking chips (AMD's 3D v-cache does this on some of their CPU models).

[anthonypasq]

brother its been 1 year since claude code released. how fast are you expecting these things to happen? the physical world and hardware are still constraints. someone has to dig shit out of the ground to build these things.