Hacker News Reader

I'm pretty sure GPUs have their own RAM installed that is usually higher spec than the motherboard RAM.

Higher in some respects (bandwidth), lower in others (latency, even though ordinary DDR5 is already no speed demon there and LPDDR5 is worse). At least from the spec sheet, these kinds of RAM are so different that I don’t really understand how demand for one can cause a shortage of the other, unless they are competing for the same manufacturing lines.

FWIW: GDDR is not higher latency than DDR. It just seems that way because the GDDR interface clock is much higher, so the number of clocks are higher too. But in terms of wall clock time, the latency is very similar.

Which makes sense: the latency is determined by the underlying storage technology and the way to access that storage, which the same for both.

They are competing for the same manufacturing lines.

RAM manufacturers are switching lines over from DDR to make HBM.

HBM works really great in GPUs too. In fact if I have to spend 5 grand on one it had better come with HBM lol.

Reply on news.ycombinator.com