density would double every year
I'm not a physicist, but I think they need to get smaller to do that.
I'm not a physicist, but I think they need to get smaller to do that.
GPU dies are quite large compared to CPU, ballpark 2X to 3X, so I guess we could get a generation to generation and a half of transistor doubling out of increasing the size of the chips at least.
But the question is, do workloads benefit from all these transistors? Server workloads, I guess so, you can get some pretty big Xeons, or lots of transistors per package with AMD and multi-chip packaging. Consumer workloads, I think the issue is more that consumers don’t have these massively parallel workloads that servers and workstations have. Consumers have stated their preference by not paying for more than a handful of cores per chip.
The obvious way to get more transistors per core is fewer, bigger cores per chip. But changing the cores is a big engineering project. And yield issues are already a big problem, which would only become worse with more complex circuits. But until we end up with one core per chip, I think we have hit an engineering/economics problem rather than a physics one.
And the engineering/economics problem is a really big one, I mean AMD has had great success by going in the opposite direction; breaking things up further with multi-chip packages.
That's not the question at all, this was about transistors getting smaller. Now you're way out in left field talking about how gpu chips being bigger than cpus somehow means that transistors don't need to get smaller, benefits, workloads, cores per chips and a lot of other irrelevant strange goal post shifting.
All I said was that transistors are still shrinking and density is increasing, I don't know where all this other stuff is coming from.
I’m just saying it is possible to follow Moore’s law. Physics didn’t limit us, engineering and economics did. Multi-chip packaging is a fantastic response to the engineering/economic problem of low yields.