Originally posted by creative
View Post
GPUs are fundamentally about maximizing perf/W. The amount of concurrency in graphics tasks allows them to take extremely efficient compute elements and achieve lofty performance numbers by stamping out thousands of them. In fact, the dies would probably be even bigger & the clocks even lower if manufacturing were cheaper. Alas, it's not. So, that forces them to use higher clock speeds than those which yield optimal perf/W.
Now, as for manufacturing tech, it's no accident that GPUs tend to lag behind most other types of chips. Something like a cellphone SoC is quite small (when compared with desktop CPUs, at least). Since the high-end models are also pretty high-margin parts, they can afford to be trail blazers on new process nodes (which have tighter supply, lower yields, and thus much higher production costs). Next, you usually see the desktop CPUs. And finally come the big server CPUs and GPUs. At least, that's how it's been for the past couple node shrinks.
If you were to look at the cost of fabricating current GPUs on say 10 nm, it wouldn't be economically viable for quite a while, yet. In fact, I think it's no accident that Nvidia chose to sell the largest of the Pascal family (i.e. the P100 and GP102) exclusively into the most price-insensitive markets, at first.
In summary: if you want GPUs to be more power-efficient, you'd make them even bigger and dial back the clocks a bit. But this would make them even more expensive, which was your original complaint. Finally, the realities of market competition and demand for the power-efficient computation that GPUs can provide means the high-end graphics cards will remain firmly pinned to the ~250 W ceiling imposed by the PCIe spec (which is pretty close to what's pragmatic from a cooling perspective). Sure, some cards push this to ~300 W, but there's not much beyond.
You'll find deeper analysis (and some recently-updated charts!), in this classic treatment of the subject:
In particular, I appreciate his analysis of memory scaling.
Comment