Except the shrinks keep yielding smaller and smaller PPA improvements. It wasn't that long ago that we expected transistor count to DOUBLE every 18 months ..... remember?
Now, it takes a couple of node shrinks (N5->N3->N2) to get what ..... 25%? At this rate, we will be looking at 2 node shrinks to get 10% before long.
Aren't you mixing up performance increases with transistor counts?
The transistor counts still go up massively, for instance:
1. NVIDIA
GA104 chip (RTX 3070 Ti among others) is 17.4 billion transistors at 392 mm²
2. NVIDIA
AD103 chip (RTX 4070 Ti among others) is 45.9 billion transistors at 379 mm²
That's nearly 3x in two years. Now granted that's going from Samsung 8nm (more akin to TSMC 12nm than 7nm) to 5nm, so almost 2 node shrinks, but then again the transistors go up nearly 3x not 2.
The same is true regarding CPUs. Intel's transistor counts per core (even without caches) have gone up multiple times since Skylake. If I'm not mistaken, it has pretty much been 2x the transistors for 20% more perf/clock.
And the biggest difference compared to yesteryear is that you can't fire up all those "newly acquired" transistors and expect them to take as little power as the previous ones at the given chip-size (which was always the case till mid 2000s). This is the part that only goes down tens of percent at sweetspot clocks. So you. have to be clever in other ways (use more to get higher clocks bloat buffer sizes, add more cores / features that aren't used together all the time ...)
In the end you might get a bigger transistor budget (as long as it's not analog or SRAM, where gains are minimal) but they are a hell of a lot more expensive and can't be fired up all at once (unless you increase power limits).
forums.anandtech.com
Facebook
Twitter