CoolPad, Scratch Wireless Partner On Budget Android Smartphone, Free Mobile Plan

Status
Not open for further replies.

SuperG

Distinguished
Jul 21, 2006
28
0
18,530
It obvious that if shader cores and TMU and Rops get much more numerous. Also Adress bus got wider. As memory speeds just don't keep up with these growing numbers. The memory bus also draw small amount of power per pin and trace and the logic at each end. But that for 512 which does add up.

The next step going to 1024 in steps. I see a routing problem of traces.
Much more layerd PCB can solve that bit. But the signal limitation get bigger which means need to get lower also latencies get bigger.

So HBM make sense.

To get a picture of this you need to know a bit about EM behavior of high frequency signals.
 
Status
Not open for further replies.