Originally Posted by Booty Warrior
Well, yeah. That's the point myself and box were making actually.
I was refuting those ideas specifically
because they are implicit in the notion of doubling bus width. A gpu with a 512 bit memory interface is capable of 512 bit data transfers to and from a single memory pool, which isn't the case in SLI. Obviously this isn't news to you, but apparently Zal didn't understand the difference.
Now you people are just making stuff up just for the sake of argument.
There are two GPUs, each one with a completely independent 256-bit, 192GB/s access to its own set of VRAM. Since it is completely independent, the total bandwidth
is effectively doubled.
Now, the efficiency
of this approach, and whether they read and write similar (redundant) stuff is a completely different thing.
We are not talking bus width
, either, we are talking bandwidth
, which is measured in GB/s
. We have two GPUs simultaneously reading and writing in their VRAM buffers all the time.
No single chip has access to double the bandwidth (same as for twice as many cores for that matter), but since there are two of them, time between frame refreshes is essentially halved, because by the time the first GPU is done with the first frame (if we are talking about AFR), the second GPU should be half way there with the second one.
In fact, in SFR mode, given that the top half of the screen is completely different
from the bottom half of the screen (unrealistic scenario, but we are not talking about efficiency here), it would indeed pretty much act
(not be) as a chip with double the cores and double the bus/bandwidth. But in reality it's not the case, of course, which is why AFR is more efficient (but with added latency issues).
Originally Posted by brettjv
Anyway, since it was more or less OT, and apparently 'over' since he stopped responding to you two ...
It's over for me, I will not be wasting time in an attempt to make myself clear another time.Edited by zalbard - 2/2/13 at 7:37am