[SOLVED] Memory bus on gpu

Beardy suburb

Commendable
Mar 7, 2022
25
4
1,535
So after some research I found that many cards have different memory bus values even when 1 of these cards in question is theoretically a way better card.
Rtx3060ti has 256bit
While rx6700xt 192bit
But then again another version of the 3060ti has 192bit
That being said does memory bus value make any difference in performance. I totally get the fact that more bus means more data transferred but I can’t see to find any info about impact in performance. Maybe the latency in which we see frames in our monitors?
 
Solution
3060 is 192, 3060ti is 256. It might have been a misprint. Anyways, this is just 1 piece of the equation when only the end result matters and that's actual performance. The main reason you won't find much info is because it's something you can't compare. Because of how different architecture utilize memory bandwidth, you can only compare it to the same architecture of gpu. You can't even compare to different series from the same company as they change memory algorithms. You'll see bandwidth increase as you get to higher end models. This is just part of being a well designed gpu and nothing notably to think about when choosing a gpu. You may look at memory bandwidth like vram depending on the usage but sometimes that's like putting on...
3060 is 192, 3060ti is 256. It might have been a misprint. Anyways, this is just 1 piece of the equation when only the end result matters and that's actual performance. The main reason you won't find much info is because it's something you can't compare. Because of how different architecture utilize memory bandwidth, you can only compare it to the same architecture of gpu. You can't even compare to different series from the same company as they change memory algorithms. You'll see bandwidth increase as you get to higher end models. This is just part of being a well designed gpu and nothing notably to think about when choosing a gpu. You may look at memory bandwidth like vram depending on the usage but sometimes that's like putting on blinders and missing real world performance.

In previous generations, there were a few gpu models that were similar with different bus width, and the effects were obvious; increase memory bandwidth and performance increased but only if that was a bottleneck of the gpu. A lot of r&d is done for them to select the corresponding bus width to gpu.
 
Solution
GPUs are data hungry processors. The amount of peak theoretical flops means that they're shuffling around terabytes of data every second. Of course, most of that is being recycled for other operations (like calculate A + B and add that output to something else) but when you can crunch that much data at once, you need to be able to feed it.

It's similar to why Xeon and EPYC/Threadripper can have quad and hexa channel memory systems. Their higher core counts means they need more RAM bandwidth to feed the cores.
 

TRENDING THREADS