Theoretical bandwidth vs EFFECTIVE BANDWIDTH

Mihai56

Prominent
Feb 21, 2017
2
0
510
Hey,
I am trying to calculate the GPU bandwidth with this articale.
https://devblogs.nvidia.com/parallelforall/how-implement-performance-metrics-cuda-cc/
I have a GeForce GT 750M and Theoretical bandwidth is 64 GB and the effective bandwidth is just 30 GB.
Has anyone any idea why the difference is so big ?
 
Solution
That was an example. The limiting factor also doesn't have to be in th GPU. Also could be a power setting. Another part running hot, and few (read nearly nothing) things scale as well in the real world as they do on paper

Supahos

Expert
Ambassador
Usually it's a bottlnecking issue. Memory bus too narrow to fully utilize all memory or something similar. Especially on midrange cards they like to pack different things into them that sound good, make people pick them, but don't really add much performance. Other issue any time you get in a mobile environment is any part could be getting warm and slowing the whole system down.
 

Mihai56

Prominent
Feb 21, 2017
2
0
510
Ok, i think i understand.
So the difference appears because my GPU has only 128 bit buss and this is not enough to fully utilize all memory ?:D

The Nvidia say :

Theoretical bandwidth can be calculated using hardware specifications available in the product literature. For example, the NVIDIA Tesla M2050 GPU uses DDR (double data rate) RAM with a memory clock rate of 1,546 MHz and a 384-bit wide memory interface. Using these data items, the peak theoretical memory bandwidth of the NVIDIA Tesla M2050 is 148 GB/sec, as computed in the following.

BWTheoretical = 1546 * 106 * (384/8) * 2 / 109 = 148 GB/s

So the formula takes in consideration also bus width.
So if i have a larger bus width i have more GB transfer, less bus width less GB transfer, and from what i understand there should not be a bottleneck because the GB transfer is calculated for a particular bus width.
 

Supahos

Expert
Ambassador
That was an example. The limiting factor also doesn't have to be in th GPU. Also could be a power setting. Another part running hot, and few (read nearly nothing) things scale as well in the real world as they do on paper
 
Solution