cancel
Showing results for 
Search instead for 
Did you mean: 

Archives Discussions

sunny_engineer
Journeyman III

High memory latency for memory copy between host and devcie

Hey,

I was trying to measure the bandwidth for the data transfers on AMD Fusion A8-3850 and I though that because of the absence of the discrete GPU, there is no cost of PCIe. So I was expecting to have very high bandwidth as compared to the numbers on a discrete GPU device. But I got these numbers for the fusion

pinned: H_to_D (4.9GB/s) D_to_D (16GB/s)

page-able: H_to_D (3.8GB/s) D_to_D (16GB/s)

and these one for the NVIDIA GPU

pinned: H_to_D (5.5GB/s) D_to_D (20GB/s)

page-able: H_to_D (4.1GB/s) D_to_D (20GB/s)

Is my understanding wrong or am I missing something here??

Thanks


Shoaib

0 Likes
3 Replies
kcarney
Staff

Hello,

I'm sorry it is taking so long to answer your question. I'm trying to find someone who knows the answer. Thanks for your patience.

Cheers!

Kristen

0 Likes

I'm sorry that I still haven't found anyone who knows the answer to your question. Since it is taking so long, let's try 2 other venues:

1. Try posting your question on the technology forums:  http://forums.amd.com/forum/categories.cfm?catid=12&entercat=y

2. Also, you can contact AMD support either by email (http://emailcustomercare.amd.com/) or by phone (http://support.amd.com/us/contacts/Pages/global-technical-support.aspx)

0 Likes
yurtesen
Miniboss

If DtoD is in between the same device, 20GB/s is quite little for any discrete GPU. There must be a problem in your program?

Your numbers does not seem very far from these did you have a look at this presentation? :

http://developer.amd.com/afds/assets/presentations/1004_final.pdf

Did you try the bandwidth tests in the SDK samples?

0 Likes