Intel® oneAPI Base Toolkit
Support for core tools and libraries to build and deploy high-performance data-centric applications
Announcements
Welcome to the Intel Community. If you get an answer you like, please mark it as an Accepted Solution to help others. Thank you!
242 Discussions

<VectorAdd sample> How can USM be more performant than Buffers on CPU for both

Anonymous
Not applicable
329 Views

It's like 150% more performant than the buffer version on my i3-6100.

And 450% more performant than the buffer version on the dedicated GPU.

The buffer version and the OpenMP version do have the same performance tho.

 

0 Kudos
1 Solution
RahulV_intel
Moderator
307 Views

Hi,


Please specify your oneAPI base toolkit version and all the devices that you have used for comparison. (CPU, integrated GPU, discrete GPU)


Also, could you let me know how exactly did you arrive at those comparison numbers?



Thanks,

Rahul


View solution in original post

4 Replies
RahulV_intel
Moderator
308 Views

Hi,


Please specify your oneAPI base toolkit version and all the devices that you have used for comparison. (CPU, integrated GPU, discrete GPU)


Also, could you let me know how exactly did you arrive at those comparison numbers?



Thanks,

Rahul


View solution in original post

Anonymous
Not applicable
298 Views

Intel(R) oneAPI DPC++/C++ Compiler for applications running on Intel(R) 64, Version 2021.1 Beta Build 20201005

CPU: i3-6100

Integrated GPU: Intel® HD Graphics 530 Gen9

Discrete GPU: GTX 1060 3gb (not detected by OneAPI)

Ok nvm, I just mess with the benchmark I did last night.

So on discrete GPU target, I have 500% more FPS with USM than Buffer.

So on CPU target, I have 100% more FPS with Buffer than USM.

I did this benchmark with the project attached.

You can select the device and buffer/USM in Core.hpp

And for comparison with CUDA on my gtx 1060 3gb, I have 300% more FPS than USM on Intel® HD Graphics 530 Gen9

Anonymous
Not applicable
294 Views

I forgot to mention that I finally have the same performance than with OpenMP, with USM on CPU target.

RahulV_intel
Moderator
246 Views

Hi,


We are forwarding your query to the concerned team since it's related to performance. The team will get back to you shortly.


Regards,

Rahul


Reply