cancel
Showing results for 
Search instead for 
Did you mean: 

Archives Discussions

thesmileman
Journeyman III

Re: Long time between enqueue and execute on GPU

Well according to you I don't have performance problems it is just my expectations.

Now that I think about it why try to optimize I'll just do what you do and set really low expectations so I will always meet them.

It is funny because I am questioning long times between two driver events (not any code I wrote) which I have no control over yet you seem to be implying I am having performance problems.

0 Likes
ikuralenok
Journeyman III

Re: Long time between enqueue and execute on GPU

We experience similar issues with kernel startup time. We've tried AMD 7970 vs GeForce 590 vs GeForce 680 configuration and found that computation intensive operations are way faster on new AMD architecture (which is much more suitable for our tasks than Kepler). But all the time we've saved on computation is then wasted on startups of cheaper kernels. In our situation this mean that we have to invest in older nVidia Tesla cards instead of waiting FirePro based on new architecture. In our particular case we speak about ~200 professional cards. Let this number be a lower bound esteem on how much this problem does really cost. Can we expect any concrete feedback from AMD developers on the issue?

0 Likes
sh2
Adept II

Re: Long time between enqueue and execute on GPU

I could confirm it. Launch latency is about 200us and constant CPU overhead is about 20us per call.

Could we expect some progress with SI and new runtime?

0 Likes