topic Quote:Robert Ioffe (Intel) in OpenCL* for CPU

Intel Iris GPU Boost in OpenCL

lucas_w_ — Tue, 19 Jan 2016 09:21:08 GMT

I have the mac mini, cpu is intel i7-4578U with Intel Iris 51000 whose max clock is 1275MHz. I find that GPU has been locked the clock, GPU clock is 750MHz. So I install the window OS in the mac mini. I found that the GPU clock just boost 800MHz when I run the OpenCL code, But for the oridinarily using, GPU can boost to the max clock. My question is that why the GPU can't boost to the max clock when I run the OpenCL code ?

Lucas,

Robert_I_Intel — Wed, 20 Jan 2016 00:01:47 GMT

Lucas,

Couple of questions:

1. Which version of the Windows OS are you running?

2. What version of the graphics driver are you running? Make sure to download the latest and greatest driver here: https://downloadcenter.intel.com/product/81494/Intel-Iris-Graphics-5100-for-4th-Generation-Intel-Core-Processors

3. How are you measuring performance and observing the frequency? Are you using Vtune? Make sure you set your power plan to performance.

Thanks!

Quote:Robert Ioffe (Intel)

lucas_w_ — Wed, 20 Jan 2016 01:46:38 GMT

Robert Ioffe (Intel) wrote:

Lucas,

Couple of questions:

1. Which version of the Windows OS are you running?

2. What version of the graphics driver are you running? Make sure to download the latest and greatest driver here: https://downloadcenter.intel.com/product/81494/Intel-Iris-Graphics-5100-for-4th-Generation-Intel-Core-Processors

3. How are you measuring performance and observing the frequency? Are you using Vtune? Make sure you set your power plan to performance.

Thanks!

Robert,

Thank you for you reply.

1. I install window 8.1 x64 ,

2.From the Device Manger, I get that the GPU driver is 10.18.14.4264.

3. From GPU Execution Units of Metrics Over time in the Vtune, I can see the GPU clock dynamically. Most of the time ,GPU clock is 800MHz.

4.I use the OpenCL event to measure execution time . Sometimes the kernel costs 6.3ms , Sometimes may cost 5.2ms.There is no other software to execute when I run the OpenCL kernel. And the GPU temperature which is measured by Intel Power Gadget is 46~70℃. Compare to AMD/Nvidia GPU, I think the fluctuation of GPU execution time is large. Why?

Hi Lucas,

Robert_I_Intel — Wed, 20 Jan 2016 02:02:25 GMT

Hi Lucas,

1. Could you please update your driver to 15.36.28.4332 from December 14, 2015? Your current driver is a bit dated.

2. Could you enqueue and run ~100 instances of the kernel and see what the average time is? It would be great if you could provide a Vtune screenshot of what you are seeing. The GPU frequency could go up and down (this is normal and this is probably the main cause of time variability of the kernel). Typically, when we measure the kernel frequency internally, we try to fix the GPU frequency with tools like PowerSpy. Your GPU frequency can go from 200 MHz to 1.2 MHz. The frequency of the GPU depends on what you are running on the CPU as well, since the total power of the platform will be maintained at a certain level.

What also could happen is that GPU couldn't ramp up to 1.2 MHz when you are running just a single kernel (that's why I am suggesting running a 100 of them). You could also try to run the following kernels https://software.intel.com/en-us/articles/optimizing-simple-opencl-kernels and see if you could see higher GPU frequencies in VTune.