Can I measure the power performance of GPU based on Samsung Chromebook2?

Hi nightskyz,

Unfortunately production devices such as the Samsung Chromebook 2 it is not possible to get power measurements without hardware modifications.  The Odroid X-U3, however, is exactly the same SoC, Exynos 5422, as the Samsung Chromebook 2 but already has the hardware modification for power measurement when you buy it.  If you are interested in power measurement, I can recommend the Odroid X-U3.

As for OpenCL on the Chromebook 2, whilst there are 6 Mali cores in the device, they are arranged into two logical core groups with 2 and 4 Mali cores respectively, that appear as two different OpenCL devices. As such it is only possible to run your kernel on either the 2 or the 4, not all 6 (unless you create two contexts, one for each device and execute your kernel on both devices simultaneously).

With regards to the performance difference between the core group with 2 or 4 cores, without knowing what code you are running I can't really speculate as to why you wouldn't see a 2x performance with 2x cores.  Could I recommend taking a read of the Mali OpenCL optimization guide for T600 series: http://infocenter.arm.com/help/topic/com.arm.doc.dui0538f/DUI0538F_mali_t600_opencl_dg.pdf.  That may give you some useful information on how compute programs can be written in such a way as to scale better on Mali GPUs.

Hope this helps,

Rich

Hi nightskyz,

Unfortunately production devices such as the Samsung Chromebook 2 it is not possible to get power measurements without hardware modifications.  The Odroid X-U3, however, is exactly the same SoC, Exynos 5422, as the Samsung Chromebook 2 but already has the hardware modification for power measurement when you buy it.  If you are interested in power measurement, I can recommend the Odroid X-U3.

As for OpenCL on the Chromebook 2, whilst there are 6 Mali cores in the device, they are arranged into two logical core groups with 2 and 4 Mali cores respectively, that appear as two different OpenCL devices. As such it is only possible to run your kernel on either the 2 or the 4, not all 6 (unless you create two contexts, one for each device and execute your kernel on both devices simultaneously).

With regards to the performance difference between the core group with 2 or 4 cores, without knowing what code you are running I can't really speculate as to why you wouldn't see a 2x performance with 2x cores.  Could I recommend taking a read of the Mali OpenCL optimization guide for T600 series: http://infocenter.arm.com/help/topic/com.arm.doc.dui0538f/DUI0538F_mali_t600_opencl_dg.pdf.  That may give you some useful information on how compute programs can be written in such a way as to scale better on Mali GPUs.

Hope this helps,

Rich

Thanks richevans

By the way.. then which means even there's no way to guess the  power performance of GPU core unit based on SS chromebook 2?

now I know X-U3 can be a good alternative, but the problem is I do not have enough time to buy and do the same job on X-U3..

Because of my project deadline.. I really need the power consumption data of one compute unit..

Is there any other way to know power data based on frequency? Or Vdd and I? I thought probably ARM has kind of datasheet of Mali GPU...

I already googled it, but unfortunately I couldn't fine it..

Thanks a lot,

Woojung Myung

Hi nightskyz,

Unfortunately there is no way to get these power numbers without an instrumented device which the retail Chromebook 2 isn't.

The power consumption of the core depends on a huge number of variables, the SoC and its process, kernel configuration, workload, environmental factors, the nature of the work being done and as such we can't provide any estimates, even for where things like frequency are known.  You can of course calculate the power consumption if Vdd and I are known for the core, but there is no way for you to obtain this information from the device without making hardware modifications.

My only suggestion is, unfortunately, to get yourself an Odroid X-U3 which is already instrumented and will allow you to take power measurements yourself.

I'm sorry I can't be of more help in this case,

Rich