VoluMark

BTW - could you change Exynos 5433 from Quad to Octa core? It uses all available cores if app support 8 threads
 
OlegSH said:
BTW - could you change Exynos 5433 from Quad to Octa core? It uses all available cores if app support 8 threads
Click to expand...
You are sure about that? The 4 A53 should do over 40. Adding to that 4 A57 we should get around 120 instead of 97.
 
Last edited:
I'm sure. You can take a look at my 7zip results for reference - http://forum.ixbt.com/topic.cgi?id=8:24730-15#335
8 threads total score - 8075 MIPS (total score is avg for compression + decompression mips)
4 threads total score - 5640 MIPS
8 threads are 43% faster
+ I checked System Monitor stats during the run and they show that all 4 A53 were fully occupied during the test
+ Samsung itself claims that Heterogeneous Multi-Processing is available starting from Exynos-5430
"The Exynos 5430 also supports big.LITTLE HMP (Heterogeneous Multi-Processing) processing – in simple terms, this means that all the eight cores of the processor can be active at the same time, and can also be individually turned on and off as and when required"
 
For your task gains could be even higher because 7zip is pure integer workload and A53 cores have pretty weak integer performance, but NEONs perf difference can be smaller for A53 and A57 clusters, especially if your render is cache friendly(since its 4 times smaller for A53 cluster)
 
Voxilla said:
Adding to that 4 A57 we should get around 120 instead of 97
Click to expand...
I think the only reason we got 97 score on 3W mobile chip is Heterogeneous Multi-Processing, A57 cluster frequency should be well below maximum, hence the huge run to run deviation
 
BTW - Could you add thread # option for your benchmark?
This way everybody will be able to verify whether his device supports HMP
 
Currently I'm using android_getCpuCount() to find the number of CPU cores.
I suspect that this returns 4 for the Exynos 5433 and it runs with only 4 threads.

I've been thinking about adding some 'pro' features like configuring the number of threads and run duration.
 
Thanks, that's more like it, nearly as fast as an A9. Must be due to all the Neon as A9 is in order for Neon.
I've removed the Allwinner score.
 
I've been busy porting VoluMark to iOS.
Preliminary results are:
  • 14 for iPad 2
  • 100 for iPad Air 2
The latter runs in 64 bit.
I'm using OpenGL to upscale the 500x500 rendered image to full screen.
That may make it a bit faster compared to Android (it may be Android internally already uses GPU for scaling bitmaps, iOS doesn't).
 
My Galaxy S5, running on the testing version of Cyanogenmod 12 (Lollipop) gets 76 points... which I find a bit strange, as the other posts for the Krait 400 only get 55. Phone was really cold, though.
 
I've posted a major new version:
  • Now runs in 64 bit if supported.
  • Possibility to configure number of cores used.
  • Analysis of CPU throttling by running x10 or x100 times.

An Exynos 7420 reportedly scores, with 8 / 4 / 1 cores
  • 123 / 93 / 27 in 64 bit
  • 101 / 74 / 24 in 32 bit
64 bit brings about 20% speedup
The additional 4 little A53 cores attribute about 33%.
 
Got up to 94 score with 8 cores
up to 60 with 4 cores
up to 45 with 2 cores
Seems I was right about 8 cores used by default in previous version after all, otherwise there is a major regression
 
OlegSH said:
Got up to 94 score with 8 cores
up to 60 with 4 cores
up to 45 with 2 cores
Seems I was right about 8 cores used by default in previous version after all, otherwise there is a major regression
Click to expand...

Indeed you were right about the 8 cores.
 
You must log in or register to reply here.
Back
Top