Sunday, 27 May 2012

Samsung Galaxy S2 (ARM Phone) vs Ubuntu PC performance


(this post has been updated in 2016)

It seems that many people assume that 1.2 GHz dual core mobile ARM CPU should be almost as fast as a PC CPU running on a similar frequency. They're wrong.

ARM cores are indeed more power efficient per square mm of surface on a same production process than Intel x86 and AMD64 architecture processors. Most of the efficiency comes from a simpler and more space efficient instruction set, but that advantage typically benefits only front-end of the CPU, which is not the biggest spender of those precious miliwatts.

The other reasons why modern dual or quad core mobile phones can run on a fraction of power that notebook or desktop (PC) CPUs need:

  • less computation units on CPU die (less SIMD, ALU, etc units)
  • smaller cache than PC CPUs
  • power gating parts of CPU (but laptop and desktop CPUs also do this for a number of years)
  • significantly slower DRAM interface  than PC CPUs, using slower DDR RAM (LPDDR2)
RAM speed significantly impacts many parts of phone performance. Executing complex JavaScript, image or video processing, Web page rendering are just some of the tasks that significantly benefit from having more RAM bandwidth. 

Your ARM device having significantly less of RAM bandwidth is also a big reason why you will probably avoid developing software on your new shiny ASUS Transformer Prime tablet/laptop (though I would certainly try:) )

So how much slower is your Android cell phone RAM than your PC RAM?

Unfortunately, I couldn't find any RAM bench-marking software that would run both on a Linux PC and on a un-rooted android device. There is a nice port of NBench, but NBench is a bigger benchmark and it needs some time before it prints out the one thing we need, the memory index. Also, it doesn't output MB/sec number, which is kind of unfortunate, since it's a really clear metric. 

So I found the really simplistic mbw (apt-get install mbw), made it even more simple (removed memcpy tests and left only the dumb array assignment part), and made Android NDK version of it.


Source here. Be sure to close any apps before running it on a PC or your phone. Default array size being copied is 20 MB (the app needs 40 MB to perform the test) to better support low memory devices. 

Here are some results (20MB array size, 20 repetitions avg, run "mbw -t1 20 -n 20", default settings on RAMbandwidth, on some larger boxes 200MB size was used ):
~12500 MB/sec Intel Core i7-6700, (DDR4 x2 2133 MHz), dedicated GPU
~12300 MB/sec -Intel Core i7-9700 (DDR4 x2 2133 MHz), driving 2560x1440@60Hz display, Ubuntu 19.04, Asrock H310M-STX DeskMini 310
~9000 MB/sec - Intel Core i7-8550U (DDR3 x2 2133 Mhz, Asus UX430UNR)
~9000 MB/sec - Intel Core i7-5600U (DDR3 x2 1600 MHz)
~8200 MB/sec - Asus N56JR (Intel  i7-4700HQ, 2x DDR3 1600 Mhz memory)
~6800 MB/sec - Intel Xeon E5-1650 v2 4x DDR3 1600 MHz)
~5400 MB/sec - Intel Xeon X3430, DDR3 memory, under moderate MySQL load( 2009)
~6000 MB/sec - Thinkpad X230 Core i5 3320M (2x  DDR3 1600Mhz)

~3800 MB/sec - Core i3-2310M 2x DDR3 1333Mhz
~2200 MB/sec - Intel Core 2 E8200, PC 6400 DDR2 RAM, Desktop PC (2008).
~1100 MB/sec - Intel Core duo L2400, PC 5300 DDR2 RAM on a  Thinkpad X60S laptop (2006). 

and our mobile contenders

~6000 MB/sec - Xiaomi Pocophone F1 (Snapdragon 845 varies between 5700-7000)
~6000 MB/sec - LG G5 (Snapdragon 820 4 GB LPDDR4 2016, varies between 5800-6500)
~1500 MB/sec - LG G3 (3GB D855 - It varies from 800-1700)
~1200 MB/sec - Raspberry Pi 3
~690 MB/sec - Doogee Valencia2 Y100 Pro
~530 MB/sec- Raspberry Pi 2
~500 MB/sec - Samsung  Galaxy S2 (2011)
~250 MB/sec - HTC Desire (2010)
~120 MB/sec - Raspberry PI (2012, under X, fbdev 720p it falls to ~90 MB/sec) 
~55 MB/sec - HTC Magic (2009, had to use smaller 10MB array size because of limited RAM available) 

Samsung Galaxy S2 sometimes reports around 440 MB/sec, and sometimes 550 MB/sec. I guess it depends where kernel allocates the memory, maybe one of the memory banks shares the bus with the GPU, GSM CPU or some other greedy device. 

It should be easy to post some test results of your own hardware, so please share. 

EDIT: Check comments for some more results