Quad-Core ODROID-X Tested Against PlayStation 3

[const]

And you are so sure because you tried yourself to implement DST on ARM or have a reference that proves you're not just doing unsustained claims?

i've tried it on high-end x86 CPU, as well on PS3 and i already told the results:

* PS3: significantly faster than real-time, i.e. few minutes of multichannel DST data are decoded for less than those few minutes; of course that's with its SPE optimized decoder.

* high-end x86 CPU with the reference decoder:

http://standards.iso.org/ittf/PubliclyAvailableStandards/c043465_ISO_IEC_14496-5_2001_Amd_10_2007_Reference_Software.zip

few minutes of multichannel DST data are decoded for several hours, yes several hours. so, it's like PS3 is hundreds if not thousands times faster

so, unless you state that there are ARMs with faster computational power than high-end x86 CPU then ARM performance would be worst than x86 CPU with the reference decoder - in fact i can try it for sure since i have several Samsung and Texas Instruments ARMs (probably one of the fastest at least those 2 companies are making), but i see it as pointless. once again, optimizing the reference decoder for x86 CPUs and multithreading it gives big performance bump and high-end x86 CPU get closer to PS3, but then is the price - such x86 system costs times the costs of PS3. so, bottom line is that ARM will be 3rd for sure and doing real test only will show 3rd by how big margin compared to the 2nd - i believe it would be very big.

[ldesnogu]

takes several hours to decode few minutes of audio. of course, there is optimized version for x86, which performance is close to that of PS3. however, my points here are two - first, the importance of optimizations; second, i'm quite sure ARM CPUs in ODROID-X will need at least hours if not days to deal with DST multi-channel decoding, showing with that their real performance for multimedia applications. so, also that as suggestion - build the reference DST decoder and run it on ODROID-X - even if you wish try to optimize it as much as possible for ODROID-X hardware and see the results.

And you are so sure because you tried yourself to implement DST on ARM or have a reference that proves you're not just doing unsustained claims?

[const]

let me put it simple - i'm absolutely sure based on the real facts that those numbers are just plain none-sense and are not even close to reflect the real power of PS3, not only as gaming console and graphic experience, but as computational power too. so, it's delusional to expect ODROID-X either Quad or Dual core to give anything close to PS3 as gaming experience. last, but not least, there are 2 simple reasons for those numbers: PS3 HyperVision (i.e. Linux is not running directly on the PS3 hardware, but through the HV) and lack of (SPE) optimizations - and if there are people that already forgot, doubt PS3 computational power and believe those numbers (which are nothing more than just result of pure wrong testing) give any real judgement of the performance experience the 2 compared devices give - a reminder that Linux cluster of several PS3 running properly written software are exceeding the power of "100+ Intel Xeon core based traditional Linux cluster":

http://en.wikipedia.org/wiki/Cell_%28microprocessor%29#Cluster_computing

which is easily verifiable by anyone having several PS3 with Linux in his home and following: http://www.ps3cluster.umassd.edu

so, those numbers result of the benchmark are excellent example of benchmark that is so wrong and not even remotely close to real-life experience with those hardware devices in question.

i have another example too - there is 'Direct Stream Transfer' (DST) - a lossless compression variant for DSD-encoded audio data (as how audio is stored on Super-Audio-CDs) - when DST contains multi-channel data it's very computational demanding to decode: PS3 do it faster than real-time, very fast x86 computer using reference DST decoder implementation as part of the ISO standards:

http://standards.iso.org/ittf/PubliclyAvailableStandards/c043465_ISO_IEC_14496-5_2001_Amd_10_2007_Reference_Software.zip

takes several hours to decode few minutes of audio. of course, there is optimized version for x86, which performance is close to that of PS3. however, my points here are two - first, the importance of optimizations; second, i'm quite sure ARM CPUs in ODROID-X will need at least hours if not days to deal with DST multi-channel decoding, showing with that their real performance for multimedia applications. so, also that as suggestion - build the reference DST decoder and run it on ODROID-X - even if you wish try to optimize it as much as possible for ODROID-X hardware and see the results.

[elg2001]

I took "Computer Hardware Design" in the Fall of 2004, and my professor was a senior design engineer on PowerPC at IBM. His name was Pra-something (sorry, it's been a long time) Kudva. I would talk to him pretty often after class about his work at IBM, but I could never get him to admit what project he was working on. I'm fairly certain it was either the Cell or the Xbox 360 CPU (which are very similar not counting the SPEs); I think "Power5" slipped out at some point. He was pretty secretive about it.

Regardless, he was an amazing teacher. I wish hardware jobs were more plentiful. It's so fun to design!

[elg2001]

As the submitter of these results I would like to clarify a few things regarding the PlayStation3 and these tests.

The Playstation3 has a general purpose PowerPC core capable of running three threads. In addition, the system has six other threads called SPEs which ARE accessible in GNU/Linux through the CELL SDK and given that you write specific code for them. If the software is properly implemented 160 GFLOPS of FP performance are possible.

The tests that were run in PTS employed the PPE core only (one PowerPC core capable of two threads).

Thanks for clarifying; much appreciated. I thought the main core was single-threaded. You learn something new every day.

[toguro123]

How come they didn't try to invent/find a way to abstract the co-processors and make it a more general purpose chip.

Its kind of boring that we only have x86 The whole monopoly in the processor and graphics chips is kind of sad in a way.

I would like for example to see a Tilera based machine.

Well, IBM killed the CELL project short after the release of the CELL SDK 3.0, so it was not developed any further. Maybe such abstraction was too difficult. What would have been nice to see is being able to access the SPE through C++ function calls (like CUDA) instead of SPU handles. Remember the PPE unit and the SPE units are of a different architectures. Thus you need a different compiler for each, then you need to link the binaries. Its quite tricky.

I have uploaded benchmark results for the UltraSPARC architecture with GNU/Linux as well as Solaris. Check them out if you have a chance!

[89c51]

How come they didn't try to invent/find a way to abstract the co-processors and make it a more general purpose chip.

Its kind of boring that we only have x86 The whole monopoly in the processor and graphics chips is kind of sad in a way.

I would like for example to see a Tilera based machine.

[toguro123]

Clarification

As the submitter of these results I would like to clarify a few things regarding the PlayStation3 and these tests.

The Playstation3 has a general purpose PowerPC core capable of running three threads. In addition, the system has six other threads called SPEs which ARE accessible in GNU/Linux through the CELL SDK and given that you write specific code for them. If the software is properly implemented 160 GFLOPS of FP performance are possible.

The tests that were run in PTS employed the PPE core only (one PowerPC core capable of two threads).

[elg2001]

As for why the PS3 wins in some tests, it most likely has to do with the fact that it has a huge advantage in memory (RAM) bandwidth. Its RAM runs at 3.2 GHz!

Has anyone managed to fully hack a PS3 to the point where the hypervisor could be bypassed entirely?

[elg2001]

Comparing the ARM system to the PS3 is absolute nonsense. Not only is the PS3 running a linux distro from THREE YEARS ago versus the ARM system on the latest kernel/gcc, it has to do everything through a hypervisor. A huge downside to that is that the GPU is inaccessible. So you have no GPU and a hypervisor that helps throw some performance out the window. If you could ACTUALLY install linux on a PS3 (on bare metal), its numbers would be much higher. Would the ARM system still win? No idea. But this setup is far from scientific; it was probably just done for fun.