jon.nettleton at gmail.com
Sat Jun 2 03:03:13 EDT 2012
On Sat, Jun 2, 2012 at 3:12 AM, Samuel Greenfeld <greenfeld at laptop.org> wrote:
> On Fri, Jun 1, 2012 at 8:07 PM, Richard A. Smith <richard at laptop.org> wrote:
>> On 05/30/2012 03:34 AM, Yioryos Asprobounitis wrote:
>>> Most of the test had empty values but the informative ones (below) show
>>> that the XO-1.5 is better in basic integer operations and memory bandwidth
>>> while the XO-1.75 is better in float and double operations as well as in
>>> memory latency.
>>> I'm not sure how much this means for real life usage :-/
>> I'm very suspect of this measurement. The 1.5 has a hardware floating
>> point unit and the 1.75 is still using soft-float. Its extremely unlikely
>> that the floating point performance on 1.75 is better than the 1.5.
> Hard FP status depends on if Yioryos is running 11.3.1 or 12.1.0. Since he
> said "os10" by today's date I'm presuming 12.1.0.
> The Fedora 17 builds should be hard fp (armv7hl). The Fedora 14-based
> 11.3.1 builds are not (armv5tel / armv7l kernel).
Looking at those numbers I am quite certain that he is using F17 and
hardfp on the 1.75. Floating Point performance of the VIA vx855
chipset is a known limitation. It is something that they fixed in the
next generation vx900 chipset.
As for the integer math. Most the simple integer benchmarks come down
to clock frequency and integer pipelines, both of which the x86
processor beats the ARM on. Theoretically these operations could be
faster on our ARM processor as it has dual issue instructions, meaning
some instructions can be handled by the core at the same time. I
don't think we are taking full advantage of this functionality as I
believe gcc needs to be told about it and then software recompiled
with the proper tuning options. That patch was just submitted by
Marvell upstream, and we still have to test it.
Once I get everything built I will try to run some tests to compare
against your results, or possibly provide some binaries you can run on
Overall I believe the ARM chipset is a win on all fronts. Yes it
doesn't blow away the previous generation on raw processing power but
really the proper metric to look at in our usage scenario is
processing per watt. On top of that the ARM SOC is able to do things
like encode/decode 720p video using a special DSP. One of the design
principles of ARM is to not make a single chip that does everything
really fast, but instead combine a group of specialized chips that are
efficient at what they do yet provide a good overall computing
Thanks for the numbers, it should be good to revisit this as we move
forward with other optimizations.
More information about the Devel