During the process of testing and validating the Parallella boards we ran a lot of open benchmarks from http://openbenchmarking.org/. It was surprisingly hard to find benchmarks that met our requirements and in our opinion the Phoronix test suite is the best option for open and transparent benchmarks that:
- Must be open source
- Have a permissive usage and reporting license
- Run out of the box
- Don’t allow tweaking of source code and compiler switches
- Have a transparent results reporting mechanism
One of the nice features of the Phoronix testsuite is that it automates the process of comparing benchmarks against old results and reporting data back to http://openbenchmarking.org.
To install the Phoronix test suite in Ubuntu:
sudo apt-get install phoronix-test-suite
To benchmark against an existing result:
phoronix-test-suite benchmark 1302242-BY-1205272AR49
In this case, we wanted to compare the performance of the dual core 667MHz ARM A9 within the Zynq SOC on the Parallella board against the Raspberry Pi results that had been reported here.
Of course we knew from the start that Parallella results would not be earth shattering since the benchmarks don’t use the FPGA logic or the Epiphany coprocessor and are mostly single threaded.
Below you can see some of the results. The full set of results are available at http://openbenchmarking.org/result/1307314-AR-1302242BY85
In our view, there really weren’t too many surprises. We should always be very careful in drawing conclusions about benchmark data, but two things seem clear from these results: 1.) ARM A9 is significantly faster than ARM11 and 2.) Phoronix benchmark scores are strongly correlated with operating frequency.
We think it’s reassuring that the ARM sub-system on the Zynq gives pretty decent baseline performance that will be “good enough” for many applications. Once we have boards out in the field, we hope our Kickstarter backers will run their own benchmarks and report results openly. As you know, it’s imperative to verify any benchmark results that come directly from a semiconductor vendor.:-)
[NOTE]: The Phoronix benchmark reports the Zynq running at 733MHz but the Zynq is actually running at 667MHz. This was due to us using a 30MHz input clock instead of a 33MHz on the Gen0 version of the board. We moved to 33MHz for the Gen1 and decided that we didn’t want to spend time figuring out why the clock was running slow. (i.e. if you measure 30 seconds of wall time, the Parallella board reports that 28s has passed).