The performance-per-Watt numbers and other really interesting comparative benchmarks are still forthcoming (a bit preoccupied at the moment with the Linux activities in the Czech Republic) from the tests that I carried out earlier this month at the Calxeda office in Austin while uploaded today are some power consumption / thermal data, compiler benchmarks, and some I/O workloads.
The most interesting results I uploaded today to OpenBenchmarking.org were the Calxeda ARM Server Sensor Monitoring numbers. For both the 1.1GHz and 1.4GHz nodes while running Ubuntu 12.10 with the Linux 3.5 kernel, some Phoronix Test Suite benchmarks were carried out while the testing framework was measuring the SoC power consumption and core temperature for the quad-core Calxeda node.
Beginning with Phoronix Test Suite 4.2-Randaberg there is IPMI sensor handling for being able to automatically poll the reported power consumption (in Watts) per Calxeda node along with the core temperature by simply setting the MONITOR=all environment variable prior to test execution.
At the SoC level, the 1.1GHz node under full load across all four ARM cores was at about 4.2 Watts -- certainly below the advertised 5 Watt envelope for the quad-core SoC. (This though doesn't factor in the disk drive power consumption, etc.) The 1.4GHz quad-core ECX-1000 node that's still undergoing work was at 6.8 Watts for the same test.
Under the C-Ray workload, the 1.1GHz node was running safely at about 36C while the 1.4GHz node was slightly warmer at 45C.
When running Apache the 1.4GHz quad-core node power consumption spiked to 7 Watts while the standard 1.1GHz configuration was still sub-5 Watts.
Across the many tests run, the Phoronix Test Suite found the average 1.1GHz quad-core SoC temperature to be 36C and 45C for the 1.4GHz edition.
The average power consumption was 4.0 Watts and 6.4 Watts for the 1.1GHz and 1.4GHz cards, respectively. Again, see more data within its result file or run MONITOR=all phoronix-test-suite benchmark 1210188-RA-CALXEDASE25 to facilitate your own side-by-side Linux performance comparison.
For those interested in LLVM/Clang vs. GCC numbers for reference on ARM hardware, see 1210189-RA-LLVMCLANG48. However, out of this result file there isn't anything too interesting. For a more exhaustive look at different compiler options for ARM, I have more data in this earlier article.
Lastly, I have some Calxeda I/O benchmark results for one of their nodes connected to a SAMSUNG MZ7PC256 disk drive. Compared to the PandaBoard ES and other ARM development boards that are often running off SD/SDHC storage, the Calxeda ECX-1000 has a native Serial ATA 2.0 disk controller so the storage capabilities are much greater and actually worthwhile for server purposes.
Stay tuned for more ARM Linux server benchmark results.