Hands On & Initial Benchmarks With An Ampere eMAG 32-Core ARM Server

With only receiving this Ampere Osprey evaluation system earlier this week, I have only just begun with the benchmarking process. A lot more tests will come over the weeks ahead while today are some "teaser" benchmarks ahead of the more time consuming tests.

The first ARM server platform I had ever benchmarked (well, unless counting that build of a 96-core solar-powered ARM computer in a trash can) was the Calxeda EnergyCore ECX-1000... It's already been five years since the Texas ARM vendor went belly-up. As a reminder, the Calxeda servers were going to be comprised of multiple ECX-1000 nodes whereby for every four nodes were found on an "EnergyCard" that would be slotted in their servers and each server could have a varying number of EnergyCards. The multiple nodes had to communicate via MPI. Each of these ECX-1000 nodes that were leading for their time had four 32-bit ARM Cortex-A9 processors clocked up to 1.4GHz and consumed 1.5 Watt per processor or 5 Watts when paired with 4GB of DDR3 memory. In 2018 with Ampere it's now possible to have an ARM 32-core 64-bit processor clocking above 3GHz and not having to worry about "nodes" or any MPI configurations to take full advantage of the available CPU cores.

Fortunately, I still have some benchmarks archived via OpenBenchmarking.org from when I was benchmarking some of their servers in 2012 at their Austin headquarters. With the Phoronix Test Suite / OpenBenchmarking.org archiving each test profile (the test meta-data) and from there the particular software version of each program under test as well as all of the installation/setup routines, it was easy to reproduce those tests today on this Ampere eMAG server. Unfortunately though some of the benchmarks (including the MPI tests) from 2012 were not building with today's GCC compiler -- particularly with being on AArch64/ARMv8 rather than the 32-bit ARM days that were common a half-decade back. So for those wondering how an EnergyCore ECX-1000 node from this first prominent ARM server start-up compares to today's ARM server hardware in 2018:

In some cases it's a ~40x performance difference, well beyond the simple difference of core counts.

With eight DDR4 memory channels, the RAM performance is also a huge difference over older ARM server tech.

It's remarkable how far ARM server performance has evolved over the past 6+ years when Cortex-A9 cores were considered great while today with ARMv8 designs and the likes of this Ampere eMAG platform is substantially better performance even if factoring in the core differences.