Ampere Altra Performance Shows It Can Compete With - Or Even Outperform - AMD EPYC & Intel Xeon
With making use of the Neoverse N1 cores, having an up-to-date compiler is important for being able to properly exploit the performance of the hardware. GCC 10 is in good shape for Neoverse N1 hardware as well as LLVM Clang 11. For the benchmarking in this article the testing so far has been done on GCC 10.
From the kernel perspective, recent Linux kernel as found on Ubuntu 20.04.1 LTS, 20.10, and even CentOS/RHEL 8 with its back-ports is in good shape. It is worth pointing out that recent versions of the Linux kernel do not enable the CPUFreq CPPC "boost" mode by default. This is required for getting the Ampere Altra to run at its advertised frequencies, but that just means writing to the CPUFreq "boost" sysfs control.
While these days x86_64 server Linux compatibility can be safely taken for granted compared to a decade or more ago, it's very pleasant seeing the good out-of-the-box support for the Ampere Altra / Mount Jade. Just the general recommendation always hold true that the newer the Linux kernel and compiler, among other key components, generally the best experience and performance at launch for new hardware. In the case of this testing today all the Intel / AMD / Ampere results were from Ubuntu 20.10 with GCC 10 and Linux 5.8 but follow-up articles will include multi-distribution comparisons, etc.
For this initial testing the Ampere Altra Q80-33 2P was going up against the Intel Xeon Platinum 8280 2P and AMD EPYC 7742 2P. Each server was running at its maximum rated memory channel and frequency. The Ampere Altra Mount Jade setup had 16 x 32 GB DDR4-3200MT/s Samsung M393A4K40DB3-CWE, the Xeon Platinum 8280 2P was running on a Gigabyte MD61-SC2 server platform with 12 x 32 GB DDR4-2933MT/s HMA84GR7CJR4N-WM, and the AMD EPYC 7742 2P on the AMD Daytona reference platform with 16 x 32 GB DDR4-3200MT/s 36ASF4G72PZ-3G2E2. A 4TB Micron 9300 NVMe solid-state drive was used across all the servers for storage with Ubuntu 20.10 on Linux 5.8 and with the GCC 10.2 compiler. Aside from a stock Ubuntu Server install, the only modifications were setting the CPUfreq governor on each system to performance and in the case of Ampere Altra ensuring the CPUFreq "boost" mode was enabled to match the default behavior of the Xeon/EPYC processors.
During the benchmarking process, the CPU thermals were monitored as well as the combined 2P CPU power consumption based on the exposed sensor interfaces for each of the processors under test.