Ampere Altra Performance Shows It Can Compete With - Or Even Outperform - AMD EPYC & Intel Xeon
For deep learning workloads like measured by TNN, the Ampere Altra 2P did manage to come successfully in front of the EPYC 7742 2P and Xeon Platinum 8280 2P servers. Not only was the Ampere Altra Q80-33 performing the best, but the combined 2P CPU power consumption was much lower than the Intel and AMD competition.
While both the AMD EPYC 7742 2P and Ampere Altra have eight channels of DDR4-3200 support, Ampere Altra Q80-33 was delivering very compelling performance for memory bandwidth tests.
Prior to receiving the Ampere Altra Mount Jade server and prior to seeing the performance potential with Apple's M1 chip on the desktop side, I figured the Ampere Altra performance would be like that of prior ARM server chips where in best case scenarios may put up a good fight against Intel/AMD but not outright exceed in both raw performance and performance-per-Watt for a variety of workloads. After seeing the results I was very surprised with how well the Ampere Altra Q80-33 2P performance is against the Xeon Platinum 8280 and EPYC 7742 servers. The performance exceeded my expectations where the Ampere Altra was able to collect wins in not only the performance-per-Watt but in the raw performance as well. Aside from software not yet optimized for the AArch64 architecture, the worst case was generally the Ampere Altra coming a bit behind the x86_64 competition but even then it enjoyed much lower power consumption than the x86_64 processors tested.
For cloud setups the Ampere Altra Q80 processors are also ideal for providing eighty physical cores per socket without any HT/SMT, which is good news given the vulnerabilities there on the Intel side, and allowing for very dense deployments. In benchmarking ARM Linux servers since the days of Calxeda ranging to 96-core ARM clusters in trash cans, it's easy to call Ampere's Altra the best ARM Linux server platform I've seen to date - not only from the performance side but from the Linux support/compatibility as well is effortless with modern AArch64 Linux distributions without any extra headaches. Bravo to all those involved at Ampere Computing.
So while Apple's M1 has done a good job for showing the ARM desktop performance potential, Ampere Altra's Q80 series shows what the performance can be like at the top-end for ARM within server and cloud deployments now actually providing serious competition against the current generation Intel Xeon and AMD EPYC servers. Looking ahead, Intel Ice Lake Xeon and AMD EPYC Milan are both expected next quarter. In fighting that Ampere already has their announced 128-core Ampere Altra Max (Mystique) processor that will retain socket compatibility with Ampere Altra and will be in production later in 2021. Given what we have seen with Ampere Altra, it will be very interesting to see how much of a performance leap they can achieve going forward with Mystique. At the same time the AArch64 Linux software support should be further maturing and more open-source projects supporting/optimizing for AArch64 given the successes of Apple M1, the ongoing Windows for ARM work, and ARM Linux servers continuing.
Ampere Altra is proving to be quite a promising ARM server platform, so stay tuned for more benchmark results in the weeks ahead on Phoronix looking at the performance against Amazon's Graviton2 ARM processor as well as other server/cloud benchmark workloads. As always with our relentless Linux benchmarking will be looking closer at the Ampere Altra performance under varying Linux distributions, GCC vs. Clang, Linux kernel releases, etc.
If you enjoyed this article consider joining Phoronix Premium to view this site ad-free, multi-page articles on a single page, and other benefits. PayPal or Stripe tips are also graciously accepted. Thanks for your support.