POWER9 & ARM Performance Against Intel Xeon Cascadelake + AMD EPYC Rome
For those wondering how ARM and IBM POWER hardware stack up against AMD's new EPYC "Rome" processors and that of Intel's existing Xeon "Cascade Lake" processors, here is a round of tests from the POWER9 Talos II, Ampere eMAG, and Cavium ThunderX in looking at the cross-architecture Linux CPU performance currently in the server space.
Our AMD EPYC Rome benchmarks this month have been focused on the performance compared to earlier AMD EPYC and Intel Xeon processors, but given the broader architecture support on Linux and there also being significant interest in the likes of IBM POWER / OpenPOWER thanks to more open-source designs when paired with motherboards from Raptor Computing Systems, here are some initial numbers for ARM and POWER9 performance against the new x86_64 server CPUs.
On the POWER9 front was the Raptor Talos II server locally we frequently use for our POWER Linux server tests. That system, which is the only system tested in this article that is 100% open-source down to the design and firmware/microcode, was running with two 22-core IBM POWER9 processors yielding a combined 44 cores / 176 threads. These current-generation IBM processors with 22 cores per package are manufactured on a 14nm FinFET process and have a 190 Watt TDP while having a 2.75GHz base frequency and 3.8GHz turbo frequency. Thanks to SMT4 this Talos II open-source server offers competitive core counts to the AMD Rome processors. The IBM POWER9 22-core CPUs sell for about $2875 USD retail; more information via RaptorCS.com.
For the ARM testing, I used the bare metal access to the Cavium ThunderX and Ampere eMAG via Packet Host. The current generation Ampere eMAG offers 32 cores at 3.3GHz and within the Packet.com infrastructure they are using the Lenovo Falcon servers. We do have the pre-production Ampere eMAG we use for various tests on-premise though for this testing went with using the Packet hardware due to the Lenovo Falcon being their production server. Additional details on Ampere's current processors can be found via AmpereComputing.com. Also tested via Packet was the Cavium ThunderX X2 yielding 96 cores with a 2.0GHz clock speed. Packet's ThunderX servers were based on the Foxconn C2U4N_MB. Both Packet server had 128GB of RAM at their nominal frequencies, similar to the Intel/AMD server tests also being done with memory at their maximum channels and rated frequencies. All tests were done via the Phoronix Test Suite under Ubuntu Linux.
In this article we are just looking at the raw performance for these x86_64/ARM/POWER9 servers using various tests that operate well cross-architecture. Power efficiency / performance-per-Watt tests were not conducted due to the remote ARM testing. Likewise, there isn't performance-per-dollar metrics due to many variables at play when it comes to factoring in the ARM and POWER9 costs.