An Extensive Look At The AMD Naples vs. Rome Power Efficiency / Performance-Per-Watt
Since the AMD EPYC 7002 "Rome" series launch in August we have continue to be captivated by the raw performance of AMD's Zen 2 server processors across many different workloads as covered now in countless articles. The performance-per-dollar / TCO is also extremely competitive against Intel's Xeon Scalable line-up, but how is the power efficiency of these 7nm EPYC processors? We waited to deliver those numbers until having a retail Rome board for carrying out those tests and now after that and then several weeks of benchmarking, here is an extensive exploration of the AMD EPYC 7002 series power efficiency as well as a look at the peak clock frequencies being achieved in various workloads to also provide some performance-per-clock metrics compared to Naples.
This article offers a variety of AMD EPYC 1P comparisons looking at the 7001 "Naples" processors to 7002 "Rome" principally for the overall AC power consumption / performance-per-Watt but also looking at the performance per clock for various workloads. The Naples processors based upon what I had available were the EPYC 7351P, 7401P, 7551, and 7601. The new Rome processors tested were the EPYC 7302, 7402, 7502, 7642, and flagship 7742 processors.
The majority of our AMD Rome benchmarking to date has been with their "Daytona" reference server while for the purposes of power consumption comparisons we waited on a retail board to ensure the numbers are indicative of what can be experienced by end-users. For this testing it meant using an ASRockRack EPYCD8 as one of the value AMD Rome 1P motherboards currently available in retail channels. This board was paired with 8 x 16GB Micron DDR4-3200 18ASF2G72PDZ-3G2E1 ECC Registered DIMMs and a 280GB Intel Optane 900p NVMe SSD.
For cooling the various processors under test was the Noctua TR4-SP3 heatsink, which has panned out very well through out AMD EPYC and Threadripper testing. The Noctua TR4-SP3 works out well and fits within 4U height requirements. We have about eight of the Noctua TR4-SP3 heatsinks in use across different EPYC and Threadripper systems to great success. Some thermal numbers are also included in this article for reference.
The EPYC testing was done within a Rosewill 4U server chassis with three Noctua 80mm exhaust fans and three Noctua 120mm intake fans.
For carrying out the power testing, we were measuring the AC "wall power" of the entire system in each instance using a WattsUp Pro power meter that was interfacing with the Phoronix Test Suite via USB for polling the AC power consumption every second and recording that for each test. In this article just the Naples vs. Rome power efficiency is being compared due to not having an identically configured Xeon Scalable server with monitoring the wall power. A Rome vs. Cascade Lake power comparison will happen once assembling a comparable server for accurate readings. Likewise, the peak frequency across all of the processors available cores were also polled every second for seeing what the peak frequency being achieved at any given point of time was on a per-test basis. The CPU temperature was also being monitored for ensuring no thermal throttling, which did not happen thanks to the Noctua cooling.
Now let's see how the real-world Naples vs. Rome power numbers look...