#1 Benchmarks Einstein@Home
Posted: Sat Nov 15, 2025 1:20 pm
Over the next months I will use this posting to publish my Einstein@Home benchmarks for the various SBCs in my possession. I will use this single posting, a 'work-in-progress' (WIP), to prevent spamming TSBT's forum completely, being a mere guest.
Einstein@Home has various apps suitable for ARM crunchers, such as the Apple-silicon-only Binary Radio Pulsar Search (Arecibo,GBT,A) (BRP4A), for both Apple M CPU (2.08 Apple_M-opencl) and Apple GPU (2.08 Apple_gpu-opencl). This will have to wait to be benchmarked by me until I have an Apple with a M CPU.
They also have Gamma-ray pulsar search #5 (FGRP5) for 64-bit ARM (1.16 beta test for SBCs, 1.14 for Apple MACs under ARM), and Binary Radio Pulsar Search (Arecibo,GBT) (BRP4) for no less than ten(!) ARM platforms -7 of which are under Android.
The remaining three are two Linux running on ARMv6 (armhf), one of which is optimized for NEON. I'd need my oldest Raspberry Pi in order to test that -and I might do that in order to see what OS version is the best, pure for the fun of it. The final one is BRP4 for 64-bit ARM.
So, in conclusion, I will test two Einstein@Home apps, on all but three of my ARM SBCs. This is because I experienced that the boards with less than 1GB per core tended to crash during running (The 2GB version of the quad-core Cortex-A57 Nvidia Jetson Nano, the 4GB hexa-core Odroid-N2+ running Armbian Linux and the 4GB hexa-core Radxa Rock Pi 4B+, running Armbian too). I will used a modified version of the utility of @davidbam because said utility is gone from the pages and will condense it to one line per test so that the eventual table will contain a wealth of comparable data in a very compact format.
Gamma-ray pulsar search #5 (FGRP5)
Binary Radio Pulsar Search (Arecibo,GBT) (BRP4)
Three conclusions so far:
Einstein@Home has various apps suitable for ARM crunchers, such as the Apple-silicon-only Binary Radio Pulsar Search (Arecibo,GBT,A) (BRP4A), for both Apple M CPU (2.08 Apple_M-opencl) and Apple GPU (2.08 Apple_gpu-opencl). This will have to wait to be benchmarked by me until I have an Apple with a M CPU.
They also have Gamma-ray pulsar search #5 (FGRP5) for 64-bit ARM (1.16 beta test for SBCs, 1.14 for Apple MACs under ARM), and Binary Radio Pulsar Search (Arecibo,GBT) (BRP4) for no less than ten(!) ARM platforms -7 of which are under Android.
The remaining three are two Linux running on ARMv6 (armhf), one of which is optimized for NEON. I'd need my oldest Raspberry Pi in order to test that -and I might do that in order to see what OS version is the best, pure for the fun of it. The final one is BRP4 for 64-bit ARM.
So, in conclusion, I will test two Einstein@Home apps, on all but three of my ARM SBCs. This is because I experienced that the boards with less than 1GB per core tended to crash during running (The 2GB version of the quad-core Cortex-A57 Nvidia Jetson Nano, the 4GB hexa-core Odroid-N2+ running Armbian Linux and the 4GB hexa-core Radxa Rock Pi 4B+, running Armbian too). I will used a modified version of the utility of @davidbam because said utility is gone from the pages and will condense it to one line per test so that the eventual table will contain a wealth of comparable data in a very compact format.
Gamma-ray pulsar search #5 (FGRP5)
| Board | Average Run Time in sec. | Average CPU Time in sec. | Credit per WU | Credit per Day/CPU | Credit per kWH |
|---|---|---|---|---|---|
| 8GB Odroid-M1, Quad-core Cortex-A55, 6.9W measured power draw, 4 WUs/CPU, running at 2000 MHz under Ubuntu 24.04 | 64,842 | 62,322 | 693 | 3,694 | 22,304 |
| 4GB Odroid-C5, Quad-core Cortex-A55, 3.0W measured power draw, 4 WUs/CPU, running at 2500 MHz under Ubuntu 22.04 | 48,421 | 45,028 | 693 | 4,946 | 68,697 |
| 4GB Raspberry Pi 4, Quad-core Cortex-A72, 7.2W estimated power draw, 4 WUs/CPU, running at 1800 MHz under Raspberry Pi OS Trixie | 46,142 | 45,734 | 693 | 5,190 | 30,037 |
| 4GB Raspberry Pi 4, Quad-core Cortex-A72, 11.0W estimated power draw, 4 WUs/CPU, running at 2000 MHz under Raspberry Pi OS Trixie | 37,998 | 37,667 | 693 | 6,303 | 23,875 |
| 8GB Raspberry Pi 5, Quad-core Cortex-A76, 12.7W estimated power draw, 4 WUs/CPU, running at 2400 MHz under Raspberry Pi OS Bookworm | 24,367 | 24,066 |
693 | 9,829 | 32,763 |
| 8GB Raspberry Pi 500, Quad-core Cortex-A76, 11.0W estimated power draw, 4 WUs/CPU, running at 2400 MHz under Raspberry Pi OS Bookworm | 23,689 | 22,880 |
693 | 10,110 | 38,296 |
| 4GB Raspberry Pi CM5, Quad-core Cortex-A76, 9.2W estimated power draw, 4 WUs/CPU, running at 2400 MHz under Raspberry Pi OS Bookworm | 23,705 | 22,894 |
693 | 10,103 | 45,758 |
| Board | Average Run Time in sec. | Average CPU Time in sec. | Credit per WU | Credit per Day/CPU | Credit per kWH |
|---|---|---|---|---|---|
| 8GB Odroid-M1, Quad-core Cortex-A55, 6.9W measured power draw, 4 WUs/CPU, running at 2000 MHz under Ubuntu 24.04 | 19,012 | 18,074 | 62 | 1,127 | 6,806 |
| 4GB Odroid-C5, Quad-core Cortex-A55, 3.0W measured power draw, 4 WUs/CPU, running at 2500 MHz under Ubuntu 22.04 | 12,985 | 12,336 |
62 | 1,650 | 22,920 |
| 8GB Raspberry Pi 500, Quad-core Cortex-A76, 11.0W estimated power draw, 4 WUs/CPU, running at 2400 MHz under Raspberry Pi OS Bookworm | 7,101 | 6,967 |
62 | 3,017 | 11,429 |
- BRP4 doesn't pay well, Credits per KWh-wise. Nor do the WUs land on all boards equally well.
- Overclocking a Pi4 might not be the most clever thing to do, budget-wise. Better buy a Pi5.
- The Odroid-C5 looks to be the most efficient board here.
The power draw estimates are based for the Pi 4 on an (obsolete) article on overclocking the Pi 4 (Bullseye days), and for the Pi 5 variants on the power draw during the HPL Linpack score benchmarks of Jeff Geerling. I will measure the real power draw with my Brennenstuhl-meter, but that takes time.