Raspberry Pi 4 Benchmarks – Heatsink Edition

A few days ago, I ran some benchmarks in Raspberry Pi 4, and quickly found out that using the board without a cooling solution will cause serious performance issues, as in some cases my board was slower than Raspberry Pi 3 model B due to severe overheating.

After playing with LibreELEC yesterday, I’ve now reinstalled Raspbian Buster Desktop on the board, and fitted it with a largish heatsink and some old thermal paste.

Raspberry Pi 4 Heatsink
Click to Enlarge

So I’ll run benchmarks again with and without heatsink. I’ll only run sbc-bench this time.

SBC Bench Installation

Open a terminal windows or connect to the board through SSH and run:


Raspbian Buster will automatically fetch the latest operating systems packages upon first boot, but apparently not the latest firmware:


So I ran rpi-update to get the very latest firmware as well, and rebooted the board:


Normally, you should not have to do it, but Raspberry Pi 4 is really new, so I expect frequent changes at the beginning. Read warning in rpi-update:

WARNING: ‘rpi-update’ updates to pre-releases of the linux
kernel tree and Videocore firmware.

‘rpi-update’ should only be used if there is a specific
reason to do so – for example, a request by a Raspberry Pi
engineer.

Results with heatsink

Room temperature: 28°C. Uptime, idle temperature, and CPU freq:


Benchmark results:


Frequency/temperature monitoring during 7-zip:


It never went over the maximum 85°C temperature, and always under 80°C.

Results without heatsink

Now let’s remove the heatsink, reboot the board and wait a few minutes.

Room temperature: 28°C. Idle temperature:


So 4°C more at idle after about 3 minutes uptime without heatsink..

Benchmark results:


Throttling did occur, and indeed monitoring frequency and temperature during the benchmark shows temperature going close to 85°C, and real frequency dropping as low as 750 MHz to cool the system:

Comparison table and Pretty Chart

We can see that single thread benchmark are not affected by the presence of an heatsink, but the multi-threaded 7-zip compression is certainly impacted.

BenchmarkRaspberry Pi 4 “Naked”Raspberry Pi 4 “Heatsink”Ratio
memcpy2608.82540.197.37%
memset3715.63541.895.32%
7-zip44235397122.02%
OpenSSL
aes-256-cbc 16KB
64891.56k64782.34k99.83%

The differences for memcpy, memset, and OpenSSL are just statistical errors. But Raspberry Pi 4 with heasink is over 20% faster for 7-zip compression as it does not throttle during the test.

Raspberry Pi 4 Benchmarks Heatsink
Click to Enlarge
Support CNX Software - Donate via PayPal, become a Patron on Patreon, or buy review samples
Subscribe
Notify of
guest
25 Comments
oldest
newest most voted
tkaiser
tkaiser
1 year ago

The sbc-bench output also recorded that swapping happened while running the 7-zip benchmark (RPi folks for whatever reasons love swap on SD card). If %sys exceeds 1% or %iowait exceeds 0% then benchmark execution suffered for sure. The generated numbers could be slightly higher with ZRAM or if no swapping would’ve been occurred. The fastest way to switch to ZRAM on Raspbian/Debian is to: Shell wget https://mirrors.kernel.org/ubuntu/pool/universe/z/zram-config/zram-config_0.5_all.deb dpkg -i zram-config_0.5_all.deb reboot 123 wget https://mirrors.kernel.org/ubuntu/pool/universe/z/zram-config/zram-config_0.5_all.debdpkg -i zram-config_0.5_all.debreboot A better variant is to rely on Stuart Naylor’s replacement https://github.com/StuartIanNaylor/zram-swap-config though Also still undervoltage / frequency capping was reported. Based on your unboxing… Read more »

Spemer
Spemer
1 year ago

Thomas, you are a legend.

tkaiser
tkaiser
1 year ago

Well, in the past I found frequency capping always related to undervoltage (or maybe I remember wrong). Anyway, it’s ThreadX here so neither documentation nor sources available.

Only random comments on the Internet like https://www.raspberrypi.org/forums/viewtopic.php?f=63&t=147781&start=50#p972790 (which is related to the RPi 3 and 2, and the whole behavior doesn’t make that much sense any more)

blu
blu
1 year ago

As 7zip goes south with each consecutive run on ‘naked’, and the 3rd runs on heatsink/naked are scored ~5300/~4000 respectively, it’d be reasonable to expect a heatsink/naked ratio of well over 25% in the long run on this task.

Domih
Domih
1 year ago

Now, put a fan on top of the heatsink and do it again. Please. TIA!

Diego
Diego
1 year ago

Now dip it in liquid nitrogen and redo it again 😛

Marius Cirsta
1 year ago

I am also for the liquid nitrogen test. I think it’s the next thing for the rpi and pretty soon everyone will be doing it.

Domih
Domih
1 year ago

Hey, I love the smell of the spinning fan in the morning.

theguyuk
theguyuk
1 year ago

Flat bed Raspberry Pi waffle grill and Teasmaid.

TLS
TLS
1 year ago

Meh, liquid helium is what real overclockers use…

Diego
Diego
1 year ago

I’m getting old I got the impression. We should start a new OC xprize to reach -20 Kelvin.

tkaiser
tkaiser
1 year ago

When looking at the VL805 the chip itself seems not to be the problem: https://miro.medium.com/max/1400/1*KedwLPE9eXoeuwdciZpE0Q.png So most probably PCIe power management related? Ah, yes, lspci -vvv confirms ASPM Disabled (Active-State Power Management): Shell [email protected]:/# lspci -vvv 00:00.0 PCI bridge: Broadcom Limited Device 2711 (rev 10) (prog-if 00 [Normal decode]) Control: I/O- Mem+ BusMaster+ SpecCycle- MemWINV- VGASnoop- ParErr+ Stepping- SERR+ FastB2B- DisINTx- Status: Cap+ 66MHz- UDF- FastB2B- ParErr- DEVSEL=fast >TAbort- <TAbort- SERR- TAbort- <TAbort- <MAbort- <SERR- Reset- FastB2B- PriDiscTmr- SecDiscTmr- DiscTmrStat- DiscTmrSERREn- Capabilities: [48] Power Management version 3 Flags: PMEClk- DSI- D1- D2- AuxCurrent=0mA PME(D0+,D1-,D2-,D3hot+,D3cold-) Status: D0 NoSoftRst+ PME-Enable- DSel=0 DScale=1… Read more »

Stefano
Stefano
1 year ago

Which Heatsink is used? where can i get it?

Stefano
Stefano
1 year ago

Thank You 🙂

Exaga
1 year ago

Jean-Luc, Thanks for carrying out this RPi4 test. I’m so glad other people have found the same I have have regarding the device. When I first started asking questions and alerting people [on June 26, 2019] of the large amounts of heat that this device produces, not many believed me, and I was labelled a “Troll” by some others. Anyway, I’ve fitted a 28mm x 28mm x 20mm aluminium heatsink to my Raspberry Pi 4 Model B with a 30mm x 7mm @ 5v fan on top of it. The system is really stable now, and very fast. Thermal status… Read more »

boolean
boolean
11 months ago

I’ve just ran the benchmark as described and here are the results: http://ix.io/1TWC

I’m using a Wicked Aluminum enclosure which uses passive cooling to keep the temperatures down. It also helps that my workbench is made of metal, so heat from the case is passed on to the bench itself.

boolean
boolean
11 months ago

I actually did not take a closer look at the the results before posting, but you’re right indeed! I think it might has something tot do with the governor’s settings. On Raspbian the default governor is “ondemand”. I’ve set it to “performance” and did another run which shows more consistent numbers: http://ix.io/1TZA. The room temperature was ~26C. I also noticed something peculiar. If the SoC runs too cold, it also shows lower clock frequencies. When I initially fired it up, the sun hasn’t touched my workbench yet, so it was pretty cold to the touch. Since the Raspberry’s enclosure is… Read more »

willy
willy
11 months ago

It could also be something completely different, like systemd having fun in the background, or a crontab running stuff like a package update system. All of these will result in the measured CPU speed to be lower if the thread is scheduled out, and might explain why the measured frequencies are only slightly lower and not stable.

boolean
boolean
11 months ago

You would think so, but I was able to reproduce it by keeping the case (too) cool. When the SoC’s temperature drops below 33°C it doesn’t reach its maximum clock frequency (2 GHz).

Advertisements