Benchmarking Raspberry Pi 5

There is nothing so important to an engineer who gets their hands on a new bit of hardware as to immediately try and figure out how fast it can go. On the other hand, there is nothing more hotly debated between engineers as to how to measure how fast things go. But like a lot of things, what you should measure really depends on how and why you’re going to be using the hardware in the first place.

As our American friends might say, your mileage may vary.

A sneak look at the black version of the new case for Raspberry Pi 5

So let’s see how fast the Raspberry Pi 5 can go!

Benchmarking with Geekbench

There are many benchmarking frameworks, but this one is mine, and why it’s mine comes back to the question of how and why. Instead of just straight-up crunching some pointless numbers, Geekbench takes a stab at running tests that attempt to reflect how people normally use their computers: loading a website, rendering a PDF, and adding filters to an image.

Like a lot of other benchmarks, Geekbench headline scores are split into single-core and multi-core scores. However, just due to the nature of the benchmark, Geekbench scores can vary a lot from run to run. Setting force_turbo=1 in config.txt and avoiding things like software updates somewhat reduces the noise between runs, but the best strategy is to average multiple runs. Which is what we’ve done.

If you’re running the 64-bit distribution of Raspberry Pi OS, it ships with 16KB page size enabled, which gives slightly improved performance at the expense of compatibility with older ARMv7 32-bit binaries.

So we’ve also gone ahead and tested the Raspberry Pi 5 using both 16KB and 4KB page size. To do that we’ve used an internal build of Geekbench 6.2 which is patched to support 16KB page sizes (the off-the-rack release won’t work with 16KB page size, and needs you to reboot and use a 4KB page size).

For Geekbench benchmarks, bigger numbers are better!

Single core

Single-core scores measure the processing power of one CPU core and are more relevant for applications that are lightly threaded, meaning they rely mostly on a single core to process instructions.

Single CoreScore
Raspberry Pi 5 (16KB page size)774
Raspberry Pi 5 (4KB page size)764
Raspberry Pi 4340
Single-core scores for Geekbench 6.2

Over one hundred runs of Geekbench 6 we saw an average score of 764±6 for Raspberry Pi 5 using a 4KB page size, and 774±6 using a 16KB page size. That’s a ×2.4 speed increase over Raspberry Pi 4.

Multi core

Alternatively, multi-core scores attempt to measure the more realistic case where jobs are being distributed across all of the cores of the board’s CPU. These sorts of scores are more relevant for heavily threaded applications such as web browsers, which generally try to offload individual tabs into separate process threads.

Multi CoreScore
Raspberry Pi 5 (16KB page size)1588
Raspberry Pi 5 (4KB page size)1604
Raspberry Pi 4723
Multi-core scores for Geekbench 6.2

Over one hundred runs of Geekbench 6 we saw an average score of 1,604±22 for Raspberry Pi 5 using a 4KB page size, and 1588±63 using a 16KB page size. That’s a ×2.2 speed increase over Raspberry Pi 4.

Overclocking

So then: we don’t advise people to overclock their Raspberry Pi, and we aren’t necessarily going to tell you how to do it, or give you any guarantees if you do. On the other hand, there might be people who have written a blog post about doing it. Thanks Jeff.

2.4GHz3.0GHz
Single-core score774906
Multi-core score16041634
Overclocking the CPU to 3.0GHz

Overclocking the CPU from 2.4GHz to 3.0GHz, and the GPU from 800MGHz to 1GHz, we see a corresponding increase in performance with a ×1.2 increase in score on the single-core tests. Interestingly, we don’t see a similar increase in performance in the multi-core tests. This discrepancy is probably down to constraints on memory bandwidth, but it’s hard to tell for sure.

Other people, other benchmarks

So that was my benchmark, but you shouldn’t just take my word for it. You should definitely go look at what other people have done around benchmarking and performance testing on our new hardware.

Because as you’d expect, after perhaps looking at thermals, the first thing most people did when they got their hands on a preview Raspberry Pi 5 was to run some benchmarks, which gives us lots of numbers and graphs.

As always, Jeff Geerling has performance numbers for everything from Ethernet and wireless throughput, to memory throughput, to cryptographic performance — where Raspberry Pi 5 is 45 times faster than Raspberry Pi 4 thanks to the BCM2712 processor bringing Arm’s Cryptographic Extension to the Raspberry Pi.

Cryptographic benchmarks for Raspberry Pi 4 and 5 by Jeff Geerling

Then, the folks over at Core Electronics have thrown a whole mess of benchmarks at the hardware.

BenchmarkUnitsRaspberry Pi 4Raspberry Pi 5Performance Increase
Sysbench Single-ThreadedMBps6991041×1.49
Sysbench Multi-ThreadedMBps27944165×1.49
Stress-ng Single-ThreadedBogo ops/s104.78182.68×1.74
Stress-ng Multi-ThreadedBogo ops/s413.12737.21×1.78
Bzip Single-Threadedseconds44.9820.53×2.19
Bzip Multi-Threadedseconds28.5914.36×1.99
Gimp Resizeseconds67.0129.95×2.24
Gimp Rotateseconds77.2432.77×2.36
Gimp Auto-levelsseconds80.5234.64×2.32
Gimp Unsharp-Maskseconds115.1649.71×2.32
Speedometer 2.1score20.562.5×3.05
Glmark2score97202×2.08
Openarena TimedemoFPS8.7727.05×3.08
RAMspeed WriteMBps439129355×6.69
RAMspeed ReadMBps590227931×4.73
HDparm ReadMBps43.8190.05×2.06
dd WriteMBps34.4961.23×1.78
Iozone 4K RAND WriteMBps9.3815.22×1.62
Iozone 4K RAND ReadMBps4.714.6×0.98
Boot timeseconds33.419.1×1.74
Benchmarks results showing the performance improvements from Core Electronics.

But the benchmarks I found the most interesting are the ones are from Seeed Studio, who have gone out and benchmarked Raspberry Pi 5 using the ncnn framework.

The ncnn framework is a deep-learning inference framework that supports various neural network models — such as PyTorch and TensorFlow — and a range of hardware. Designed with mobile deployment in mind, it offers GPU acceleration via the Vulkan API, and it’s been really interesting to see the sort of performance increases we’re seeing between Raspberry Pi 4 and 5.

Some ncnn framework benchmarks from Seeed Studio

Remember though, unlike my own Geekbench benchmarks, here smaller numbers are better, and they were getting some really nice numbers from Raspberry Pi 5. Testing the YOLOv8n model on Raspberry Pi 5, with 640×640 pixel video stream, gave an inference rate of around 12fps.

Wrapping up

When we launched Raspberry Pi 4 back in 2019, with its quad-core Arm Cortex-A72 processor clocked at 1.5GHz, it was around forty times faster than the original Raspberry Pi model from 2012. But with Raspberry Pi 5, with its quad-core Arm Cortex-A76 processor clocked at 2.4GHz, we now have between two and three times the CPU and GPU performance again; roughly twice the memory and I/O bandwidth; and for the first time, we have Raspberry Pi silicon on a flagship Raspberry Pi device.

We’re pretty proud of the performance improvements we’ve been able to make. But best of all, though, is seeing folks out in the wild finding out exactly how far we’ve come in the last few years. If you end up running any benchmarking yourself, drop a link in the comments below: we love watching what you’re doing!

36 comments
Jump to the comment form

xeny avatar

Nice numbers.

The Geekbench scores – are they with bare board, case+fan or heatsink+fan?

Reply to xeny

Alasdair Allan avatar

Those were done with a Raspberry Pi 5 and Active Cooler.

Reply to Alasdair Allan

George Kokotis avatar

Could you please add Mathematica’s benchmark (WolframBench) for Pi5 vs. Pi4?

Reply to George Kokotis

Andy avatar

Hi, are there any benchmarks for gpio access speeds on the Pi 5?

Reply to Andy

PhilE avatar

Not that I’m aware of – what test would you like to see? It may not be much faster than a Pi 4 since GPIOs are on the other side of a PCIe link (high bandwidth, but also increased latency), but once the necessary software has been completed there’s the possibility of making use of the single PIO instance (as seen on RP2040).

Reply to PhilE

A Stevens avatar

So, broadly speaking, we’re looking at just about a 100x performance improvement, in just over a decade. Which is pretty fabulous! You average normal PC certainly hasn’t matched that, I would think? Obviously, core counts keep going up, but the sort spectacular generational performance leaps of old don’t seem to happen any more. It seems the little RPi still has plenty of future potential, but I can’t wait to see if the Pi 5 really works as a daily driver for a ‘normal’ person. The Pi 4 couldn’t quite, but it looks like this one might :)

Reply to A Stevens

Liz Upton avatar

We had one calculation kicking around where a Pi 5 was actually working about 150x faster than a 2012 Pi. I’ll see if I can dig it out.

Reply to Liz Upton

Liz Upton avatar

…and like magic, an email appeared just now in my inbox from Tim Rowledge, one of the Smalltalk small gods. He says:

I’m fairly sure your 150X comment refers to my benchmarking Squeak Smalltalk.
Comparing results over the years shows several tests at 150+X, several at 80X, and general performance to be a solid 70% of my 3.9Ghz i7 iMac or ubuntu-box. Which is amazing for $80.

Thanks Tim!

Reply to Liz Upton

xeny avatar

The faster display stack with Wayland moves a 4 significantly nearer to the “good enough” threshold. If I weren’t so fussy about JS heavy page load times in a browser it would comfortably there. I dread to think about comparing either price or performance to the quad CPU SPARCStation 20 I used about 25 years ago.

Reply to xeny

George Kokotis avatar

You brought back memories 😂
How about the SGI Indigo of ’94 that costed the equivalent of a nice car? Or the pretty penny of a decent CAD-grade GPU of the 80s, when a formal attire sales rep would come to discuss the purchase – all that before the teens game era?

Reply to George Kokotis

Claas avatar

Can’t wait to get my hands on and see what’s the actual limit.

Reply to Claas

Paul Webster avatar

In your Geekbench test results the page size didn’t seem to make much difference and in the 2nd set 16k performed worse than 4k.
What workloads would see 16k significantly outperforming 4k?
Plus – would having things set to 16k make it problematic taking SD card from RPi5 and running it on older device or is it all handled at boot time by config.txt or similar?

Reply to Paul Webster

PhilE avatar

It speeds up anything that switches RAM pages frequently – from memory, JetStream 2 saw a 7% boost, while a random memory access benchmark got a 14% uplift.

Reply to PhilE

Pete avatar

You say, “… we saw an average score of 764±6 for Raspberry Pi 5 using a 4KB page size, and 774±6 using a 16KB page size.”

That’s a measly 1.3% increase for the cost of binary incompatibility.

Reply to Pete

PhilE avatar

Fortunately it only takes a one-line, one-time-per-image config change to revert to 4K pages.

Reply to PhilE

rclark avatar

Thanks for benchmarks. Can’t wait to get hands on one of these to kick its wheels… Pre-ordered a 8GB version with 5V-5A plug and the fan/heat sink — un-patiently waiting, waiting… A test USB 3.0 2TB SSD drive has arrived today at the house just for this SBC’s boot drive. Don’t plan on using the SD card interface for anything. Also have a portable 4TB HDD drive available for testing to see if RPI-5s USB power can take it without fainting like the Pi-4 does :) . Let alone the performance difference between 3 and 4….

Assume boot times is measured booting off of a SD card, not a USB 3.0 SSD or the PCI-e SSD .

Reply to rclark

rclark avatar

Errr between RPI 4 and RPI 5…. Can’t edit response!

Reply to rclark

Leon Matthews avatar

I’ve got a venerable MacBook Pro 13″ out in my workshop. I’m using it to write this comment. Its Geekench score is only 308 single-core and 406 multi-core!
I had tried to replace it with a Pi4 and an old 1080P monitor on a swing arm, but the Pi couldn’t cope with my 3D printer’s Octoprint web-interface with its MJPEG webcam stream. It looks like the Pi5 could manage it handily!

Reply to Leon Matthews

Mark Tomlin avatar

Hum, 16k page sizes. Running Raspberry Pi OS on a M1 / M2 / M3 Macbook just got a little easier.

Reply to Mark Tomlin

Mahendra avatar

Long time user of Pi as the main and mostly only computer. The main reasons being relatively safe (no viruses since I stopped using windows), energy efficient (6 watts), lightweight (no bloatware) and compact (looks elegant in unofficial aluminium case) and low cost (no worries of warranties and repair costs).

It would be mostly fun and useful if raspberry pi high- end offerings are improved in the same areas as mentioned above such as, using smaller process node, a separate security monitoring and permissions software, software to run in ram-disk entirely.

One small and useful innovation could be cables with physical on -off button so that one need not disconnect the cables each time one has to stop using a certain peripheral. Think using two pis connected with a single monitor. The use cases are numerous.

Reply to Mahendra

rclark avatar

Adafruit has cables for this use. For example, the USB Power Only Cable with Switch – A/MicroB (Product ID: 2379)
I own several.

Reply to rclark

Mahendra avatar

Thanks. Add some more features like not only power but data too, hdmi interface and you have a hassle free multi-pi setup. And even leave the cable part, only an adaptor with a switch would be a sufficient and perfect solution – one that could utilise any cable laying around.

Reply to Mahendra

Mahendra avatar

Forgot that pi 5 has a physical on-off/sleep button. If it could be configured so that whenever it goes to sleep it instantly cuts power to the hdmi, then it would become a physical hdmi on-off button. Adding hardware features always multiplies use-cases.

Reply to Mahendra

Mahendra avatar

Or isn’t sleep means power-off to the monitor. So no need of even external configuration. At last, problem solved.

Reply to Mahendra

Dan P avatar

I’d be interested in the USB3 speeds if any one has any comments. I’m getting into astrophotography and people have commented that the Pi 4 USB 3 interface can’t cope with the high throughput from the Astro camera. I’m guessing/hoping that the new addition IO chip should help with that and the throughput should be much better?

Reply to Dan P

Alasdair Allan avatar

USB speeds are significantly faster. The Pi 5 has dedicated 5Gbps available for both the ports, courtesy of our RP1 chip. That means its ×2 faster, and as each channel is independent you’ll get an uplift above that if you’re using them in parallel.

Reply to Alasdair Allan

Dan P avatar

Excellent! Thanks for the update. It looks like the Pi 5 will be timed perfectly for my new hobby!

Reply to Dan P

Mike Redrobe avatar

For completeness, (and laughs) you should run geekbench on all current models:
Pi zero w and zero 2w

Reply to Mike Redrobe

Mark avatar

I am currently using an 11 year old laptop (i5 3210M) which I’ve just found out is slower than the Pi5 – from a quickly scurry on the net.
Seriously contemplating whether I should bite the bullet and go this.

Reply to Mark

Ashley Whittaker avatar

I am FAR from a neutral jury member but… DO IT

Reply to Ashley Whittaker

Mark avatar

Mmmmmm, thanks for the input, I’m 95% convinced this could work, and there’s only one way to REALLY find out isn’t there….

Reply to Mark

fanoush avatar

I think the speed vs price is not the real reason here. You can possibly find cheaper and faster refurbished Intel or AMD hardware (even passive with no fan and low power consumption) but it won’t be the Raspberry Pi so if you want that, just go for it.

Before Pi5 launched I got (refurbished but good as new) Wyse 5070 thin client for 70 EUR including ram, storage, power supply. It is relatively small, silent, in nice case, with upgradable RAM and SSD and is faster than Pi4 (https://browser.geekbench.com/v5/cpu/21851677) However after Pi5 launched I see it is actually slower than that! Now I see dell Optiplex 300 thin client (https://browser.geekbench.com/v5/cpu/16172037) available for about 120 EUR on ebay and other places so that may be similar price to Pi5 with power, case, storage etc. And these are both passive with low power draw. If you don’t mind the noise there are much better deals for used desktops/laptops. So it depends. I’d go with the Pi for the fun of experimenting or the community or the ARM CPU, or the size of bare board (if you don’t plan to attach extra stuff like SSDs), not because of Pi being cheaper/faster hardware.

Reply to fanoush

rclark avatar

Plus the RPI-5 isn’t refurbished/used ebay trash (yeah, I know, one man’s trash is another ones treasure) . Now brand new RPI-5, ready to rock, ready for fun, ready for some GPIO electronic projects! How does a Wyse thin-client stack up against that? :D

Of course, one has to get one in hand…. Not a whisper from my pre-order store that anything is even shipping yet …. All silent on the Eastern front.

Reply to rclark

fanoush avatar

I was answering to someone who wanted to replace 11 years old laptop by Pi5 so GPIO will probably sit there unused.

As for getting GPIO access you can always pair any computer with Raspberry Pico (or even Zero) over usb or wireless. Also I’d say that most “GPIO electronic projects” are best served with something else than Pi5 – either older Pi model, Zero, Pico or even ESP32 or other micros. Fast CPU is often not needed and not everyone can afford killing $60 thing by wire plugged to wrong pin.

Reply to fanoush

Steven Friedrich avatar

I have been doing some testing to see how much performance I need to replace my WD MyCloudEX2 Ultra which utilizes a ARMADA™ 385 processor. My Pi4B appears to be faster. I’m using an external USB 3.0 HDD and am surprised the Pi4B doesn’t stutter streaming 4K movies over gigabit Ethernet.
Has anyone benchmarked the ARMADA™ 385 against the Raspberry family?

Reply to Steven Friedrich

Uygar Yilmaz avatar

“a ×1.2 increase in score on the single-core tests” means the performance metrics reached “2.2 times” the original values. That’s clearly not the case here. The increase is “0.2 times.”

Reply to Uygar Yilmaz

Leave a Comment