Single Board Computers Benchmarks(github.com)
github.com
Single Board Computers Benchmarks
https://github.com/ThomasKaiser/sbc-bench/blob/master/Results.md
50 comments
Definitely a useful table, but two additional dimensions I would really love to see here are cost, and power consumption per unit work, especially the latter.
And year introduced, so we cant look at trend.
Personay it's been crushing to me that Cortex A53 (2012) and A7 (2011, a slightly updated A5, 2009) have been around for so so so so long & still dont have a worthwhile replacement. A55 is out but debateably a 10% win at best, and the almost 2 year old A510 doesnt actually exist/isnt for sale. A7 has some A35s that maybe sort of compete.
I want to see what is changing over time, if anything. I never thought I'd be anywhere as thamkful as I am that RPi has seemingly singlehandedly forced some value delivery, some progress. I dont know that we'd be seeing things like A72's (2015) for under $100 otherwise. The low end market has been rough.
Personay it's been crushing to me that Cortex A53 (2012) and A7 (2011, a slightly updated A5, 2009) have been around for so so so so long & still dont have a worthwhile replacement. A55 is out but debateably a 10% win at best, and the almost 2 year old A510 doesnt actually exist/isnt for sale. A7 has some A35s that maybe sort of compete.
I want to see what is changing over time, if anything. I never thought I'd be anywhere as thamkful as I am that RPi has seemingly singlehandedly forced some value delivery, some progress. I dont know that we'd be seeing things like A72's (2015) for under $100 otherwise. The low end market has been rough.
I think the problem with the in order designs is that there is not much to improve on. Most of the gains for OoO cored designs are coming from the branch predictor.
You mentioned A55 and A35. There are good boards shipping those cores, now. rk356x and various Amlogic SoCs come to mind.
You mentioned A55 and A35. There are good boards shipping those cores, now. rk356x and various Amlogic SoCs come to mind.
The A55's have been around for a bit, yeah. They just dont seem to be noticably better as cores (but there's been ongoing gains in uncore). You may be right that there's just not much headroom perhaps. But supposedly the A510 and announced A515 offer some real gains.
I havent tried the A35. I see them as more targetted towards MID devices than the A7 was butb hats product-market segmemtation maybe.
I havent tried the A35. I see them as more targetted towards MID devices than the A7 was butb hats product-market segmemtation maybe.
Is that all from one author ? If the results are submitted by multiple people I'd imagine it would be hard as not everyone have setup to measure power usage accurately.
For anything battery powered idle usage also is useful stat but that I'd imagine would vary widldly depending on whether kernel supports all of the chip's power saving features
For anything battery powered idle usage also is useful stat but that I'd imagine would vary widldly depending on whether kernel supports all of the chip's power saving features
> whether kernel supports all of the chip's power saving features
I mean that is genuinely relevant info for you if you're gonna be using that distro. It doesn't matter what it supports in hardware if there's no software support for it, it's like it doesn't exist.
I mean that is genuinely relevant info for you if you're gonna be using that distro. It doesn't matter what it supports in hardware if there's no software support for it, it's like it doesn't exist.
On top of that it might support it but need special settings, or have certain optimizations enabled to get the lowest power.
Also would be a lot more useful if you could sort by columns.
Visidata has no trouble importing the table:
vd 'https://github.com/ThomasKaiser/sbc-bench/blob/master/Results.md'Bruh. How have I've never heard of this tool before, this looks amazing!
VisiData link for the curious
https://www.visidata.org/
https://www.visidata.org/
> Visidata
Wow, thanks for this, never heard about it before, installs super easy, super useful.
Wow, thanks for this, never heard about it before, installs super easy, super useful.
I added some sorted table variants: https://github.com/ThomasKaiser/sbc-bench/blob/master/result...
That makes it a bit harder, most people don't have access to accurate power measurement device.
Kinda shame USB hosts don't have that functionality
Kinda shame USB hosts don't have that functionality
I’d also like to see some form of GPU compute and network bandwidth capabilities.
Yeah, notably the Radxa ROCK 5B who has outperformed the Apple M1 on performance/power.
Wow.
Wow.
How should RK3588 be able to outperform any recent Apple SoC?
I fail to see how the Apple M1 is a SBC. I suppose a Macbook may technically be a single board because it's all soldered together to prevent people from upgrading their RAM and that sort of thing but it's not like one can get it as an actual dev board.
Also interesting that the Pi 4 running in 32 bit armv7 mode outperforms itself in 64 bit mode by up to almost 2x.
Also interesting that the Pi 4 running in 32 bit armv7 mode outperforms itself in 64 bit mode by up to almost 2x.
> I fail to see how the Apple M1 is a SBC.
SBC is single board computer. Why it wouldn't be ?
> it's not like one can get it as an actual dev board.
There is nothing in the term that specifies it is supposed to be hackable dev board. There is also no findamental difference between "laptop built on single board that is connected to peripherals" and "a rPi4/CM4 connected to peripherals".
It was always used to signify "you don't need any extra IO/memory than the board to run the thing". I think term itself came from microcomputers.
SBC is single board computer. Why it wouldn't be ?
> it's not like one can get it as an actual dev board.
There is nothing in the term that specifies it is supposed to be hackable dev board. There is also no findamental difference between "laptop built on single board that is connected to peripherals" and "a rPi4/CM4 connected to peripherals".
It was always used to signify "you don't need any extra IO/memory than the board to run the thing". I think term itself came from microcomputers.
I think a likely reason for the Pi marks is that there’s lots of work to be done on the 64-bit OS. They still recommend the 32-bit version even though newer Pi’s have 64-bit CPU’s.
I researched this a bit when I configured a Raspberry Pi 400 as a server in my bedroom.
https://joeldare.com/private-analtyics-and-my-raspberry-pi-4...
I researched this a bit when I configured a Raspberry Pi 400 as a server in my bedroom.
https://joeldare.com/private-analtyics-and-my-raspberry-pi-4...
Yeah afaik they also had some issues with MMAL using some hardcoded address jumping that while it did work on 64 it would on occasion go outside the new virtual address range and crash, which made the Pi Camera and the GPU very unreliable. I suppose by now Broadcom's already fixed that given that there are actual 64 bit builds on the official Raspberry site again.
While there's a lot left to fix I'm sure that part of the slowdown is also just physics. Even if it's native, there's an extra layer of abstraction to process, you need to fetch twice the address bytes, and write twice the address bytes. Makes sense it's roughly a 2x slowdown since it's doing 2x the data shuffling (especially since it's in single threaded performance).
Or it's far more complicated than that and it's just a coincidence that the numbers align...
While there's a lot left to fix I'm sure that part of the slowdown is also just physics. Even if it's native, there's an extra layer of abstraction to process, you need to fetch twice the address bytes, and write twice the address bytes. Makes sense it's roughly a 2x slowdown since it's doing 2x the data shuffling (especially since it's in single threaded performance).
Or it's far more complicated than that and it's just a coincidence that the numbers align...
The Mac mini is closer to a dev board than a MacBook and it's as close to a dev board Apple will get other than the Apple 1. Arguably, a Mac mini is comparable with a NUC and also comparable with a Pi if all you do with the Pi is put it in a case and use it like a PC.
As soon as you desire GPIO, tough luck.
As soon as you desire GPIO, tough luck.
It also same as many laptops.
Would probably help to include price points, or at least links to product pages. Otherwise I don't necessarily know if I'm looking at a $200 intel thing or a Pi Zero clone.
Great to see the this SBC benchmark. I have done one benchmark for several available SBCs from an edge data stack viewpoint (Edge Data Stack Capabilities Ranking: https://joinbase.io/ranking/, but the article is still in progress...so sorry for no content now)
What I have tested : $20 RockPi S(minimum squared Armv64 dev board), $0 unused IPTV box (S905L2 chip)(should be bought from secondhand market in ~ $7), #30 NanoPi Neo3(the minority of cheap SBC with DDR4 on), the Allwinner D1(the first batch avaiable Risc-v 64 board) and my own Xeon server (and more cloud based instances).
A possibly surprising fact is that, if the modern software (such as my database here) can take full advantage of the modern hardware, then the performance of even a small 40mm squared SBC board (such as RockPi S) can be compared to the large-scale-used traditional softwares running on a Xeon Platinum 8260 based bare-metal server(such as PostgreSQL).
What I have tested : $20 RockPi S(minimum squared Armv64 dev board), $0 unused IPTV box (S905L2 chip)(should be bought from secondhand market in ~ $7), #30 NanoPi Neo3(the minority of cheap SBC with DDR4 on), the Allwinner D1(the first batch avaiable Risc-v 64 board) and my own Xeon server (and more cloud based instances).
A possibly surprising fact is that, if the modern software (such as my database here) can take full advantage of the modern hardware, then the performance of even a small 40mm squared SBC board (such as RockPi S) can be compared to the large-scale-used traditional softwares running on a Xeon Platinum 8260 based bare-metal server(such as PostgreSQL).
Mainline support is important for many use cases, performance only for some.
Agreed, mainline Linux support with no closed drivers (and, ideally, no closed component firmware blobs) is often important, along with an open source toolchain, if you need the device to be sustainable.
I hear you.
There's a page on the FSF site[0] where they give some pointers on the kind of support the various SoCs have, but it's very terse.
For the records, I have an Odroid C1 that's collecting dust on a drawer...
[0] https://www.fsf.org/resources/hw/single-board-computers
There's a page on the FSF site[0] where they give some pointers on the kind of support the various SoCs have, but it's very terse.
For the records, I have an Odroid C1 that's collecting dust on a drawer...
[0] https://www.fsf.org/resources/hw/single-board-computers
Great list. Thank you
Rockpi 5B is looking good!
Don't think the title is correct though. That's a mix of arm boards, some decided not in the SBC class. e.g. H270-T70 is more of an enterprise server.
Rockpi 5B is looking good!
Don't think the title is correct though. That's a mix of arm boards, some decided not in the SBC class. e.g. H270-T70 is more of an enterprise server.
sbc-bench started to overcome the limitations of common benchmark tools with SBC few years ago but obviously has evolved. And the results collection tries to collect everything ARM to get a perspective (like Apple's efficiency cores outperforming the big cores on Rock 5B for example).
And some operational modes (like testing individual core types / clusters also individually) make sense on 'generic PCs' too, see this Alder Lake i9-12900T thingy where efficiency and performance cores are measured individually: http://ix.io/4dzo
And some operational modes (like testing individual core types / clusters also individually) make sense on 'generic PCs' too, see this Alder Lake i9-12900T thingy where efficiency and performance cores are measured individually: http://ix.io/4dzo
I'd like to know how fast the networking is of such boards. As in: read a file and distribute it over ethernet or wireless, in MB/s.
What's the next best indicator? Clock speed? CPU?
What's the next best indicator? Clock speed? CPU?
None. You need to look at interface the NIC is connected to the CPU. Sometimes MAC is built into SOC and board just needs PHY, that can be easily line-speed.
Sometimes it's via PCIe (usually pcie to usb+ethernet) chip, then it's usually also line speed. Well, at least till you start using USB3 heavily
Sometimes you just get USB3 NIC or worse, USB2.
Wireless can also be funky, I've noticed using BT and Wifi at same time absolutely destroys wifi speeds on my rPi4
Hell, in theory even builtin MAC doesn't guarantee line speed, all depends on internal architecture.
Sometimes it's via PCIe (usually pcie to usb+ethernet) chip, then it's usually also line speed. Well, at least till you start using USB3 heavily
Sometimes you just get USB3 NIC or worse, USB2.
Wireless can also be funky, I've noticed using BT and Wifi at same time absolutely destroys wifi speeds on my rPi4
Hell, in theory even builtin MAC doesn't guarantee line speed, all depends on internal architecture.
Seeing the Macbook on there was a bit of a surprise, never really considered those to be in the same realm as RPis. But that list also has some other big hitters like the Honeycomb, so that's fair.
Is anybody aware of any SBCs capable of accepting a full-size PCI-E GPU? I've been wanting to throw together a server for some simple AI home assistant stuff (like running Stable Diffusion), but I haven't dug in to understand if ARM is capable of this yet or if I should just go buy some used parts on eBay.
(I have an old 1080 TI which is sufficient for everything I need, but I only have a dual Xeon server from 2011 that sucks watts.)
(I have an old 1080 TI which is sufficient for everything I need, but I only have a dual Xeon server from 2011 that sucks watts.)
Why not build a cheap 4 core ryzen setup and not deal with all the pain... I know socket am4 is on the way out... But you could do it and have a system that works......
Used ebay parts. Intel generally have lower idle usage. Also remember that total TDP doesn't really indicate idle usage so you can often get away with getting stronger part and not pay that much in power.
Also on side note it's insane how much less newer servers use when idle. Few of our never Xeon servers use below 50W (single CPU + 6 NVMe drives), while comparable dual SSD server from 6-7 years used good 100W.
AMD on the other side uses like double.
Also on side note it's insane how much less newer servers use when idle. Few of our never Xeon servers use below 50W (single CPU + 6 NVMe drives), while comparable dual SSD server from 6-7 years used good 100W.
AMD on the other side uses like double.
A card like a 1080 TI needs a PCIe x16 slot. There's no SBC with an x16 slot. Though there's plenty of Mini-ITX motherboards, which is pretty small at 6.7-in x 6.7-in (17.0 cm x 17.0 cm).
Logically you can use even a 1x slot for loads that don't have a lot of traffic on the bus without much impact. There are adapters for the physical aspect of it.
Jetson TX/TK/Nano/Xavier. The older ones should be quite affordable now
The TX and TK are garbage at this point. The Nano is (IIRC) going to be similar to a 1050Ti. Unfortunately, prices for the Nano are the same kind of crazy as GPU prices last year: All the examples I could find on eBay were 2-4x original retail price. It’s an okay board, but frustrating to the point that I gave up and built a cheap AMD box using mostly spare parts.
> but frustrating to the point that I gave up
Same experience here, with a 2G Nano sitting in my drawer because developing on it is like pulling teeth compared to a Pi 4, despite all the extra horsepower. Nothing is standard on that thing. They couldn't even keep the GPIO header as-is, they had to reshuffle the pin locations while keeping it misleadingly in the same form factor.
Same experience here, with a 2G Nano sitting in my drawer because developing on it is like pulling teeth compared to a Pi 4, despite all the extra horsepower. Nothing is standard on that thing. They couldn't even keep the GPIO header as-is, they had to reshuffle the pin locations while keeping it misleadingly in the same form factor.
This is super cool, I'd love a purchasability index or benchmark for these too
> a purchasability index
All get a zero these days.
All get a zero these days.
I wish there were more SBCs with PoE support.
It's kinda... expensive to add. Proper POE is 48V so need some bigger pin spacing and you kinda want to have galvanic isolation too and that's some extra parts that are not small.
Just look how big even the smallest rPi POE shields are (and I think some of them don't do the power isolation.
Also, realistically if you're also not booting from network you want some kind of UPS to not get dirty shutdown every time. I've used UPS Pico which has option for supercap but feature-wise it is a bit overkill and also doesn't do isolation nor power negotiation, just plain 24V passive PoE.
Just look how big even the smallest rPi POE shields are (and I think some of them don't do the power isolation.
Also, realistically if you're also not booting from network you want some kind of UPS to not get dirty shutdown every time. I've used UPS Pico which has option for supercap but feature-wise it is a bit overkill and also doesn't do isolation nor power negotiation, just plain 24V passive PoE.
The interesting metric for me would be usability with mainline kernel.
Easy with ARM: Once the hardware is obsolete upstreaming work is close to be finished.
TIL about ix.io, that is awesome