The ultimate Pi 5 arrives carrying 16GB ... and a price to match The Raspberry Pi has come a long way from its early days, as demonstrated by the single-board computer maker's latest iteration of the Pi 5 in 16GB guise. A 16 GB Raspberry Pi 5 board Is this the ultimate Pi 5? The company gave us one to review ahead of today's launch; and frankly, it's a bit of a headscratcher. The extra …
I don't think it's publicly known how much RAM ChatGPT would require to run, but I've seen estimates starting from 45GB, up to around 80GB. OpenLLaMA models can apparently work within 16GB, although I'm not sure that's total system or required free RAM[1]. The Pi also does not have a GPU and the processor is rather limited, so it's not ideal for that application in many ways. (There is a "machine learning" module, but it uses a bespoke architecture and is mostly only suited to 2D vision tasks.)
That said I've been wondering for a while about trying a Pi for some medical imaging research applications, since they are a lot cheaper and less power hungry than what we normally use, even if it might be slower. Some of that is deep learning and some traditional algorithms, both are RAM hungry but not usually to those 80GB extremes, 8GB probably wasn't enough for some of it, but 16GB is a safe-ish bet. (Although less for our own use than for education and outreach and maybe making this kind of technology more available to researchers with less resources.)
[1] I had some fun recently getting a 3D-UNet type model working on a 8GB laptop GPU, would just about fit, but required running with desktop stopped to ensure as much free GPU RAM as possible.
From where I work right now we looked at the Pi for similar applications to you (albeit not medical), and the devs opted to use the nvidia Jetson Nano thing. That is far more ideal to LLMs than the Pi, albeit 3x the price (comparing it to the 16GB RAM one).
Personally I've ran some of the OpenLLaMa things on serveral machines. Processor is critical really, but they'd run within 16GB RAM.
Thanks, the Jetson looks interesting and probably more practical for serious use. Might be an good next step if we find a real application, I'm looking at the Pi as a proof of concept for SoCing this type of thing (and the GPU is only useful for some of these loads, the single threaded benchmarks for ARM Cortex-A78AE 8 Core from Orin NX and BCM2712 from Pi 5 don't look too far apart, although twice the cores isn't to be sniffed at).
"The Pi also does not have a GPU and the processor is rather limited, so it's not ideal for that application in many ways."
I have to correct you there. The Pi does have a GPU and you can use it through a few interfaces. OpenGL and Vulkan are available, and some of the boards support OpenCL 1.2 at least. It's true that most normal acceleration systems will not support this, but that is a very different thing than the board not having a GPU at all. I wouldn't recommend it for running an LLM, but you have the ability to do so if you don't mind slow output.
The main difference from a typical GPU is that the VPU runs the master proprietary OS that actually controls the hardware, which means the Raspberry Pi does not respect your freedom (but what did you expect from Broadcom factory e-waste).
At least adding support in the LibreRPi project is possible as by default RPi doesn't activate the digital handcuffs unless you ask for them.
You might be right that demand for the 16 GB model will be lower than demand for the lower-spec versions. That would just mean that they make fewer production runs of that version. It wouldn't really change what price they can charge for either of them or whether they have enough buyers to make the 16 GB version profitably. I'm guessing that there are enough people who want to use the Pi 5 as a desktop or have a memory-hungry process they already run on Pi 5s who will buy a 16 GB version, so they won't be losing the money for this production run.
No-one needs to run ChatGPT.
You'd want 16G otherwise for VMs. If there are no VMs, then 4G is fine. My Pi4 with 2G RAM can run Writer, Thunderbird and Firefox at the same time, if not too many tabs on Firefox. A 4G would be far better, but the 2G was a bargain.
There are maybe better solutions a little more costly than a Pi 5 with 16G RAM, if you really need 16G RAM.
Did I mention that no-one needs ChatGPT?
> Who wants to wait multiple seconds or even minutes to get a response?
Jeff Geerling has tested the 16GB RasPi with an LLM, and found that it generates about one word per second...
The extra memory is always welcome, however, the souped up board is strides apart from the computer's humble origins as something that buyers could pick up for around $30.
I don't know why people are struggling with this, you don't have to buy the 16GB one. There is a range of options now. If you want humble origins then there is a 2GB version, there is a Zero 2W which is far cheaper than $30 and far more capable than the humble origins, plus the single core Zero. And the 4 is still for sale and I can still see the Pi 3 on sale today at the approved seller. In real terms, these are cheaper than they were and cheaper than the original Pi.
So the original Pi idea is still there, but a choice is available too.
I don't think the D0 stepping CPU on this 16GB was on the original Pi 5 4GB and 8GB at release. It arrived later with the 2GB version.
Agree about the misdirection.
Personally, I see little reason why the more performant Pi variants don’t have a DDR4 SODIMM socket, this would enable reuse of modules from old laptops (and there are going to be a lot of these around when the world moves to W11). Then if you really wanted you could run with 64GB of RAM…
Trouble is the Pi uses LD DDR4, which has to be soldered to the motherboard. I assume as the 16GB is a Pi specific package of 4*4GB chips, the price increase isn’t too bad. However, if you accept that large memory Pi’s won’t be used in very low power applications, then it should be possible to replace that custom soldered on LD DDR4 module with a slightly more power hungry SODIMM socket and conventional mass market DDR4 SODIMM.
Yes, the price has risen significantly *for the top of the range* but you can still by a Pi 1 A+ for under GBP16 (around USD20) excluding local taxes, and a Pi 3 for not much more.
I can't see many people wanting the 16GB, I'm happy with the 8GB for my dev system, but I'm sure there'll be a market for them.
I bought a few of the earlier Pi - they were nice enough, pretty ok for my use cases. One of them does the 3d printer stuff, one is doing pihole, and some sit in a box and are not used. I was amazed how much the price has increased over the years. Admittedly, they now offer much more power, RAM, whatnot, but it is - for me - ultimately a hobby board, and I would like it to be priced accordingly (i.e. such that I can just get one without thinking too much for a project that will ultimately fail). That's why I now do many things with arduinos, cheap as chips ;) - and the new ones have WiFi and all sorts of stuff, and with things like Tasmota setting up sensors is now really easy.
Now that I think of it: I actually prefer my attiny stuff (but that's a different use case, admittedly).
Cheaper, older models are still available and I've got a couple of projects earmarked for some of my older units. For me, the Pi4 is the sweet spot with much higher bandwidth for network and file access. The RPi5 provides more oomph but requires a cooler and a bigger power supply as well. I can see people using these in some situations but you are getting towards the kind of environment where blades might make more sense. But they wouldn't be offering them if there wasn't a demand.
I attached an M.2 to a Pi5 8gig and tried to run it as a desktop. No getting away from it, it needs more oomph especially to watch a decent ripped DVD. (even with recompiling VLC) and at this price they are now crossing into cheap laptop territory. So I won't be making the jump.
I too will not be using an R'Pi as a desktop - instead, I'll keep on using my desktop for that. Because I am very well aware that what is required from "a desktop" varies wildly from person to person (e.g. I chew up loads of cores & RAM at the desk but stick to watching videos at 720p or smaller, off in the corner - go downstairs to the telly box for bigger).
Meantime, I'm partway rigging up a set of recently acquired Pi's (3B+'s as I decided that spec best fits this use case) to run as specific servers, alongside the other varied Pi's already on the LAN. I've also just bought an RP2040 Pico-alike without WiFi but with a W5500-driven Ethernet jack onboard - it'll fit a niche but I wouldn't expect everyone else to leap out and buy one (you pay a premium for wired these days).
In other words, horses for courses - that you won't buy any particular model because it doesn't fit your particular use case isn't really worth saying.
Depends how that ripped DVD is being decoded. The Pi has hardware decompression capabilities, so will depend if VLC is using the hardware or just the main CPU. I use my Pi 4 as a media centre with Kodi and it can happily handle 1080p H265 videos without breaking sweat. Of course, Kodi is much more lightweight than a full Linux distro...
I think this looks more like a display driver issue or something else going on.
I've being playing DVD ISOs since the Pi3 era and Pi4 was happy doing full 1080 blu-ray ISOs.
I'm happily running H265 4K video files on my Pi5.All with plenty of CPU headroom left.
I'm likewise using a LibreELEC Kodi to be TV friendly but should still be capable on a RaspberryPi OS.
Having just ripped a couple of DVDs (MP4) over the holiday and watched them both on the RPi-4 Kodi and my MBP 2016, I can confirm the RPi-4 is fine, better than the old Mac, which did the ripping and encoding. It even handles x265 well, but you must give ≥ 4 bigger power supplies.
As usual, it depends. About a month ago I posted (ElReg) a Pi5 set-up for on-line use by retirees. Subsequent work has shown that a 2GB Pi5 with a Flirc passive case and a ”High Endurance" microSD card was satisfactory for normal Chromium browser use (<12 open tabs). Playing a nominal 1080 YouTube video (with Bluetooth speaker) was OK.
The 2GB Pi5 runs consistently 4-5C cooler than the 4 and 8GB versions. We could have saved AU$9 by using the official RP plastic case; but aesthetics, robustness, and not needing a fan, favoured the Flirc. The fan in the standard RP case was usually running, even on idle, with the 4GB Pi5. With the 2GB Pi5 the fan was normally off, but came on when the browser was opened (summer room temp of 25-26C). It seems that the 16GB version has the same "more efficient" chip revision as the 2GB Pi5, so it may run cooler for those who need it.
We had to get an assistant to approve a bunch of bananas that we'd just weighed on a recent visit to Morrisons.
When the assistant came over the terminal flashed up black and white images of the bananas so I'm guessing they do some image processing to check that the thing you just weighed is what you claim it is.
Of course bored assistant just tapped through the confirm procedure at lightning speed without particularly seeming to confirm that they were indeed ordinary bananas and not organically hand-reared avocados so you may still get away with your cunning plan...
Yes, because all of the companies who now use RPi's as the basis for their own products will be only too happy to embark on a significant mechanical redesign in order to accommodate a thinkpad (or any other alternative system you might care to mention), as well as an equally significant code rebuild/write in order to get whatever they're running on the RPi to now run instead on a completely different architecture...
Then you've got the myriad of hobbyists similarly quite happy with the RPi form factor and ecosystem, who might not have as pressing a need to not change away from that, but who'd still very much prefer to continue using something they already know and understand and are perfectly happy with *except* for it not having quite as much RAM as they now find thsemselves desiring.
You seem to have fallen into the same trap as the article author themselves, of not being able to comprehend that the RPi is no longer *just* a cheap and cheerful board for hobbyists to play around with, but is *also* now an increasingly important component in both higher performance personal setups (in some cases replacing those older desktops/laptops you're now suggesting could be used, in scenarios where they used to get given a second lease of life because there weren't any other viable platform options available), as well as professional products being developed the world over.
So yes, there's probably zero business case for a 16GB Pi5 if all you care about is churning out cheap and cheerful playthings for hobbyists, but given how few mis-steps RPi as a company have taken over the years, I think anyone questioning the relevance of this particular model is doing them a serious dis-service, because to do so is to rather strongly imply that they haven't done their market research correctly and are just chucking out this update because they can, not because they already know people want it.
Would a Pi be any good to use for routine simple tasks as a desktop replacement? I would only want to use it for browsing the internet (with an adblocker), POP3 email, LibreOffice? Zoom?
Since dumping Windows and switching to Linux Mint more than a decade ago I've used relatively cheap, small, fanless, solid state computers from China via Amazon. They've worked great, but I'm getting more suspicious of such items from China as I don't know what nasties may be lurking in the firmware etc nowadays.
Have a look at the Argon ONE v3 case with M.2 support (link below). A Pi5 in one of those, with a cheap M.2 SSD, would make a great desktop machine for Libre Office and web browsing, I think. Removing the SD card improves the performance noticeably.
https://thepihut.com/products/argon-one-v3-m-2-nvme-raspberry-pi-5-case
GJC
I have an 8GB Pi in an Argon 5 Neo case with a Crucial 500GB M.2. It was a bit fiddly to put together (old fingers and eyes); but works well. Against that, I found that if you only need web browsing, light office work, and nominal storage, a 4GB Pi with an "Official" 64GB microSD is not *that much* slower - YMMV.
That's what I've got connected to the big TV's in the house. I can sit down in any room, pick up a wireless mini keyboard and do just about anything I can do on my powerful Linux Mint laptop, albeit slightly more slowly. All the data is on a NAS which is another Pi 5 with a larger SSD. I'm using 8GB ones are the moment which seems to be sufficient for desktop type use, where as my the 4GB Pi 4B I bought just before the 8GB ones came out struggled with lots of tabs open.
I'm pondering whether to get a 16GB Pi 5 to replace an 8GB Pi 4 which acts as my data collector, gateway and webserver, or wait for the 16GB compute module and put together a system based on that.
Chinese firmware dodginess should be the least of your concerns these days. And frankly, if you're concerned about that there's nothing that anyone can offer that will satisfy your healthy suspicion.
You're probably safest buying used hardware off eBay, piggybacking off a neighbour's WiFi and running through a VPN. That way at least you know that the eBay seller, your neighbour and the VPN provider all have the possibility of snooping on your activities.
(Seriously though, I'd be more concerned about US snoops than Chinese ones)
Can't say I trust the US or the Chinese nowadays. China in particular has been caught so many times lately hacking into systems around the world, particularly in the US, they have shot themselves in the feet from a credibility and integrity perspective.
I don't like being spied on any more than the next man, but given a choice I'd rather be spies on by the guys on the inside pissing out than the guys on the outside pissing in. The Chinese are on the outside pissing in because they want to piss all over you, while the US just wants to make sure we aren't going to wake up one day to find you pissing in our boots. That would really piss us off.
@Andy Non:
If you didn't design, verify, and fabricate your CPU and support chips, you don't know what's in it.
If you didn't design, verify, and fabricate your GPU, you don't know what's in it.
If you didn't design, verify, and fabricate your IME, PSP, or equivalent, you don't know what's in it.
If you didn't code or completely review your toolchain and IME/PSP/etc. code and assemble/compile and link it, you don't know what's in it.
If you didn't write or completely review, and assemble/compile your BIOS, you don't know what's in it.
If you didn't code or completely review, assemble/compile and link your OS, you don't know what's in it.
It all comes down to who/what you trust. Microsoft, Intel, and AMD all have potentially-spy-assisting features in their products.
Some techies have given me the hairy eyeball because for decades I've used some Chinese rando's ("chenall"'s) GRUB4DOS boot loader, some Russian rando's TUI partitioning program - Ranish Partition Manager, and the Taiwanese National Supercomputing Lab's Clonezilla.
I've had my personal details leaked to bad guys four times so far -- all due to corporate and government cock-ups.
Most of that should work. The only thing I'd be concerned about is Zoom. The Pi is capable of all the other tasks as a low-end desktop, and if you use an NVME drive, load times are significantly reduced.
Zoom, however, is making me nervous. It doesn't look like they have native support for ARM Linux for their native app (using this system requirements page). That means you'll have to use Zoom from a browser. Browsers and graphics have always been a weak spot for the Pi. For a long time, Chromium was the only browser that could use hardware video anything on the Pi. Theoretically, Firefox can do it now. That wasn't necessarily a problem when playing streaming low-resolution video, but it will likely be a much bigger problem when handling multiple input streams, an output stream if you're on camera, and screen sharing. If you need Zoom for lots of video, this would be worth testing. Maybe you can find someone with a Pi 5 already who can report on this.
Of your tasks, web browsing is now probably the most demanding: one process per tab; multiple threads per tab, etc. Video is all about the availability of HW codecs and this is likely be down to the browser, but the more recent RPis have pretty decent HW codecs so worth checking.
Thanks, I just had a look on YouTube and there is a good demo of someone using a Pi to run Ubuntu quite successfully. Granted he used an overclocked version, which I'm a little uneasy about, and he still wasn't able to play video's back at higher resolution / frame rate. But for everything else it seems fine. I wonder how well it would cope with a screen full of Zoom participants?
Quote: 'and he still wasn't able to play video's back at higher resolution / frame rate.'
Sounds like a driver issue perhaps?
I'm using a Pi 5 as a media player (LibreELEC+Kodi, no plugins/extensions) plugged direct into my TV (old model LG 4k), and it plays back 4k 60fps files (from a NAS) just fine for me.
PS: The downvote wasn't from me.
...it's starting to get close to the little N100 boxes that are basically everywhere these days. I guess if you particularly needed arm rather than x86 for some workload it's nice to have that option, or maybe if you have a lot of PI compatible HATs or similar that you want to reuse, but I can't see that for most people the pi would be a better way to go. Once you're up in the £150 area (which appears to be what one would cost from pihut including a case, SD card, and power supply) the intel machine is just that much more capable and basically the same price. If you're not constantly thrashing the hell out of it it's not even significantly more power hungry.
Tell me, what is the Pi about?
Because for me, my massive collection of Pis is not about ARM. ARM is how they were able to make such a thing, but I did not buy a Raspberry Pi because I was annoyed with Intel and had some objection to running my workload on those. I bought mine for the following reasons:
1. Cheap, but powerful enough to run a non-stripped OS.
2. No noise, low power, and low heat during operation.
3. Compatibility with Linux with sufficient support that I don't have to go to weird efforts to make things keep working or replace the box if this one breaks.
And the Pi does all of those things very well. A couple minor glitches on number 3 because of nonstandard kernels for the Pi, but those are mostly erased because the Raspberry Pi people have done a fantastic job at maintaining support for all their variants, something other ARM-based SBCs don't do. However, benefit 3 is also available on X86. Benefit 2 is available with small Intel-powered boxes. And that means that, for me and many others who have used Pis over lots of cases, a 16 GB Pi and a 16 GB Intel box may end up looking more similar after all. The Pi has GPIO support that the Intel box doesn't have, but I do not have any use cases that need 16 GB of RAM and the GPIO at the same time, and I could always buy a Pico or Zero and use it to drive GPIO devices for me.
I do want to know what the Pi is about for you, though. There are many use cases where the Intel boxes we're comparing with wouldn't be the best option, but that's not what you said. If there's some other important distinguishing factor, I am curious what you had in mind.
Well, that's kinda my point. The Pi is about - for me at least - cheap tinkering type devices where you get easily accessible GPIO in a low power low cost platform. I don't really see how putting 16G on a pi will be much use to anyone who isn't trying to use it as a "desktop" and if you are trying to use it as a desktop it seems like you'd be better served with something else. It just seems a bit... unnecessary. There's always a use somewhere for basically anything under the sun, but I'm kinda struggling to see the point of this one.
That's not what the Pi is about.
Well no. But that's somewhat the point. If you want to play with LLMs or some form of ML, then an x86 processor with fast RAM is going to be head and shoulders above a Pi - even a 16GB RPi 5 with an "AI" hat (which will run you £60-120, putting a total system cost >£200).
RPi benefits from GPIO and the hobbyist ecosystem around it. In truth, it's hard to see many applications for the Pi 5 at any RAM level - the original "coding"/educational angle (a Pi for every child, a resource in school labs) has largely fallen by the wayside, leaving it split between hobbyist use and industrial. Some people might use it as a light desktop, but a refurbed uSFF desktop is highly competitive there.
In truth there's not really much that a 5 will do that the 4 wouldn't. Better performance on some computer visiony type stuff for sure, but what else?
This is where I think people get a bit confused - for the vast majority of hobbyist applications (robotics, SDR, Home Assistant, PiHole, etc), a 3B will suit. Heck, I've got an original 1B running PiHole. If you're playing with running lots of containers or LLMs, then in honesty an N100 box or Ryzen NUC type thing will blow the pants off a Pi for £200-300 (i.e. about the same price point), and still top out at 20W.
It's hard to see the market for any of the Pis over about £50. I can only assume the 16GB Pi5 already has a business/industrial customer with a specific requirement that made it worth developing.
Things the Pi5 brings over the Pi4B... Built in RTC with a battery option. Fan header separate from the GPIO block. The ability to put real mass storage (NVMe SSD) within the footprint of a standard case, rather than a fat, very stiff, external cable and storage device. There is even at least one M.2 adapter that lets you use the official case and put the lid on for a very tidy package.
My money to sense ratio is a little bit skewed when it comes to Raspberry Pi. I already have a 8G model 5 booting from a 256G NVMe with the active cooler - it's running an OS - quite fast - but nothing else!
If I wasn't job hunting right now, I'd have my 16G pre-order in already ;-)
As the price rises, there is a diminishing return vs performance gained. I've abandoned the pis for anything beyond very simple tasks; with less than 32 gb they're just never going to be the desktop I need.
Instead, I've switched to 8 core, dual thread AMD Ryzen-based NUCs. 550$US/unit delivers 64GB on-board, ROCm Radeon graphics (on-board), dual GB eth, BT and WiFi, with 2 TB NVmE.
The clincher: the Ryzen mobile technology on which they are based consumes a maximum of 20 watts when going full bore. Full bore for this machine means e.g., hosting a 70B LLM while also playing a AAA first-person sandbox shooter.
I have two, which have replaced a half dozen raspberry Pis (8 GB) which I will still use for certain other applications for which they are more apropos: on-site DNS, SDR Radio controllers, and similar.
I don't need 64 GB or even 16 GB for such applications, so I don't expect to be purchasing the 16 GB version at all, especially not at that price point.
In my eyes, this 16 GB Raspberry Pi represents the manufacturer trying to bring the device on par with a typical desktop, and largely failing, due to the constraints imposed by the form factor and relatively high cost vs total compute delivered. In my application, it makes far more sense to up the payout a bit and get some serious high octane bang for my buck.
Best of luck, Raspberry Pi, I think you'll need it.
"Some people just don't get maths."
True, a single machine like that could replace several Pi 5s (more RAM, faster processor, faster storage and more built in) so you need to multiply the cost of the Pi 5's by a factor that is likely to be more than five to compare prices, so they would be more expensive.
OP has currently replaced them with a one to three ratio at the moment (so PIs would still be cheaper) but could increase the load a lot from there.
I'm not sure why you even considered the Pi for that use case in the first place. If you need 64 GB, and I'm not sure why you do but I'll take your word for it, it seems obvious that a board with a maximum of 8 GB until today, which has only been available for a couple years, wasn't going to do it. You're comparing two very different types of computer, as you've described in detail, which are suitable and realistic for completely different uses.
I'm not sure why you do but I'll take your word for it
Well, you probably do need 64GB if you want to run 70B LLM while playing a AAA sandbox shooter. Although that might just move "I'm not sure why you do" one level up, as well as adding "why do you think that's the use case for a Pi?".
Also, as someone with a Ryzen based laptop, mobile Ryzen power use is good, but I wont entirely believe 20W at full whack without seeing some specs. AMD Ryzen 7 5800H is 45W TDP on its own, without including the rest of the NUC. Beelink SER6 (just picking one of this class of device as an example) seem to come in around 55ishW on load with occasional 80W excursions under heavy loads.
The Pi isn't a useful LLM device. Shouldn't even be part of the conversation if you want to do LLM. It can, with the HAILO hat do a fair job of optical AI runtime tasks.
LLM isn't the only use of RAM. I've no idea why it's mentioned in the article.
16GB allows me to stuff more roles onto a single device that sits there with a tiny footprint sipping a few Watts. It is going to allow me to consolidate multiple Pis to a single Pi, removing all the cables and stuff. I'm quite pleased that one has appeared. I have my preferences and I also see a lot of other stuff that people do that I don't get. I don't need to question them.
The people at Raspberry Pi know their market. They are aware that other stuff like Orange Pi and Rock Pi are available with 16GB.
I was looking for the revised version of this one:
https://www.raspberrypi.com/products/compute-module-4s/
More suited to a blade server deployment…
(Charlie Clarke didn’t specify exactly what type of blade server, so given the Raspberry theme….)
Lesson: clearly Raspberry are into confusing names just as much as other vendors such as Dell with their revised computer range name conventions.
Actually, it's not really the cost of the memory that's the difference but the integration on the SoC with the various CPUs and GPUs, including the ability to avoid copying between them. Still, although it's a hefty premium that many are happy to pay, because nothing else provides the same performance for some tasks. Probably going to get a new Mac this year myself to replace the current 2020 MBP (the 2016 reserve is still going strong) and while I'm not keen on splashing out, at the end of the day it's essentially a bookkeeping exercise for any professional user.
I'd bet against it. They are not building the x00 models to be the highest-end computers they can imagine. If they were, they would have put a connector for NVME drives into the 500 they actually made, since that is useful to making the desktop substantially faster. They decided that wasn't what they were going for, which probably means that 16 GB is also not what they're going for.
I am curious though, why do you want one? What is the value to you of having an upgraded 500 over just having a Pi 5 attached to a keyboard of your choice? I ask because they don't look very different to me and thus the 500 and 400 were never very appealing to me, but enough people seem to like them that I'd like to know why.
"I want to see someone really show off and show us all how to create an arm64 virtual machine without just yelling qemu and running off into the distance. A step by step will be fine."
Step 1 - download Alpine Virtual aarch64 ISO image: https://dl-cdn.alpinelinux.org/alpine/v3.21/releases/aarch64/alpine-virt-3.21.2-aarch64.iso
Step 2 - make sure you have qemu-system-aarch64 installed on whatever x86_64 Linux distro you use
Step 3 - create a aarch64 VM in whatever hypervisor/hypervisor frontend you use on your Linux distro, whether qemu-system-aarch64 on the CLI or libvirtd etc as a "frontend". For the new VM specify the downloaded Alpine ISO to appear as a CDROM device, specify a block device (ideally virtio-blk or virtio-scsi) to be created as a disk of suitable size (maybe 700MiB or so)
Step 4 - start the VM. NOTE: running a aarch64 VM on a x86_64 host will be sloooowwww (could take multiple minutes to boot) as no hardware virtualisation acceleration is used.
Step 5 - once Alpine has booted then run "setup-alpine" and install using "Sys mode" to the block device (i.e. /dev/vda or /dev/sda depending on what type of block device you specified earlier)
Step 6 - once Alpine is installed then run "poweroff", remove the CDROM device from your VM's definition and then power on the VM again and (eventually) it should boot from the block device.
That's just a quick off-the-top-of-my-head step-by-step. I do run Alpine aarch64 VMs occasionally, I just personallty don't use "setup-alpine" to create them.
My current desktop is a Lenovo Tiny M900 i7 which I picked up with 32GB for $100. I replaced the drive in it, but I would need to get a disk if I wanted to use a Pi5 as a desktop replacement, but also a case and a PSU as well. Its getting harder and harder to justify picking up a sbc especially as a desktop replacement once I realised you can just pick up 2nd hand enterprise kit for a fraction of their original price and in many cases cheaper than the sbc equivalents when factoring in the tco. Same goes for any of my server projects too.
The only home project that I have that has survived getting virtualised into a kvm vm has been my pikvm box that allows me to remote control my other M900 that runs all my servers. With the advent of all those tiny kvm products, if it dies I'll prob get one of those as they are a fraction of the price of my pikvm (I'm running it on pi4 with usb hdmi and old school usb splitter).
It's worth noting that second hand kit comes in a variety of conditions from barely used to having been run 24/7, looked after very well or abused to near death in harsh environments. Corporate kit is usually refreshed every 5 years or so because the likelihood of failure has increased. You spin the wheel and take your chance. Or you can buy a Pi brand new which may suit your needs and be pretty confident of a decent lifespan from delivery.. But only if it suits your needs. I've gone both routes for various reasons which suit my needs :-)
I bet there's hundreds of thousands of them just gathering dust.
Then there'll be ones that have only ever been used for pure compute tasks - iow, the GPIO never used.
That's great if the device is the same price as the PI used to be.
However, for $120 if all you are going to do is use it as a mini computer for pure compute tasks and not do any GPIO stuff, a second hand NUC is a way better option.
I dabbled using the PI for media servers years back, but found them not powerful enough.
They sure are now, but I opted for a second hand NUC and never looked back.
I can't say I'll buy another PI again - got four of them in various states of slow decay.
"However, for $120 if all you are going to do is use it as a mini computer for pure compute tasks and not do any GPIO stuff, a second hand NUC is a way better option."
Not if, for roughly the same outlay in terms of component costs, the Pi *also* gives you a factory fresh piece of hardware and consistency with an ecosystem you're already familiar with.
These may not be concerns for the amateur dabbler who just wants the most bang for their buck and is happy to spend hours of their own time learning a new system, and dealing with any potential wear and tear issues that come with using second-hand hardware that's been through the hands of unknown former users in unknown former use cases.
They very much ARE the sorts of concerns professional users of RPis will be taking into consideration if they either have existing designs based around the current RPi5 which are becoming constrained due to the memory size, or if they're looking to develop new designs where they already know attempting to use the current 5 would be a non-starter due to the lack of memory - in both cases being able to continue using a Pi means being able to take advantage of all the built-up experience, code base etc. they've already got, rather than having to spend time (and money) on learning and certifying a completely different platform, and in the former case being able to simply swap out the current Pi5 for a larger memory variant, without needing to worry about redesigning any other part of the product, could also be a significant time/money saving.
And even if they're happy to switch to a new platform with all of the costs that would incur, they'd also need a VERY compelling reason to switch to a platform being supplied as second-hand parts - in the world of commercial engineering, using second-hand parts is typically a matter of last resort if you've been unable to source newly-built parts and need to maintain production until either you EOL the product entirely, or you've managed to redesign it around parts which are still available new. So instead of comparing the cost of this new Pi against equivalent performance used parts, I rather suspect that a lot of its potential users would be comparing it against the cost of equivalent performance new parts, and coming to a rather different conclusion than you've done here.
The early commercial ARM servers were a pile of shit. Headless and the NIC's would die if nothing else died sooner which turned them into paperweights.
The PI's were slower but kept on keeping on.
The problem with ARM wasn't so much porting or supporting software, it was having machines fast enough to be able to run build and test without slowing everything to a crawl.
Except that their comments are a closer analogy to "for the price of a new motorbike, I can buy a secondhand car". The thing they're comparing it to tends to be faster and more expandable at the cost of higher power usage and being secondhand. I don't consider most of these comparisons very useful, but your analogy misses their point by a wide margin. From their perspective, they are getting more for their money because the systems they're buying can deliver computation faster, and computation speed is one of the most commonly used metrics when assessing computers. We can argue why this comparison is not good enough for us, but arguing that they're wrong for making it in the first place doesn't make sense to me.
Except that thinking about the RPi as being merely a "computer" is the wrong way to go - do that and sure, you can find a myriad of more attractive looking alternatives once you head up into this performance bracket.
Whilst *some* people are using it just as a low-cost low-power alternative to pressing an old desktop/laptop into service, there's now a *lot* of commercial users of Pis who are using them as the computational core of their products, but *also* making use of at least some of the additional features that you don't get so easily from those alternatives. For this class of user, the Pi is anything but a mere "computer", it's a fairly specialised component of their products, and one which has few rivals at present.
And no, their comments aren't a closer analogy to that - most if not all of the ones I've seen so far are along the same lines as IGotOut states, i.e. comparing the cost of a *new* RPi against the cost of a *used* alternative with equivalent/better performance. So they're getting better performance at a lower cost, but at the expense of it being used (and potentially abused by its previous unknown owners running previous unknown workloads). Which makes a BIG difference to the comparison, because whilst some Pi users might be quite happy to switch to an alternative from the used marketplace, others (like commercial users) absolutely will not, at which point you can't simply continue to treat these new vs used comparisons as the sole arbiter of whether or not this new Pi makes sense.
Put it this way, RPi the company seem to know what they're doing, and they've clearly come to the conclusion that releasing this variant makes sense. So anyone questioning this really needs to take a step back and ask themselves why they've come to a different conclusion than the people who make the things and probably know rather more about the market for RPis than anyone else. Just because it might seem like a nonsensical thing to do based on what you as an individual might know about the RPi market, doesn't mean it's nonsensical based on those parts of the market you're not aware of, and it's painfully obvious from some of the comments being made here that some people really don't know just how big and varied that market now is.
"Put it this way, RPi the company seem to know what they're doing, and they've clearly come to the conclusion that releasing this variant makes sense. So anyone questioning this really needs to take a step back and ask themselves why they've come to a different conclusion than the people who make the things and probably know rather more about the market for RPis than anyone else."
I think this explains why we view those comments differently. I do see some people making those statements who should consider the question you stated. However, many of those comments don't seem to be saying "Why would they make that?" but rather "Why would I or someone like me buy that?". They're not buying them for embedded purposes, so they're not saying that a secondhand computer is suitable for every case where a Pi is. They're saying that larger machines could do the same things they'd do with a Pi and they assume it is for most others posting here, which based on the comments is not an unfair assumption in general.
I don't agree with them when they're comparing it to secondhand devices. I find that most of the ones that are available have much higher power usage or some kind of damage, and also the people who make that comparison seem better than I am at finding secondhand machines with impressive specifications at really low prices; the ones I see are less impressive. However, I do acknowledge that for them and for many of us, the Raspberry Pis we buy are used in a way that a larger computer could fill in for. Therefore, when they express a view about the price of one, I can understand why they're thinking that. It's the same way that I might call out a restaurant for being ridiculously expensive. Maybe you have a client who loves that restaurant and always buys something expensive from you if you take them to lunch there, and that would justify nearly any price, but you know that I'm thinking of the average customer when I make my comparison. I don't have to agree with them to understand why they're making the comparison and how it could apply to plenty of other perspective buyers.
I picked up a complete Intel N100 based NUC style mini PC, with 16gb of DDR4, a 512gb m2 ssd (sata admittedly) and a seemingly legit Windows 11 pro UEFI embedded licence... for £99 a fortnight before Christmas from Amazon.
If they are going the cluster path - a caddy to hold a crap tonne of compute modules would seem like a better option surely?
I'm sure these will sell out to people who will leave them in a drawer or use them to run retroarch with no meaningful improvement in capability or performance over the 4 or 8gb varients.
Every time Pi release a new product, many people say “Too expensive”, “Abandoned the hobbyists” etc. Everyone seems to forget that Pi still make ALL (well, most, the very original Pi1 was superseded) of the previous boards, and all at the same original prices. So taking inflation into account they have got cheaper over time! People also seem to want these newer highly specced boards at the same price as the older ones. But that simply isn't possible - you need to pay for all that extra performance. You simply cannot make and sell a quad core 8GB device for the same money as a single core 512MB device, for reasons which seem to be obvious except to these commentators.
Here is the current Pi price list, for those non-believers. And for all those allegedly abandoned hobbyists, these are the prices EVERYONE pays, commercial, industrial, maker, education, and hobbyist. No-one gets a discount for quantity and all those commercial sales keep the cost down for everyone else. The hobbyist benefits hugely from the industrial sales of the Pi - they have certainly not been abandoned.
Raspberry Pi Zero $5
Raspberry Pi Zero W $10
Raspberry Pi Zero 2 $15
Raspberry Pi Model B $35
Raspberry Pi Model A $25
Raspberry Pi Model B+ $25
Raspberry Pi Model A+ $20
Raspberry Pi 2 Model B $35
Raspberry Pi 3 Model B $35
Raspberry Pi 3 Model B+ $35
Raspberry Pi 3 Model A+ $25
Raspberry Pi 4 Model B $35/$45/$55/$75
Raspberry Pi 5 $50/$60/$80/$120
I want to run a Korean MMORPG server. The game install setup files are a mere 2GB. I think the game runs its databases in SQL. I could load the whole game in the RAM. The game itself runs in a shoe these days, 2.5D parametric view of 2D sprites.
Having it in a tiny box with an assigned IP running at home would be better than some old crude PCs I got laying around, burning several watts more.
That would be neat.
The problem is that if you had to support ARM Linux there weren't really better options.
Yes, you can build with a cross compiler which works but is hard to maintain, test, emulator, well sort of but you rely on the emulator being accurate (and they weren't for ARM).
Mac wasn't really a good option apart from using them for OS/X support. Accuracy was an issue if nothing else.
Buy one of the server boxes ?. Well we tried that, they were relatively expensive, hard to source and died young. And the larger ones were really expensive and very hard to source.
So that left a 4G Pi. Which worked mostly, slower than we wanted because we didn't typically ship until all platforms built and passed FVT/performance. (Burnt a few times by failures on one platform being ignored but actually warnings of failures that would bite someone on one of the other platforms eventually).
So, the 16G Pi would be a godsend in that situation, in practice it'd be a lot faster, not just CPU speed but lack of memory pressure.
It's fairly niche admitted but faster is faster and it's actually ARM so testing was mostly an accurate indicator of what would happen when scaled up.
This post has been deleted by its author
Needing a low power security camera capable of capturing car number plates at night, I may finally have a practical use for a Pi. You see, nearly every security cam prioritises low ISO over graininess, which results in a visually pleasing but useless image when anything is moving at any speed faster than glacial at night. And it's extremely annoying that most security cameras don't provide access to ISO or shutter speed via software UI, or any other way. In case I need to sue a driver to death, I record every one of my ebike rides using a GoPro Hero 7 Black, and at 1/960s and ISO800, there is very little number plate blur. I can record every nearby number plate, even at night. Setting min ISO to 100 and max ISO at 800 means I don't have to change any settings, as day turns to night, and I'm hoping I can get a Pi with Pi Camera to pull the same trick. Here's an example from last week, and this wasn't close or fast. Underexposed, has to be brightened in post:
https://www.youtube.com/watch?v=KkVZw-1Ip-0
I mean, have you seen all the blurry faces and screengrabs of blurry cars cops send to the media, hoping the public will recognise the blur or some blurry part thereof? No security camera I've seen is fit for purpose! Not if that purpose is shooting unblurry video at night. Every dashcam I've ever seen also suffers the same issue. I don't mind graininess if there are a lot of frames per second, but I don't want blur because that's utterly useless. This won't need 16 freaking GB of RAM.
The logical use seems to be a NAS controller. Assuming you have a UPS, you could disable synchronous writes and get really good performance with ~12 GB of RAM cache. It could handle several security cameras, computer backups, remote sync, and file serving all at once with no problem.
My daily driver is a Pi5 with 8GB of ram. The game changer for me was adding the SSD hat and installing Ubuntu OS for Raspberry Pi on the drive. I use the machine for mainly email and web surfing. It plays music and videos just fine. The CPU only seems to start up on start up and reboot. I rarely turn it off, except on vacation.
The Register 
Biting the hand that feeds IT
Copyright. All rights reserved © 1998–2025
