So, I saw this link and immediately thought, "I bet the discussion will be about how we should use N100s, instead." I wasn't disappointed. Even on dedicated Pi forums, you see that happening.
I guess I understand this point of view if you were trying to use Pis to experiment with Kubernetes or something. You'd have built your own (desktop) PC or a personal server rack for that kind of thing, years ago. But for the vast majority of typical uses (Home Assistant, VPNs, etc.) a Pi is going to be way more than you need. It will sit there and silently and reliably run, for years at a time, powered by a USB cable. I know mine have.
Why would I consider replacing those with a bigger box, fan noise and a power brick? Maybe I'm missing something?
I think the reality is that a lot of people don’t use or care for the main Pi features like the GPIO pins or camera connectors, and aren’t using them in places where space or power is constrained. They’re purely looking for a cheap amount of compute.
For those use cases, it makes sense that performance per dollar is the top metric that many people look for. Pi used to be king there too, but as Pi prices have gone up and the price of other cheap compute has come down, it’s harder to justify the Pi. The Pi is still a great board, but a Pi5 with case, power supply, SD card will run you $100+ (throw in an SSD and you’re at $200+ easy), meanwhile an N100 PC can be had for the same price.
It's kind of like a keyboard without LED lights for status, pop up tabs for the angle, and a recessed guide for the USB cable. Would it technically be cheaper without those things? Sure... but it also costs next to nothing to add them. Once a keyboard gets to $10 almost all of the cost is already put into "being the best keyboard for $10".
There are certainly cheaper, <$100 TCO, alternatives to the Pi or N100 PCs... but they are accordingly worse performing. One thing Pi is really good at is having well supported options on the ultra low end. The Pi Zero 2W for $15 - it's crap performance but cheap. The Pi Pico microcontroller starts at $4 if you don't need a full OS (or just need to augment an existing box with GPIO over USB). If you're building a PC out of a Pi it's just not a differentiated option vs an actual PC is all.
N100 are often no fan and the power brick need not be much different than the one that feeds the Pi. At the end of the day the max wattages aren't much different.
The biggest thing the Pi line gets you is good GPIO and premade hats using it (with prepackaged OS images). If you're just running standard software like a home assistant or VPN then it doesn't make much sense. Doubly so on articles about OCing it.
I think a lot of people on the internet try to use the Pi as a home media server or file store and can’t envision it being used for anything else.
It’s versatile but really it has always sucked at those things. Yes, get a used N100 if that’s your goal.
The Pi shines in education and as an embedded device where power, size and maintainability are important. I spotted one the other day in one of those presses that crush pennies as a souvenir. It displayed an instructional video on a screen and also seemed to control the card reader device on the front. Why would you use a N100 in that context?
2W, maybe. Pico, no. Driving a payment system and displaying an interactive video is still a job for a SBC not a microprocessor, even a modern one.
Raspberry Pi have found a price level and form factor that works for customers. The raw speed of the 5 is probably overkill for most user applications (the 4 is too IMO) - though edge AI and robotics need it, so there is demand. But ultimately the speed is there because a faster ARM chip fits within the cost-power-size requirements of the board.
Pico was meant to shine for much of the education half (e.g. micro-python controlled mini robots without the hassle of complex and poorly supported developer environments, OS management and configuration, or hardware complexity at a couple $ cost point), not the interactive media half (which would target the Zero 2W usage case over a full blown PC style Pi).
The compute module variants do tend to shine in the edge robotics niche where you need to do some processing in a custom form factor and don't want to design or source a SoC directly... but you don't really need the "rest of the PC" the standard Pi models provide. Pis (any model) are a bit of a shit fit for edge AI in general though. The memory bandwidth is god awful (even compared to cheaper SBCs) and the hardware to meaningfully offload matrix operations just isn't there like it is in a lot of other embedded boards fit for purpose. This leaves you spending more $ to use more Watts for less AI compute vs typical embedded AI options.
The Pi foundation's big break was targeting ultra low cost markets (i.e. sub $50 TCO) with a well supported device that has/had a lot of hacker community mindshare. Any product they've made beyond that focus (i.e. not the embedded or microcontroller variants) has been a pretty bad technology fit which ends up in more drawers than deployments despite the hype.
And they sold NUCs to ASUS. For mini PC servers I would always go intel for QVS transcoding. AMD is not on the same level as Intels media engine.
Intels iGPUs also support SR-IOV these days so you can get GPU acceleration to multiple VMs.
Streaming (local network - Sunshine/Moonlight) these days is really good. So good I consider moving my gaming machine to the garage and just stream it to a thin client or steam deck.
Interesting, but why bother. Pi's were great when they were cheap, but one can get a better performing Intel N100 box for a similar price and power consumption now.
The only reason I can think of to use a Pi is GPIO.
As a person who uses both, I won't leave an N100 system where I can leave a RPi5.
First, the overall footprint of an RPi5 is way smaller, plus RPi is easier to cool passively than an N100, also it allows RPi5 to be lower maintenance on the longer run.
Most N100 are noisier, hotter and bigger than a Pi5 (I have a smaller N100, but it's still actively cooled).
When paired with a decent A2 card (e.g.: Kingston Canvas Go), the SD-card related lag disappears. Add a good external SSD, you have tons of solid state storage. At that point you can just forget it.
I like my N100, and companies like GMKTec routes 2 PCIe lanes to NVMe slots allowing ~1.5GB of throughput, but for most home use, that's overkill. I rather use Pi as a server and N100 as a desktop (which is exactly what I do).
> When paired with a decent A2 card (e.g.: Kingston Canvas Go), the SD-card related lag disappears. Add a good external SSD, you have tons of solid state storage. At that point you can just forget it.
At that point you can just underclock an N100 NUC for the same price and almost the same TDP (it's pretty much even with the SSD). Most vendors configure the CPU well into diminishing returns for marketing reasons and even a small undervolt can make them passive (the same is true of many laptops).
I get the form factor issue but exposed RPis without the enclosure (more $$) are a burned out board waiting to happen. I use them in enclosed hardware that needs gpio like 3d printers but the N100 wins every other time.
You can't rule out a fan sucking in dusty air from side or worse, bottom of the case. Most N100 systems have too much air circulation inside to forget them in the long term.
Many N100 systems come with 5W TDP as default. However, this is more of an "average" TDP. It doesn't prevent the CPU to enable its turbo and exhaust its thermal budget. Forcing it to lower envelopes via cpufreqd would work, but you're limited by your fan again, which will make everything dusty, eventually.
I have found a pretty cheap completely aluminum case [0] which sucks heat from every IC and bottom of the case via silicon thermal pads. It keeps the processor at 37 degrees, cools USB bridge and power controller too.
With no moving part, case is almost airtight. So, it's sitting snug and cool in that case. You can't burn RPi in that configuration.
I went from a passively cooled RPi5 to a N100. The RPi idled at around 50°C and throttled under load. The N100 keeps 56°C under light load (the heaviest things are rtl_433, grafana, nextcloud). To keep it from throttling in heavy load I zip tied a small Noctua NF-A4x20 to the side of the case over the passive heatsink and adjusted the fan curve to only run at above 60°C. This works super well, the N100 doesn't get hotter than about 62°C under full load and the fan spins slow enough that I can't hear it at all even if I put my ear directly next to it.
> To keep it from throttling in heavy load I zip tied a small Noctua NF-A4x20 to the side of the case over the passive heatsink and adjusted the fan curve to only run at above 60°C.
I applaud your creativity but surely this is an indication of the limitations of the architecture!
Nice solution, but with the case I linked, my Pi idles around 37, and reaches 50-ish (both degrees C) at most (if I load it hard), since the case is sucking heat from all chips and bottom of the PCB.
How does it handle electrostatic discharge when touching it in that case? I used a case that was really more of an oversized heatsink[1] than a case and it was really sensitive. Managed to crash the rpi a couple of times when I touched it without touching ground first a couple of times.
edit: I replaced the thermal sticker for thermal paste, not sure if that could have affected it, the paste wasn't the conductive kind.
Thats somewhat surprising. I have an old 14nm Celeron I use as a home server and under typical usage it is a few degrees above ambient. Checking now it is 26C. There is no fan, and it just has whatever tiny stock heatsink it came with.
That will depend on what other hardware they put on the board. In general though N100 is just a standard x86-64 so you have a lot better odds of good support.
The n100 is an intel board and all the parts used in the cheap nucs I have tried are supported on quite old mainline kernels.
At about twice the price for just the pi, I got a n100 machine (with a case and power supply and 500mb ssd and 16gb ram). Draws about 10 watts, is faster on every axis, and once you setup the pi with a case/psu it’s very close to the same price.
10 watts isn't idling, that's working. The chip has a TDP of 6 watts, but the RAM uses some too. I turn mine down to a lower clock speed, which takes off about 3w.
> You need a power brick for the N100 box.
Yes, but it was included (along with the case) for $106 USD, delivered.
Granted, that was on special, but AliExpress has sales that get them down to $115 USD pretty regularly.
I find it quite interesting that we're beyond the point of having an Intel processor being a massive advantage. I think 10 years ago most people would have jumped at an x86 processor with similar specs to the equivalent Raspberry Pi.
That a distro's principal maintainer steps down is unfortunate for anyone who is downstream from them, but it is not the end of the world. As the article mentions, other distros for Rockchip continue to exist.
> the official Linux kernel is more based on SoC vendor's BSP (Board Support Package) that can utilise full SoC features. And Hardkernel is able to consult with Rockchip in case we found missing SoC features only when the issue is discovered in their BSP, otherwise it's very hard. Therefore, we are managing the Linux kernel with most similar OS with the version from Rockchip, even Ubuntu 20.04 is not from Rockchip.
> The images with upstream kernel would not be stable enough and features are not fully supported, so please do not use this image for your important project or to store important data.
Since late 2023, ODROID-M1 is supported by mainline u-boot (odroid-m1-rk3568 target) and you can also use a mainline kernel (6.4 or later) which provides working USB 3.0.
GMKtec Mini PC N97 [1] is 160 EUR on amazon.de (with the 70 EURO voucher that is). Got mine with free shipping to Romania, as the purchase was over 50 EUR.
Also, the N97 has better performance than the N100, but the power consumption is higher. For me its a plus as it's more powerful than the usual N100 mini PC's while much smaller and fan noise is quite low (for now at least...). Not really intended for high workloads and 24/7 anyway. More like a backpack PC.
Over at AliExpress there are tons of options, like this[1] or this[2] fun 4x M.2 box.
Yes they're about twice the price of a bare Pi5, but they come with metal cases and power supplies, can fit an M.2 NVME without add-ons, and like the first example also includes some RAM and storage.
While likely not the best quality RAM and storage, what I got in mine has held up to light use for several months so far without issues, so better than nothing. YMMV.
I bought a few 2GB RPi4's when they were $35. At that price point they were great. I don't get the RPi5, especially 4GB+ variants.
I haven't got any personal experience of them but aren't there a number of Raspberry Pi clones which offer the same functionality at much lower cost? Seems like that would open the gap up.
The issue with nearly all Raspberry Pi clones is documentation. Boards tend to be undocumented, sparsely documented, or documented with errors. It makes for a frustrating experience when you run into problems, especially when those problems are with bog-standard Raspbian (or whatever). Raspberry Pi hardware has occasional quirks (eg, don’t try to hotplug USB on a Raspberry Pi Zero), but quirks are the norm for other SBCs (anything from Pine64 is an exercise in frustration). So the modest upcharge for a Raspberry Pi is usually worth it.
There are some, but a lot of them have very spotty Linux support. In most cases either you're stuck with some ancient kernel provided by the manufacturer, or you're relying on community support from Armbian at best.
Take for example the OrangePi Zero 3[1], which you can get[2] from around $25.
Armbian provides community support status[3], so you're relying on volunteer efforts to provide updates and support. Also drivers for peripheral devices might be missing[4].
There are others, like say the NanoPi Zero2[5][6] for around $20, but again, support might vary. The Zero2 is so new it's not listed on Armbian, and most of the other NanoPi's are under community support.
That said, I've been running some NanoPi ZeroPi's (the Zero2's predecessor) for several years now, without any issues. Just using them for light loads, one is my PiHole for example, but yeah can work great as long as you do a bit of research in terms of what's supported and not.
Right, it's twice the price but like 10x the capability (unless you need GPIO).
That's why I think the lower-end, lower-priced RPi4 made more sense. Fast enough to have decent network speed and run a lot of functional workloads, cheap enough that you could "spam" them.
I use it in a 3d printer. I don't need 10x the capabilities. For my usecase it's just 70 EUR more, with the only benefit it might boot and update faster.
No, I think that was the price from the Spain warehouse and included tax. Just 10 EUR shipping. Aliexpress got better with showing you the real prices.
Yes, as a learning-first (teaching) platform, the custom DE is an advantage.
But further, the R.Pi does not need to be an over-sized solution or lead a benchmark with a top score. It just needs to be fast enough for learning and hacking with GPIO.
A computer doesn't need to have a gigahertz burst mode just to run an efficient little OS for coding.
As for Wayland, it only recently became the default WM. But it is nice to have.
> N100 is inexpensive, has impressive performance, and can support 16+ GB RAM.
Yep runs 32 GB just fine, even if not advertised. Also great Software support as usual on the PC Platform. Also more I/O, esp. PCIe than a Pi and you actually can buy boards with standard connectors. Also at least for server type use cases x86 is most of the time probably still a lot less painful than ARM.
With the appropriate license, there is no particular reason Apple's team couldn't have built a good x86 processor. Switching to ARM was a matter of necessity, not due to a superior architecture.
Architecture largely doesn't matter these days, just how you implement it and what fabrication you build it with.
> = current Debian supported "out-of-the-box" with a custom D.E. and Wayland
Effectively the same thing for basic desktop use, disregarding wayland, as doing a basic bare bones debian x86-64 CLI only install (totaling something like 1.2GB disk space used) and then "sudo apt install xorg xfce4 xfce4-goodies libreoffice gimp thunderbird firefox-esr gthumb nautilus konqueror", or whatever other list of applications one wants to install along with xorg and the desktop environment.
Why buy Intel N100 with Intel Management Engine (which is disabled on US Government computers)? I rather buy RaspberryPi where I don't need to disable Intel ME by myself...
You ca get GPIO on the cheap from a Arduino, ESP8/32, or RP2040 microcontroller. RPi itself is far too wasteful for that.
> People using them as cheap computers ruined it.
False, industrial users using them in all commercial embedded products ruined it.
Almost nobody is using RPis as cheap computers at home since they suck at that: too expensive, too underpowered and too limited in functionality to be a computer for the average joe
My RPi 2 becomes an unusable stuttery mess when I open a web browser and for the price of a RPi 5 I got a used quad core Intel 8th gen laptop with 8GB of RAM and 256GB SSD which not only beats the RPi at any and all computer tasks, but it also comes with built in keyboard, 1080p monitor, battery, charger and I can take with me to do work on the go.
The only thing I see the RPi good at at home is a headless PiHole/VPN or whatever, not as an actual desktop computer replacement.
A Raspberry Pi was a full Linux computer with GPIOs. You can run whatever daemons and still control a relay board. No microcontroller fills that niche. There still isn't really anything in that niche other than single-board computers (i.e. Raspberry Pi clones) - for example, NUCs don't have GPIOs on them.
Of course you can fall back to the previous standard idea - connecting an Arduino to another computer and communicating by the serial port.
They were cheap computers, so people used them as cheap computers, I don't understand your point at all. For a quite a long time they were also the de-facto standard way of making a LED blink over the internet without having to dive into the embedded space.
IMO it is the people using them as embedded controllers for their products that ruined it, in combination with the poor decision by the Raspberry Pi folks to prioritize commercial demand for the products over hobbyist demand during the pandemic. We hobbyists "made do" with other solutions in that time, and now there's no real need to go back.
Good luck writing software for one of those N100 boxes only to find out they're sold out and you now need to extend the code for new model over and over.
N100 is standard x86. Write generic code that runs on a generic linux kernel and you won't have to worry about if this is a N100, N200, AMD, or even something weird like a 30 year old RS6000. Well if you need GPIO you have to worry a little, but you are forced to spend an extra day to support a generic GPIO framework instead of directly writing to PI pins - that extra effort means you don't care.
Even in embedded where you have contracts to ensure the same hardware will be available for a decade everyone has moved that way because eventually the hardware you rely on goes out of production and so you need to support new hardware.
As a desktop, one can get a used x86-64 dell ultra small form factor quad core Intel core i5 something, 16GB RAM, something 128 to 256GB SSD on eBay for a hundred bucks... It'll have integrated video capable of driving one 4k at 60Hz as native displayport. And a decent amount of other gigabit Ethernet, usb3, analog in and out audio, etc.
Yeah it'll be 7 years old but it'll still be much faster than a rpi5.
Even a mediocre Intel 120GB sata3 ssd from 8 years ago will run circles around a microsdxc card in disk performance.
These mostly come from corporate lease environments and have very little write wear on the ssd and are very clean.
The raspberry pi is more useful if you need the gpio pins or want something really small and perfectly silent for some embedded application.
From a sata3 ssd in a usb3 enclosure? adds cost and complexity and cabling... I understand there's also a raspberry pi5 sata "hat" which is something like $20 and then figure at least $35 to buy a decent quality small sata3 ssd to add to it.
Not the author, but on a Pi 5 small LLMs such as tinyllama or qwen2.5:0.5b run at over 25 tokens per second (haven't tested the performance boost with it yet). It could be useful if you wanted a local assistant or an AI server for devices you don't want to run LLMs on
Is there some kind of list that shows comparable performance over all of these changes? There's that ARM 4_0_X_X quant, NUMA patch, now this SDRAM tweak. Plus if there's any delta with your llamafile speedups vs. openblas. I feel like we should be approaching over 3 tok/s tg for 7/8B models at 4 bits by now?
Last I tested, the ARM quants were actually slower on a Pi 5 than Q4KS, but I later learned I messed up some compile flags and haven't had the chance to retest lol.
> You can actually monitor the temperature of the sdram and it reports if refresh at half or quarter the rate can be done. That allows the overhead due to refresh to be reduced by a half or a quarter which does improve benchmark results.
I wonder why this is not available on PCs. tREFI tuning has a big impact on performance, and having an automated guided optimization based on temperature would be nice.
Wow... I haven't heard of someone using refresh timing to speed up a system outside of retro hardware in ages. I guess I'd just assumed that after we'd transitioned to DDR that refresh cycles were just such a small part of overall cycle time that it didn't matter. Goes to show what I know!
Faking NUMA can help enforcing access patterns that make things faster:
> In the worst case, two sdram clients accessing different pages of the same bank, may cause repeated closing and opening of the pages, harming the usable bandwidth you can achieve.
Which pages belong to a bank depends on the physical address lines of the buffers used. We have found that it is pretty common for buffers to be allocated in pathologically bad ways.
> NUMA allows us to have more control over this. We can split our sdram, into, say, 8 NUMA regions, and configure the kernel to interleave the allocations between the regions.
In addition the sdram controller allows the address bits used for segmenting the banks to be reconfigured (which we've exposed through eeprom config SDRAM_BANKLOW).
If it works on a Pi, wouldn't this make sense to also implement with DDR5? Or is it already there? Now that the chips are split into 2x mux of 4x32bit busses with dual rank on dual channel, it would really suck if the data was defragmented in such a way that only one 32 bit bus could access it and only half the time (if my understanding of this is right anyway).
I'd like to see the numbers around the power usage added too, I see the Pis and audrinos personally as a low wattage device that I don't need to worry about the cost, I'm sure this overcloclocking isn't too much but would be nice to see to weight up viability
Jeff has a N100 vs. Raspberry video where he compares power consumption.
I think he used a GMKTec G3, and the numbers were not bad, but active cooling adds a constant whine to background. Some companies' fans are more whiny than others, too.
N100 is not a "cool" chip by any passive cooling standard, esp. under load. However, its standard (software limited) TDP is around 5W IIRC. You can increase that limit for more speed and added noise.
If I'm going to remove the case and do my own modding, then all best are off. I want something working well out of the box. I have enough computers to play, and not enough space to put Frankenstein systems securely around the house.
Fair enough, but the RPI 5 also won't reach max performance out of the box until you add a third party cooling case on it, be it active or passive. N100 systems at least come with a case already on out of the box that let them reach max performance albeit with some noise.
And we're talking about tinkerers after all, which is the target market for the RPI. Parents won't be buying RPis en-masse for their kids from Walmart to use as their main computer for school instead of Macs, iPads, Chromebooks or Windows laptops.
Average Joes who aren't into tinkering, aren't daily driving a RPI as their main computer. It's a DIY toy/tool for those who want to tinker, learn and make stuff, so some added effort is implied anyway.
Yes, that's sold as a DIY computer, so getting it a case is a must. OTOH, I'll be honest and say, both the official cooler and official case for RPi allows it to run at full-steam, almost all the time.
I especially bought that passive case for my Pi, because it's destined to be a "tuck away server" from get go, and it's trucking along just fine. I'm adding services on top of it slowly, and it's not showing any signs of fatigue.
A2 card and faster SD support really makes that one shine, without adding hats on top of hats. Again, OrgangePi 5B has better hardware, but way worse support.
Yeah GMKTec is pretty bad when it comes to cooling solutions. Literally just had the K9 with the 125H delivered yesterday and while I was expecting the fan setup to be meh, it's honestly like a bad case of tinnitus. Constant eeeeeeee at the most annoying frequencies possible.
They put a tiny 40mm fan on top which resonates with the plastic box and screeches like an absolute banshee all_the_time. That one is a must-remove since it also blocks installation of any heatsinks on the nvme/ram, and a mesh top results in roughly the same temps. But they gotta put their logo on top, priorities.
The second problem is the main cpu cooling, which is actually a really well made bottom intake blower side exhaust setup... but it's set up to run at like half speed even when you're at 30 degrees and there's no bios fan curve control for it except for extreme temp steps. I'm currently pondering adding a pico or attiny that reads the board PWM and retransmitts it so it only runs the fan when it's set above usual. Or maybe just a full passive copper block replacement, but that would require a case redesign...
Otherwise it's a really capable little machine if they put a little effort into making it shut the fuck up at idle.
G3's fan is mostly silent. You might hear a small clicking on lower RPMs, and mostly a whoosh on higher RPMS, OTOH, it' still moving air and sucking dust from the bottom.
Looks like K9 needs to move more air since it's packaing an U5 in that small box, so cooling them silently is much harder.
Iirc the N100 idles at 6W and boosts to 25W, whereas the 125H idles at 35W and boosts to over 100W, which the cooler has to be designed for of course. The Ryzen boxes have similar cooling problems from what I hear.
But if you run the N100 flat out, which is likely to happen given that it's not the fastest thing, and disable turbo on this one (which I've done for now since my use case is a low power server) they are actually surprisingly close in heat output. The Radxa X4 tried to cool the N100 like a Pi 5 and it almost melts.
So, I saw this link and immediately thought, "I bet the discussion will be about how we should use N100s, instead." I wasn't disappointed. Even on dedicated Pi forums, you see that happening.
I guess I understand this point of view if you were trying to use Pis to experiment with Kubernetes or something. You'd have built your own (desktop) PC or a personal server rack for that kind of thing, years ago. But for the vast majority of typical uses (Home Assistant, VPNs, etc.) a Pi is going to be way more than you need. It will sit there and silently and reliably run, for years at a time, powered by a USB cable. I know mine have.
Why would I consider replacing those with a bigger box, fan noise and a power brick? Maybe I'm missing something?
I think the reality is that a lot of people don’t use or care for the main Pi features like the GPIO pins or camera connectors, and aren’t using them in places where space or power is constrained. They’re purely looking for a cheap amount of compute.
For those use cases, it makes sense that performance per dollar is the top metric that many people look for. Pi used to be king there too, but as Pi prices have gone up and the price of other cheap compute has come down, it’s harder to justify the Pi. The Pi is still a great board, but a Pi5 with case, power supply, SD card will run you $100+ (throw in an SSD and you’re at $200+ easy), meanwhile an N100 PC can be had for the same price.
I'm wondering why there isn't a cheapest alternative to the pi without those features that most find useless. Or a more powerful for the same price
The N100 is the more-powerful-same-price option.
It's kind of like a keyboard without LED lights for status, pop up tabs for the angle, and a recessed guide for the USB cable. Would it technically be cheaper without those things? Sure... but it also costs next to nothing to add them. Once a keyboard gets to $10 almost all of the cost is already put into "being the best keyboard for $10".
There are certainly cheaper, <$100 TCO, alternatives to the Pi or N100 PCs... but they are accordingly worse performing. One thing Pi is really good at is having well supported options on the ultra low end. The Pi Zero 2W for $15 - it's crap performance but cheap. The Pi Pico microcontroller starts at $4 if you don't need a full OS (or just need to augment an existing box with GPIO over USB). If you're building a PC out of a Pi it's just not a differentiated option vs an actual PC is all.
You can still buy the earlier pi models. The zero is $10 and the B+ is $30 - both in stock at adafruit (the first supplier I happened to check)
N100 are often no fan and the power brick need not be much different than the one that feeds the Pi. At the end of the day the max wattages aren't much different.
The biggest thing the Pi line gets you is good GPIO and premade hats using it (with prepackaged OS images). If you're just running standard software like a home assistant or VPN then it doesn't make much sense. Doubly so on articles about OCing it.
I think a lot of people on the internet try to use the Pi as a home media server or file store and can’t envision it being used for anything else.
It’s versatile but really it has always sucked at those things. Yes, get a used N100 if that’s your goal.
The Pi shines in education and as an embedded device where power, size and maintainability are important. I spotted one the other day in one of those presses that crush pennies as a souvenir. It displayed an instructional video on a screen and also seemed to control the card reader device on the front. Why would you use a N100 in that context?
The Pi Zero 2W and Pi Pico make a ton of sense for those kinds of use cases. The Pi 5 (and 3/4) try too hard to be a PC to just fall short anyways.
2W, maybe. Pico, no. Driving a payment system and displaying an interactive video is still a job for a SBC not a microprocessor, even a modern one.
Raspberry Pi have found a price level and form factor that works for customers. The raw speed of the 5 is probably overkill for most user applications (the 4 is too IMO) - though edge AI and robotics need it, so there is demand. But ultimately the speed is there because a faster ARM chip fits within the cost-power-size requirements of the board.
Pico was meant to shine for much of the education half (e.g. micro-python controlled mini robots without the hassle of complex and poorly supported developer environments, OS management and configuration, or hardware complexity at a couple $ cost point), not the interactive media half (which would target the Zero 2W usage case over a full blown PC style Pi).
The compute module variants do tend to shine in the edge robotics niche where you need to do some processing in a custom form factor and don't want to design or source a SoC directly... but you don't really need the "rest of the PC" the standard Pi models provide. Pis (any model) are a bit of a shit fit for edge AI in general though. The memory bandwidth is god awful (even compared to cheaper SBCs) and the hardware to meaningfully offload matrix operations just isn't there like it is in a lot of other embedded boards fit for purpose. This leaves you spending more $ to use more Watts for less AI compute vs typical embedded AI options.
The Pi foundation's big break was targeting ultra low cost markets (i.e. sub $50 TCO) with a well supported device that has/had a lot of hacker community mindshare. Any product they've made beyond that focus (i.e. not the embedded or microcontroller variants) has been a pretty bad technology fit which ends up in more drawers than deployments despite the hype.
The mini underpowered PC market appears to be the only new market Intel is winning these days.
And they sold NUCs to ASUS. For mini PC servers I would always go intel for QVS transcoding. AMD is not on the same level as Intels media engine.
Intels iGPUs also support SR-IOV these days so you can get GPU acceleration to multiple VMs.
Streaming (local network - Sunshine/Moonlight) these days is really good. So good I consider moving my gaming machine to the garage and just stream it to a thin client or steam deck.
> But for the vast majority of typical uses (Home Assistant, VPNs, etc.) a Pi is going to be way more than you need.
Ah, but if a Pi is more than you need,
you wouldn't be reading an article about boosting performance with SDRAM tuning
or indeed caring about the RPi 5 when the RPi 3 was more than good enough.
Interesting, but why bother. Pi's were great when they were cheap, but one can get a better performing Intel N100 box for a similar price and power consumption now.
The only reason I can think of to use a Pi is GPIO.
As a person who uses both, I won't leave an N100 system where I can leave a RPi5.
First, the overall footprint of an RPi5 is way smaller, plus RPi is easier to cool passively than an N100, also it allows RPi5 to be lower maintenance on the longer run.
Most N100 are noisier, hotter and bigger than a Pi5 (I have a smaller N100, but it's still actively cooled).
When paired with a decent A2 card (e.g.: Kingston Canvas Go), the SD-card related lag disappears. Add a good external SSD, you have tons of solid state storage. At that point you can just forget it.
I like my N100, and companies like GMKTec routes 2 PCIe lanes to NVMe slots allowing ~1.5GB of throughput, but for most home use, that's overkill. I rather use Pi as a server and N100 as a desktop (which is exactly what I do).
> When paired with a decent A2 card (e.g.: Kingston Canvas Go), the SD-card related lag disappears. Add a good external SSD, you have tons of solid state storage. At that point you can just forget it.
At that point you can just underclock an N100 NUC for the same price and almost the same TDP (it's pretty much even with the SSD). Most vendors configure the CPU well into diminishing returns for marketing reasons and even a small undervolt can make them passive (the same is true of many laptops).
I get the form factor issue but exposed RPis without the enclosure (more $$) are a burned out board waiting to happen. I use them in enclosed hardware that needs gpio like 3d printers but the N100 wins every other time.
You can't rule out a fan sucking in dusty air from side or worse, bottom of the case. Most N100 systems have too much air circulation inside to forget them in the long term.
Many N100 systems come with 5W TDP as default. However, this is more of an "average" TDP. It doesn't prevent the CPU to enable its turbo and exhaust its thermal budget. Forcing it to lower envelopes via cpufreqd would work, but you're limited by your fan again, which will make everything dusty, eventually.
I have found a pretty cheap completely aluminum case [0] which sucks heat from every IC and bottom of the case via silicon thermal pads. It keeps the processor at 37 degrees, cools USB bridge and power controller too.
With no moving part, case is almost airtight. So, it's sitting snug and cool in that case. You can't burn RPi in that configuration.
[0]: https://www.amazon.com/GeeekPi-Aluminum-Raspberry-Cooling-Th...
We use fanless creditcard-sized N100 from upboard without issues although we haven't done undervolting yet.
Oh, I didn't see that form factor, can you share a link?
Oh, upboard. That thing starts from $229 as a box. Not exactly cheap, and 8GB/64GB version is $279.
[0]: https://up-shop.org/default/up-7000-edge-series.html
True, it's not as cheap but significantly faster with the same approx power usage
Even with printers, you can connect via usb if we're talking Klipper..
Duet 3 with daughter boards for tool switching and multiple extruders (customized Voron 2.4)
> Most N100 are noisier, hotter
I went from a passively cooled RPi5 to a N100. The RPi idled at around 50°C and throttled under load. The N100 keeps 56°C under light load (the heaviest things are rtl_433, grafana, nextcloud). To keep it from throttling in heavy load I zip tied a small Noctua NF-A4x20 to the side of the case over the passive heatsink and adjusted the fan curve to only run at above 60°C. This works super well, the N100 doesn't get hotter than about 62°C under full load and the fan spins slow enough that I can't hear it at all even if I put my ear directly next to it.
> To keep it from throttling in heavy load I zip tied a small Noctua NF-A4x20 to the side of the case over the passive heatsink and adjusted the fan curve to only run at above 60°C.
I applaud your creativity but surely this is an indication of the limitations of the architecture!
Nice solution, but with the case I linked, my Pi idles around 37, and reaches 50-ish (both degrees C) at most (if I load it hard), since the case is sucking heat from all chips and bottom of the PCB.
How does it handle electrostatic discharge when touching it in that case? I used a case that was really more of an oversized heatsink[1] than a case and it was really sensitive. Managed to crash the rpi a couple of times when I touched it without touching ground first a couple of times.
edit: I replaced the thermal sticker for thermal paste, not sure if that could have affected it, the paste wasn't the conductive kind.
[1]: https://www.amazon.com/GeeekPi-Aluminum-Heatsink-Raspberry-I...
The case I'm using is this [0], and my Pi never crashed while inside it, even though I bear handled it.
The room the Pi lives has a wool carpet, too.
[0]: https://www.amazon.com/GeeekPi-Aluminum-Raspberry-Cooling-Th...
> The N100 keeps 56°C under light load
Thats somewhat surprising. I have an old 14nm Celeron I use as a home server and under typical usage it is a few degrees above ambient. Checking now it is 26C. There is no fan, and it just has whatever tiny stock heatsink it came with.
I wonder why N100 is running so hot for you.
Is the N100 better supported by Linux than the RPi? My RPi gets the TAINT_CRAP flag due to broadcom drivers from the staging area.
That will depend on what other hardware they put on the board. In general though N100 is just a standard x86-64 so you have a lot better odds of good support.
It just works.
Plenty of Rockchip bargains out there but how long are they supported https://www.jeffgeerling.com/blog/2024/popular-rockchip-sbc-...
Meanwhile the original Pi model only just went EoL https://endoflife.date/raspberry-pi
The n100 is an intel board and all the parts used in the cheap nucs I have tried are supported on quite old mainline kernels.
At about twice the price for just the pi, I got a n100 machine (with a case and power supply and 500mb ssd and 16gb ram). Draws about 10 watts, is faster on every axis, and once you setup the pi with a case/psu it’s very close to the same price.
> Draws about 10 watts
Idling?
OK. EU prices seems to be at least two times higher for a N100 mini PC.
Energy consumption seems to be 2x-3x higher. The RPi5 idles at ~3.5w and the N100 at 10w. But yeah, the N100 is faster.
You need a power brick for the N100 box.
Depends on what you need. The N100 isn't always "better".
10 watts isn't idling, that's working. The chip has a TDP of 6 watts, but the RAM uses some too. I turn mine down to a lower clock speed, which takes off about 3w.
> You need a power brick for the N100 box.
Yes, but it was included (along with the case) for $106 USD, delivered.
Granted, that was on special, but AliExpress has sales that get them down to $115 USD pretty regularly.
I find it quite interesting that we're beyond the point of having an Intel processor being a massive advantage. I think 10 years ago most people would have jumped at an x86 processor with similar specs to the equivalent Raspberry Pi.
As someone who runs a lot of ARM stuff... I still find things which aren't packaged for ARM/linux with surprising frequency.
I have an M1 at work. Anything off the beaten path fills me with dread and / or poor performance because I end up emulating it.
Intel Edison was released 10 years ago.
In EU you add 40$ for delivery and 40$ for taxes/dues
ODROID-M1 with RK3568B2 is supported until 2036 at least https://forum.odroid.com/viewtopic.php?f=29&t=44218
That a distro's principal maintainer steps down is unfortunate for anyone who is downstream from them, but it is not the end of the world. As the article mentions, other distros for Rockchip continue to exist.
From the same forum:
> the official Linux kernel is more based on SoC vendor's BSP (Board Support Package) that can utilise full SoC features. And Hardkernel is able to consult with Rockchip in case we found missing SoC features only when the issue is discovered in their BSP, otherwise it's very hard. Therefore, we are managing the Linux kernel with most similar OS with the version from Rockchip, even Ubuntu 20.04 is not from Rockchip.
https://forum.odroid.com/viewtopic.php?f=212&t=47522
> What are not working: USB 3.0 host ports
> The images with upstream kernel would not be stable enough and features are not fully supported, so please do not use this image for your important project or to store important data.
https://forum.odroid.com/viewtopic.php?f=217&t=44462
>The only thing missing is a firmware capable of booting the operating system.
https://forum.odroid.com/viewtopic.php?p=374526#p374526
That was the situation in 2022 and early 2023.
Since late 2023, ODROID-M1 is supported by mainline u-boot (odroid-m1-rk3568 target) and you can also use a mainline kernel (6.4 or later) which provides working USB 3.0.
Is the GPU also supported when using a mainline kernel?
Yes, 3D acceleration support is in mainline Linux and Mesa using the panfrost driver.
I think VPU (video encode/decode) support is still partial and NPU (AI acceleration) requires an out of tree driver (rknpu2).
VPU support in general seems to me to be blocked on an update to the V4L2 API, there is some code in the staging directory.
Can you link a few that are in the Pi5 price range?
Pi is ~70 EUR including ~25% VAT.
Cheapest N100 I found in 5 minutes of searching was a ACEMAGIC T8 Plus for ~160 EUR. (Currently sold out)
They're only really comparable if you want an 8 GB Pi 5.
Right now I could buy a 4 GB Pi 5 setup (Pi 5, case, power supply) from CanaKit for $82. $97 if I didn't have an SD card to spare.
I don't buy mini PCs from random companies. The last N100 I bought was the NucBox G2 for $139:
https://www.amazon.com/dp/B0CCDL6VS3
If I didn't need the second ethernet port I probably would've grabbed the NucBox G5 for $136:
https://www.amazon.com/dp/B0CZRJL7JM
Cheapest G5 in Europe I could find was 180 EUR, but sold out. And on amazon.de it's 280 EUR.
GMKtec Mini PC N97 [1] is 160 EUR on amazon.de (with the 70 EURO voucher that is). Got mine with free shipping to Romania, as the purchase was over 50 EUR.
Also, the N97 has better performance than the N100, but the power consumption is higher. For me its a plus as it's more powerful than the usual N100 mini PC's while much smaller and fan noise is quite low (for now at least...). Not really intended for high workloads and 24/7 anyway. More like a backpack PC.
[1] https://www.amazon.de/-/en/G5-Computer-Generation-12GB-256GB...
For me it shows as €233.82
Prices in the US dont show sales tax which is somewhere between 0 and ~10% (or a little higher).
Presumably your prices have VAT?
Oh yeah. I have no idea what's available there. Sorry!
GMKtec sells on AliExpress if you don't mind going that route:
https://www.aliexpress.us/item/3256806972422176.html
The price for those is now $200 and $170.
That's weird. The prices haven't changed for me. I wonder if it's because of "Exclusive Prime price."
It doesn't look like the AliExpress prices have changed if you don't mind going that route.
Over at AliExpress there are tons of options, like this[1] or this[2] fun 4x M.2 box.
Yes they're about twice the price of a bare Pi5, but they come with metal cases and power supplies, can fit an M.2 NVME without add-ons, and like the first example also includes some RAM and storage.
While likely not the best quality RAM and storage, what I got in mine has held up to light use for several months so far without issues, so better than nothing. YMMV.
I bought a few 2GB RPi4's when they were $35. At that price point they were great. I don't get the RPi5, especially 4GB+ variants.
[1]: https://www.aliexpress.com/item/1005007511663921.html
[2]: https://www.aliexpress.com/item/1005007900081466.html
I haven't got any personal experience of them but aren't there a number of Raspberry Pi clones which offer the same functionality at much lower cost? Seems like that would open the gap up.
The issue with nearly all Raspberry Pi clones is documentation. Boards tend to be undocumented, sparsely documented, or documented with errors. It makes for a frustrating experience when you run into problems, especially when those problems are with bog-standard Raspbian (or whatever). Raspberry Pi hardware has occasional quirks (eg, don’t try to hotplug USB on a Raspberry Pi Zero), but quirks are the norm for other SBCs (anything from Pine64 is an exercise in frustration). So the modest upcharge for a Raspberry Pi is usually worth it.
There are some, but a lot of them have very spotty Linux support. In most cases either you're stuck with some ancient kernel provided by the manufacturer, or you're relying on community support from Armbian at best.
Take for example the OrangePi Zero 3[1], which you can get[2] from around $25.
Armbian provides community support status[3], so you're relying on volunteer efforts to provide updates and support. Also drivers for peripheral devices might be missing[4].
There are others, like say the NanoPi Zero2[5][6] for around $20, but again, support might vary. The Zero2 is so new it's not listed on Armbian, and most of the other NanoPi's are under community support.
That said, I've been running some NanoPi ZeroPi's (the Zero2's predecessor) for several years now, without any issues. Just using them for light loads, one is my PiHole for example, but yeah can work great as long as you do a bit of research in terms of what's supported and not.
[1]: https://www.cnx-software.com/2023/07/03/orange-pi-zero-3-all...
[2]: https://www.aliexpress.com/item/1005007673457198.html
[3]: https://www.armbian.com/orange-pi-zero-3/
[4]: https://linux-sunxi.org/H616
[5]: https://www.cnx-software.com/2024/09/13/nanopi-zero2-tiny-he...
[6]: https://www.friendlyelec.com/index.php?route=product/product...
[1] is 150 EUR from EU, not bad. I can see myself give it a go at some point. Still twice the price for my needs.
Right, it's twice the price but like 10x the capability (unless you need GPIO).
That's why I think the lower-end, lower-priced RPi4 made more sense. Fast enough to have decent network speed and run a lot of functional workloads, cheap enough that you could "spam" them.
I use it in a 3d printer. I don't need 10x the capabilities. For my usecase it's just 70 EUR more, with the only benefit it might boot and update faster.
And then I have to add 25% VAT because the order is more than 1500 SEK.
No, I think that was the price from the Spain warehouse and included tax. Just 10 EUR shipping. Aliexpress got better with showing you the real prices.
> Pi is ~70 EUR including ~25% VAT.
Now add a power supply and case.
What if you don't need a box or power supply? And what if you like then gpio pins?
FT232H breakout board is a couple of bucks.
Thanks for that!
Thank you.
No one ever talks about just how easy it is to add GPIO to anything.
No, but thank you.
I put them into a 3D printer that already has all that.
But I need to add a 16 or 32GB SD card. Around 10 EUR, maybe less.
Is it a home-built 3D printer?
Look for ASUS Alder Lake-N, roughly 100eur incl vat. (that's mobo with N100)
You do need to buy SODIMM DDR4 (12-14eur for 8gb), m2 NVME for storage and power supply (which you have to buy for rpi5 also).
You end up with a board with way better ports and IO, better audio dac, and significantly faster performance all across the board.
Raspberry PIs 5 and such offer poor value and are very overpriced.
Raspberry Pi Zero 2 W, Orange Pi Zero 3 and RP Pico still provide good value. (for 3d printers and such)
N100 is inexpensive, has impressive performance, and can support 16+ GB RAM.
That said, I do also use an R.Pi 4 and an R.Pi 400.
The significant advantages of the R.Pi 4 and R.Pi 5 are
= R.Pi is a learning-first platform
= current Debian supported "out-of-the-box" with a custom D.E. and Wayland
= excellent community for software and hardware
= the R.Pi 400 (and the expected R.Pi 500) "keyboard computer" are unique in the marketplace, with all the community and support
= R.Pi is not an X86 platform, which may be favourable depending on your project requirements
> current Debian supported "out-of-the-box" with a custom D.E. and Wayland
I’m not sure I understand the point, because Debian supports N100 PCs natively, too. Or is a custom DE the advantage?
> Or is a custom DE the advantage
Yes, as a learning-first (teaching) platform, the custom DE is an advantage.
But further, the R.Pi does not need to be an over-sized solution or lead a benchmark with a top score. It just needs to be fast enough for learning and hacking with GPIO.
A computer doesn't need to have a gigahertz burst mode just to run an efficient little OS for coding.
As for Wayland, it only recently became the default WM. But it is nice to have.
> A computer doesn't need to have a gigahertz burst mode just to run an efficient little OS for coding.
If efficiency is your goal, you should prefer the N100 which is built on a 7nm class process compared to the Pi 5's 16nm.
> N100 is inexpensive, has impressive performance, and can support 16+ GB RAM.
Yep runs 32 GB just fine, even if not advertised. Also great Software support as usual on the PC Platform. Also more I/O, esp. PCIe than a Pi and you actually can buy boards with standard connectors. Also at least for server type use cases x86 is most of the time probably still a lot less painful than ARM.
> for server type use cases x86 is most of the time probably still a lot less painful than ARM.
This isn't true if Debian is supported.
As for the architecture per se, we should consider what Apple has accomplished with its switch to ARM.
With the appropriate license, there is no particular reason Apple's team couldn't have built a good x86 processor. Switching to ARM was a matter of necessity, not due to a superior architecture.
Architecture largely doesn't matter these days, just how you implement it and what fabrication you build it with.
> = current Debian supported "out-of-the-box" with a custom D.E. and Wayland
Effectively the same thing for basic desktop use, disregarding wayland, as doing a basic bare bones debian x86-64 CLI only install (totaling something like 1.2GB disk space used) and then "sudo apt install xorg xfce4 xfce4-goodies libreoffice gimp thunderbird firefox-esr gthumb nautilus konqueror", or whatever other list of applications one wants to install along with xorg and the desktop environment.
When you're running off a battery and solar panels, a couple Wh makes a big difference.
GPIO - I have a small "robot" running off of a Pi5, this will speed up compile times.
You mean a refurbished Intel box.
That's not a fair price comparison.
Why buy Intel N100 with Intel Management Engine (which is disabled on US Government computers)? I rather buy RaspberryPi where I don't need to disable Intel ME by myself...
You just have the undocumented VideoCore which also happens to boot the system and have full access to RAM at all times...
I’m of the same opinion now. It’s so much more pleasure to work on an n100 vs rpi4/5 too. Everything is snappy and responsive, apt-update fast
The GPIO was always the point. Pis were supposed to be educational tools. People using them as cheap computers ruined it.
>The GPIO was always the point.
You ca get GPIO on the cheap from a Arduino, ESP8/32, or RP2040 microcontroller. RPi itself is far too wasteful for that.
> People using them as cheap computers ruined it.
False, industrial users using them in all commercial embedded products ruined it.
Almost nobody is using RPis as cheap computers at home since they suck at that: too expensive, too underpowered and too limited in functionality to be a computer for the average joe
My RPi 2 becomes an unusable stuttery mess when I open a web browser and for the price of a RPi 5 I got a used quad core Intel 8th gen laptop with 8GB of RAM and 256GB SSD which not only beats the RPi at any and all computer tasks, but it also comes with built in keyboard, 1080p monitor, battery, charger and I can take with me to do work on the go. The only thing I see the RPi good at at home is a headless PiHole/VPN or whatever, not as an actual desktop computer replacement.
A Raspberry Pi was a full Linux computer with GPIOs. You can run whatever daemons and still control a relay board. No microcontroller fills that niche. There still isn't really anything in that niche other than single-board computers (i.e. Raspberry Pi clones) - for example, NUCs don't have GPIOs on them.
Of course you can fall back to the previous standard idea - connecting an Arduino to another computer and communicating by the serial port.
They were cheap computers, so people used them as cheap computers, I don't understand your point at all. For a quite a long time they were also the de-facto standard way of making a LED blink over the internet without having to dive into the embedded space.
> People using them as cheap computers ruined it.
IMO it is the people using them as embedded controllers for their products that ruined it, in combination with the poor decision by the Raspberry Pi folks to prioritize commercial demand for the products over hobbyist demand during the pandemic. We hobbyists "made do" with other solutions in that time, and now there's no real need to go back.
Wouldn't say it ruined it. People used them as cheap computers because at the time there barely were good alternatives for that cheap of a computer.
Ruined precisely what?
The price.
Because I got a handful of them when they were cheap? 30% free faster computing? I'll take it anytime
Because people like to tinker
A Rasberry Pi 5 is no match for a tiny-mini-micro 1 litre PC at nearly the same idle power draw
good documentation
Ecosystem of HATS
GPIO
camera connector (better than most usb cameras)
Good luck writing software for one of those N100 boxes only to find out they're sold out and you now need to extend the code for new model over and over.
N100 is standard x86. Write generic code that runs on a generic linux kernel and you won't have to worry about if this is a N100, N200, AMD, or even something weird like a 30 year old RS6000. Well if you need GPIO you have to worry a little, but you are forced to spend an extra day to support a generic GPIO framework instead of directly writing to PI pins - that extra effort means you don't care.
Even in embedded where you have contracts to ensure the same hardware will be available for a decade everyone has moved that way because eventually the hardware you rely on goes out of production and so you need to support new hardware.
As a desktop, one can get a used x86-64 dell ultra small form factor quad core Intel core i5 something, 16GB RAM, something 128 to 256GB SSD on eBay for a hundred bucks... It'll have integrated video capable of driving one 4k at 60Hz as native displayport. And a decent amount of other gigabit Ethernet, usb3, analog in and out audio, etc.
Yeah it'll be 7 years old but it'll still be much faster than a rpi5.
Even a mediocre Intel 120GB sata3 ssd from 8 years ago will run circles around a microsdxc card in disk performance.
These mostly come from corporate lease environments and have very little write wear on the ssd and are very clean.
The raspberry pi is more useful if you need the gpio pins or want something really small and perfectly silent for some embedded application.
> Even a mediocre Intel 120GB sata3 ssd from 8 years ago will run circles around a microsdxc card in disk performance.
You don’t need a microSDXC card for the pi, you can boot it off an SSD, which is what my old PI4 does to run Proxmox Backup Server.
From a sata3 ssd in a usb3 enclosure? adds cost and complexity and cabling... I understand there's also a raspberry pi5 sata "hat" which is something like $20 and then figure at least $35 to buy a decent quality small sata3 ssd to add to it.
Sure, but then the DELL will already be several times bigger anyway.
[dead]
Just tried the SDRAM tweak and got myself a respectable 10% speedup with LLMs. Hat tip.
What are you doing with local llms on a pi5, genuine question btw.
Not the author, but on a Pi 5 small LLMs such as tinyllama or qwen2.5:0.5b run at over 25 tokens per second (haven't tested the performance boost with it yet). It could be useful if you wanted a local assistant or an AI server for devices you don't want to run LLMs on
Just curious, what can a 0.5B model be used for?
I use mine mostly as a context-free code copilot. It's not perfect, but it knows how to write a template and for most tasks that's all I need.
Are there any vision models that will work on the Pi 5? Something similar to minicpm-v?
I don't believe you can get any useable results. Perhaps only for R&D.
Is there some kind of list that shows comparable performance over all of these changes? There's that ARM 4_0_X_X quant, NUMA patch, now this SDRAM tweak. Plus if there's any delta with your llamafile speedups vs. openblas. I feel like we should be approaching over 3 tok/s tg for 7/8B models at 4 bits by now?
Last I tested, the ARM quants were actually slower on a Pi 5 than Q4KS, but I later learned I messed up some compile flags and haven't had the chance to retest lol.
Points of further interest:
= TFA also mentions that the R.Pi 4 can be tweaked [1]
= an R.Pi 500 is in development [2]
[1] _ "The tweaks can also give the Pi 4 a boost, but the Pi 5 improves more dramatically"
[2] _ "Pi 500 Geekbench results starting in September—all seemingly with these tweaks in place already!"
>Changes may roll out in a firmware update for all Pi 5 and Pi 4 users soon.
Where can I sign up to not miss the new official firmware release?
Edit: well, I'll just do the firmware upgrade now, why wait...
Hacker News doesn't have a RemindMeBot... ;-)
you can subscribe to releases on their github page
Good to read they're doing this for the older CM4 too!
> You can actually monitor the temperature of the sdram and it reports if refresh at half or quarter the rate can be done. That allows the overhead due to refresh to be reduced by a half or a quarter which does improve benchmark results.
I wonder why this is not available on PCs. tREFI tuning has a big impact on performance, and having an automated guided optimization based on temperature would be nice.
> That allows the overhead due to refresh to be reduced by a half or a quarter which does improve benchmark results.
Does this increase susceptibility to rowhammer?
Wow... I haven't heard of someone using refresh timing to speed up a system outside of retro hardware in ages. I guess I'd just assumed that after we'd transitioned to DDR that refresh cycles were just such a small part of overall cycle time that it didn't matter. Goes to show what I know!
Raspberry pi 4 has NUMA? The same NUMA thing large servers with 1Tb of RAM have?
It looks like it does https://github.com/raspberrypi/firmware/issues/1854#issuecom...
Faking NUMA can help enforcing access patterns that make things faster:
> In the worst case, two sdram clients accessing different pages of the same bank, may cause repeated closing and opening of the pages, harming the usable bandwidth you can achieve. Which pages belong to a bank depends on the physical address lines of the buffers used. We have found that it is pretty common for buffers to be allocated in pathologically bad ways.
> NUMA allows us to have more control over this. We can split our sdram, into, say, 8 NUMA regions, and configure the kernel to interleave the allocations between the regions. In addition the sdram controller allows the address bits used for segmenting the banks to be reconfigured (which we've exposed through eeprom config SDRAM_BANKLOW).
https://forums.raspberrypi.com/viewtopic.php?t=378276
If it works on a Pi, wouldn't this make sense to also implement with DDR5? Or is it already there? Now that the chips are split into 2x mux of 4x32bit busses with dual rank on dual channel, it would really suck if the data was defragmented in such a way that only one 32 bit bus could access it and only half the time (if my understanding of this is right anyway).
Exactly what I was thinking too.
I'd like to see the numbers around the power usage added too, I see the Pis and audrinos personally as a low wattage device that I don't need to worry about the cost, I'm sure this overcloclocking isn't too much but would be nice to see to weight up viability
Jeff has a N100 vs. Raspberry video where he compares power consumption.
I think he used a GMKTec G3, and the numbers were not bad, but active cooling adds a constant whine to background. Some companies' fans are more whiny than others, too.
N100 is not a "cool" chip by any passive cooling standard, esp. under load. However, its standard (software limited) TDP is around 5W IIRC. You can increase that limit for more speed and added noise.
You can still passively cool it if you want to.
Which limits its performance severely. Even with active cooling it can reach 60 degrees C without much effort.
On the other hand RPi5 idles around 37 and reaches 50 degrees C while passively cooled with a heat-sink case.
>Which limits its performance severely.
Not if you put a large fin block on it.
If I'm going to remove the case and do my own modding, then all best are off. I want something working well out of the box. I have enough computers to play, and not enough space to put Frankenstein systems securely around the house.
Fair enough, but the RPI 5 also won't reach max performance out of the box until you add a third party cooling case on it, be it active or passive. N100 systems at least come with a case already on out of the box that let them reach max performance albeit with some noise.
And we're talking about tinkerers after all, which is the target market for the RPI. Parents won't be buying RPis en-masse for their kids from Walmart to use as their main computer for school instead of Macs, iPads, Chromebooks or Windows laptops.
Average Joes who aren't into tinkering, aren't daily driving a RPI as their main computer. It's a DIY toy/tool for those who want to tinker, learn and make stuff, so some added effort is implied anyway.
Yes, that's sold as a DIY computer, so getting it a case is a must. OTOH, I'll be honest and say, both the official cooler and official case for RPi allows it to run at full-steam, almost all the time.
I especially bought that passive case for my Pi, because it's destined to be a "tuck away server" from get go, and it's trucking along just fine. I'm adding services on top of it slowly, and it's not showing any signs of fatigue.
A2 card and faster SD support really makes that one shine, without adding hats on top of hats. Again, OrgangePi 5B has better hardware, but way worse support.
Yeah GMKTec is pretty bad when it comes to cooling solutions. Literally just had the K9 with the 125H delivered yesterday and while I was expecting the fan setup to be meh, it's honestly like a bad case of tinnitus. Constant eeeeeeee at the most annoying frequencies possible.
They put a tiny 40mm fan on top which resonates with the plastic box and screeches like an absolute banshee all_the_time. That one is a must-remove since it also blocks installation of any heatsinks on the nvme/ram, and a mesh top results in roughly the same temps. But they gotta put their logo on top, priorities.
The second problem is the main cpu cooling, which is actually a really well made bottom intake blower side exhaust setup... but it's set up to run at like half speed even when you're at 30 degrees and there's no bios fan curve control for it except for extreme temp steps. I'm currently pondering adding a pico or attiny that reads the board PWM and retransmitts it so it only runs the fan when it's set above usual. Or maybe just a full passive copper block replacement, but that would require a case redesign...
Otherwise it's a really capable little machine if they put a little effort into making it shut the fuck up at idle.
G3's fan is mostly silent. You might hear a small clicking on lower RPMs, and mostly a whoosh on higher RPMS, OTOH, it' still moving air and sucking dust from the bottom.
Looks like K9 needs to move more air since it's packaing an U5 in that small box, so cooling them silently is much harder.
Iirc the N100 idles at 6W and boosts to 25W, whereas the 125H idles at 35W and boosts to over 100W, which the cooler has to be designed for of course. The Ryzen boxes have similar cooling problems from what I hear.
But if you run the N100 flat out, which is likely to happen given that it's not the fastest thing, and disable turbo on this one (which I've done for now since my use case is a low power server) they are actually surprisingly close in heat output. The Radxa X4 tried to cool the N100 like a Pi 5 and it almost melts.
Find 30% is quite a feat
Really interesting, thanks for sharing!
This is really interesting, thank you for sharing this
[flagged]