non-x86 open-source-hardware laptops, let's take stock of Jul 2025

[brucehoult]

The only problem I have with it myself is that it's limited to 8 GB RAM which make e.g. building current GCC challenging unless you restrict the build to '-j1' which is a shame on a quad core board. Older GCC (e.g. 9) is fine. Current GCC is *just* ok to build on a 16 GB RAM machine with '-j4' or even '-j8' if you have swap enabled -- it uses a few hundred MB, but not enough to slow it down.

Don't talk to me about it, they seem to have gone crazy with gcc!
It takes forever to compile unless you have modern machines.

You can't have more than 8GB of ram on machines like the HP-PARISC C36xx C37xx, which only have 1 core!
The C8000 can have to physical CPUs, so 2 cores, and uses non-proprietary ram, but at most it supports 16GB!

Machines like PowerMac G4s, even [email protected] (the fastest), are doomed as they don't physically address more than 2GB.

Of course you're going to have a problem with any 32 bit machine. All the Linux distros were dropping them already five or six years ago, largely because of the difficulty of building more and more packages on them -- as well as few people using them now.

[Nominal Animal]

My 24 core 5.4 GHz Core i9 laptop just feels like magic, and is very spoiling.

From the cheaper side of laptop world, the Ryzen 7 7730U (8 cores, 16 threads) I recently bought with 16G RAM and 512G (Samsung PM9C1, MZVL8512HELU-00BH1) feels that way to me too.  All my normal activities are absolutely instant, even booting is so fast I only need to type in my Grub/LUKS password, reach for my cuppa, take a sip, put the cup back, and it's ready for my username and password.  (All my complaints about HP Laptop 15-fc0xxx are physical; like the keyboard could be nicer, the chassis more rigid, cooling fan quieter, display backlight a bit more uniform; all explained by the affordable price.)

I have noticed that the storage bandwith is absolutely critical to me.  In 2008 – 2015 or so, I mainly used my desktop machine with mixed RAID0/1 setup with spinny rust, and an Acer Aspire One minilaptop/netbook for a carry-along.  (I still have it in working condition in storage.)  The storage was a slow 2.5" HDD, but moving my browser profile to a limited-size RAMdisk made all the difference.  (I automated all that, and discussed the details and related bugs at gnome display manager mailing list.  Instead of fixing GDM, it was easier to interpose /usr/bin/firefox, so that first invocation (when no processes with the binary name are running as the current user), the profile is unpacked from storage to the ramdisk; and when the last instance has exited, the profile is packed back to storage.  It's absolutely transparent to the user, but made Firefox really responsive, even with such a feeble processor.)

Similarly, the physical chassis is extremely important to me.  I'd be willing to trade quite some weight for a full-aluminium underside, especially if also used as a heatsink; that requires that the processor and other getting-hot-bits are on the underside of the board, so they can be in direct contact with the chassis/heatsink.  There are a few SBCs like this already, and they are much easier to passively/quietly cool than the stupid "standard" option of a fan embedded in the heatsink, but their peripherals tend to not be that well suited for a laptop.

Multiple keyboard layouts is a plus, as is availability of replacements.  (For whatever reason, I tend to go through keyboards... although for full-size external keyboards, taking them apart and cleaning them every six months or so, made them survive for years.  Laptop keyboards just don't have the tolerances, longevity, cleanability I need.)

This all tracks back to non-x86 laptops.  I know for a fact that there are many ARM-based SBCs I'd be happy to use in a laptop, and I suspect the aforementioned RISC-V's would be too.  The M.2 M-key PCIe NVMe is somewhat critical (number of lanes – I'd much rather have one four-lane one than two two-lane ones, or a two-lane one rather than two one-lane ones).  For example, for Odroid M1S and Odroid M2, HardKernel dropped it from M1's two PCIe 2.1 lanes to just one; this hugely dropped their practical value for me.  It is also definitely worth it to spend a bit more on an SSD with excellent random 4K read performance; you rarely need the maximum bandwidth, but you can easily be limited by maximum IOPs.

Processor speed does make a huge difference in startup latencies and how snappy everything feels, but all that can be compensated a lot by good systems integration, cutting off the porky bits.  Gnome and KDE desktops might feel sluggish, but switch to XFCE or LMDE (or to a lesser degree, to Cinnamon or MATE), and it'll feel much better.  And adjusting desktop environment and application defaults to conform to your workflow and preferences instead of the default opposite makes even more of a difference.  For me, the machine becomes as much an extension of my hands as a chisel or screwdriver or wrench does: I don't need to think about the tool at all, I just do what I want.  This helps me, because my cognitive power is very limited, and I can fully focus on what interests and benefits me.

You know, like when typing: you too probably don't need to see your fingers or the keyboard to type, you just think of the things you want written, and your subconscious/motor memory takes over the mechanical part of typing.  You only note if you make a typo, right?

If I were to build my own non-x86 laptop, I'd start by collecting information on keyboard, battery, display alternatives, focusing on easily available known-good OEM parts.  Then, I'd design the base chassis, assuming full-aluminium underside at least doing double work as a heatsink, then see what kind of space is left over for the motherboard.  While the selections themselves aren't that important –– they'd just reflect my personal preferences –– it is all that information that would be useful for anyone else building their own.  For example, are there multiple keyboard modules that can be used with the same chassis?  Perhaps just replacing the upper (keyboard) plastic surface?  Or with an insert of some kind?  Could a single motherboard support more than one type of battery, by just having more than one connector (obviously only one in use at the time), or perhaps via a "jumper" adapter?

Because the design of the motherboard is way outside my current skill range, I haven't even doodled any designs yet myself.

[DiTBho]

do you know how much it takes from "load kernel and exec" to /sbin/init, and then to /bin/login via agetty?

On Zaurus-SL/C1K, Standard Gentoo/OpenRC ... 17 bloody minutes
On rb532a, Standard Gentoo/OpenRC .... 15 bloody minutes
On Apple Pismo, standard Gentoo/OpenRC .... 7 bloody minutes
On Apple MacMini/x86-2009, standard Gentoo/OpenRC ... 55 seconds

It's so bloody slow because there are many interactions with Python and bash-scripts

Do you know how much it takes with my superfast booting process - all in one - all written in C/89 ?

On Zaurus/PXA270@400Mhz... 39 seconds
On rb532/mips32@400Mhz ....  34 seconds
On Apple Pismo/G3@400Mhz .... 10 seconds
On Apple MacMini/x86-2009 ... 2 seconds

So, it's also a matter of *what* you run 

[brucehoult]

My 24 core 5.4 GHz Core i9 laptop just feels like magic, and is very spoiling.

From the cheaper side of laptop world, the Ryzen 7 7730U (8 cores, 16 threads) I recently bought with 16G RAM and 512G

Not THAT much cheaper ... you paid 700€ for 16GB/512GB, I paid (Feb 2024) ~1400€ for Lenovo Legion Pro 5i (2023) 32GB/iTB and WQXGA (2560 x 1600) screen, i9-13900HX (24 cores, 32 threads), RTX 4060.

I have noticed that the storage bandwith is absolutely critical to me.

It's really not to me on my SBCs. Even on the ones that have eMMC or M.2 PCIe SSD capability I usually just run from SD card. It's much more convenient if you want to try out different OS images. And as long as they have "enough" RAM, it's only a boot time / first run problem -- once everything is cached by Linux it's the same speed no matter what storage you have, at least for software development.

You need to get gcc and the project and library headers into the RAM cache, but after that even a 30 MB/s SD card can easily keep 24 cores running gcc fed with C source code, let alone 4 or 8 Arm or RISC-V cores.

Different story if I was doing video editing or even maybe image editing, but I'm not.

Even scp'ing big things between machines only needs about 20 MB/s if both machines are on the 5GHz WiFi -- I just tried something between my M1 Mac and that i9 laptop to check, and both of those have very fast SSDs (multiple GB/s)

[Nominal Animal]

My 24 core 5.4 GHz Core i9 laptop just feels like magic, and is very spoiling.

From the cheaper side of laptop world, the Ryzen 7 7730U (8 cores, 16 threads) I recently bought with 16G RAM and 512G

Not THAT much cheaper ... you paid 700€ for 16GB/512GB, I paid (Feb 2024) ~1400€ for Lenovo Legion Pro 5i (2023) 32GB/iTB and WQXGA (2560 x 1600) screen, i9-13900HX (24 cores, 32 threads), RTX 4060.

Not a big difference in absolute terms, but a big difference (50% / 2×) in relative terms!

I have noticed that the storage bandwith is absolutely critical to me.

It's really not to me on my SBCs. Even on the ones that have eMMC or M.2 PCIe SSD capability I usually just run from SD card. It's much more convenient if you want to try out different OS images. And as long as they have "enough" RAM, it's only a boot time / first run problem -- once everything is cached by Linux it's the same speed no matter what storage you have, at least for software development.

For testing, yes: the difference is that first run.  For an everyday carry, the initial latency gets old quick.  It definitely did for me, when using just the minilappy on vacations up north.  (I only paid 200€ for it back then, new with warranties, so it was a nice stress-free travel carry: even if it got stolen or broken, no biggie at all.)

For desktop GUI stuff, I tend to have a browser, Gimp, Inkscape, and a couple of terminals running all sorts of stuff, with typical workflows using about 8G of RAM, the rest used for caching.  16G shows clear benefits of caching; typically about half is used for caching.  I do not currently have any device dedicated for compiling and such stuff, and currently run any large builds in the background (starting the build from a nice'd and ionice'd shell).  If I were to build the kernel often, 32G of RAM would help with that.

For non-GUI stuff like servers and appliances and compile farms and such, and for non-portable GUI stuff like media players, we do already have plenty of SBCs to choose from!

Some Chromebooks with ARM A78/A76/A55 cores with fullHD resolution displays are sold here for around 300€ or so (Lenovo Ideapad Slim 3 Chromebook 14" 82XJ000YMX, Lenovo IdeaPad Duet 3 Chromebook 11" hybrid tablet 83HH0009MX) – note that we pay 25.5% VAT here.  I haven't touched or seen these in real life, so have no idea how they feel, nor how well actual Linux runs on these, but it is interesting hardware.  They're not OSHW, obviously.

Even scp'ing big things between machines only needs about 20 MB/s if both machines are on the 5GHz WiFi -- I just tried something between my M1 Mac and that i9 laptop to check, and both of those have very fast SSDs (multiple GB/s)

That's why I prefer wired GbE for large data transfers.  90 MBytes/sec is definitely achievable, and has been for a very long time even with cheap SBCs supporting SATA (like my Odroid-HC1, running Samsung Exynos 5422 octa-core big.LITTLE 4×Cortex-A15 2GHz + 4×Cortex-A7 1.3GHz SoC, a dedicated-SATA version of Odroid XU4).

I also like to have Cat6a or better S/FTP (shielded with braid and foil) cables in different lengths nearby (hanging inside a closet door, large figure 8 loops).  Not because my home environment is that noisy wrt. EMI, but because the shielded cables can handle walking or rolling over them with my chair much better.  Nowadays it's the small plastic lock tab that tends to break before the cable gets damaged enough to affect transfer rates (due to retransmissions or dropped packets).

[DiTBho]

What do you use for caching?
Tmpfs? Anything else?

[brucehoult]

What do you use for caching?
Tmpfs? Anything else?

Just what the kernel automatically has built in.

Here's my Milk-V Megrez (my current main workhorse, taking over from VisionFive 2 for the previous 2 years). Some GCC 15 build happened during that uptime.

Barely any swap used, but if I don't enable swap at all then the gcc build fails with -j4 when it tries to link four big programs at the same time. LLVM has a specific build setting to limit the number of parallel link jobs, distinct from the general level of parallelism, but gcc doesn't that I know of.

[Nominal Animal]

Ditto.  With desktop stuff, using tmpfs for /tmp and /var/tmp helps, but otherwise I too leave caching to the kernel.

[DiTBho]

A young dude, Byran Ee, made his Own ARM laptop

Blog, see here
GitHub, see here


Byran Ee, intervied by LTT

So, no interest/any comment for this project? 

[DiTBho]

Talking about "ARM-laptop", I found something amazing: The Acorn-A4 RiscPC laptop!

I've never seen one in person, but it's amazing that there's a laptop designed to run RiscOS (v3)!

[DiTBho]

It looks like there's a "kickstarter" trying to make a laptop based on the RPI CM5. See here.

Dunno 

[DiTBho]

[DiTBho]

Any comment?

[brucehoult]

Any comment?

Not without price, battery life, knowing the quality of the screen and keyboard and trackpad.

Also, pity it's the Pi chip not RK3588S.

[JPortici]

pity it's the Pi chip not RK3588S.

Yup. Anything PI in it is an automatic pass from me.

This thread had made me thinking.. I have a 2012 macbook i keep repairing, last year i finally installed debian on it and literally gave it a second life. Even 5-6 years ago it was borderline useless for how it was slow at everything, but it was actually the OS. Same applications now are usable, even productive... pity the current state of UI don't like anything below full HD (and the scaling factor in KDE can only do so much.)

However at some point i will have to upgrade. If only apple wasn't this hostile to the community i would have already bought an Air with Mx. But i have to install debian on it, period. OSX has been going downhill since after snow leopard and i really find things to be much more difficult than they should if you want to go outside the walled garden (compiling not using apple's compilers, using the terminal at all, every update having to undo all new UI features i didn't want)

But i'd like my next laptop not to be intel-based... because i already have a PC and the laptop is a secondary device, it doesn't *need* to be powerful. I already have a PC or a Switch for games, i can do some light programming, or browsing, on the laptop. nothing special.

Ideally it would be awesome if the framework had an option for ARM (or RISC-V. Yes, i know there is a RISC-V motherboard but it's not exactly consumer ready, or on par with say a RK3588 or simillar in performance. I don't know if there is a RISC V chip that can provide a smooth desktop experience right now, but it's only a matter of time)
Let's see what happens in the next year or two. Hopefully there will be more diversity when i'll decide to make the upgrade

[brucehoult]

Ideally it would be awesome if the framework had an option for ARM (or RISC-V. Yes, i know there is a RISC-V motherboard but it's not exactly consumer ready, or on par with say a RK3588 or simillar in performance. I don't know if there is a RISC V chip that can provide a smooth desktop experience right now, but it's only a matter of time)
Let's see what happens in the next year or two. Hopefully there will be more diversity when i'll decide to make the upgrade

Sure, the current RISC-V Frame 13 main board isn't very fast -- something like a late Pentium III or PowerPC G4, but quad core, but without an Altivec/SSE equivalent -- but it's only a shake-down product with more coming.

There is already the DC-ROMA "RISC-V AI PC" main board announced with 8x SiFive P550 running at 2.0 GHz. That's announced to be shipping in Q3 2025 which is ... well, nowish. That's the same CPU cores as in the HiFive Premier or Milk-V Megrez (which I have), but twice as many and clocked a bit higher. It's still lacking SIMD, but has a TPU so idk. The P550 is a bit better than Arm A72 (other than no SIMD) and I think also similar to Core 2 at similar speeds. It's hard to find the same benchmark run on Core 2 in the late 2000s and also on e.g. Pi 4 in the early 2020s, but it seems the A72 is better in Coremark and PassMark (better core pipeline) but the Core 2 wins in SPEC due to larger caches.

If you think of the new DC-Roma board as somewhere around a 1.83 GHz to 2.26 GHz Mac Mini or MacBook but with 4x as many cores ... you'll be close.

Tenstorrent, which is headed by Jim Keller [1] and has a couple of key members of Apple's M1 team, is intending to tape-out their 8-wide M1-class TT-Ascalon CPU. The first chips may be only 2.5 GHz vs the M1's 3.2 GHz but with very similar performance per GHz. Keller has stated that they want to see as wide distribution of Ascalon as possible, explicitly including in laptops. I expect we'll see hardware available to buy sometime in 2026.

[1] helped design the last of the VAXes, Alpha, AMD Athlon and Athlon 64 (was responsible for x86_64 instruction set), Chief Architect of PA Semi doing high performance PowerPCs then I think he moved to Apple followed later by Apple acquiring PA Semi who then turned to designing Apple's first Arm-ISA cores, including the first Arm64 chip anywhere. Then Keller was back at AMD and is the father of Zen after they were floundering a bit with "Bulldozer" and "Piledriver". Then he was at Intel in 2018 and drove the 12th gen strategy with P and E cores, and also getting Intel to outsource production. It is thought his resignation in 2020 was over differences about outsourcing. In between, Keller was at Tesla designing the FSD hardware. As with Elon, he is either a god, or the world's luckiest bullshitter. It doesn't actually matter which as long as both of them keep being in the right place at the right time.

[JPortici]

Sure, the current RISC-V Frame 13 main board isn't very fast -- something like a late Pentium III or PowerPC G4, but quad core, but without an Altivec/SSE equivalent -- but it's only a shake-down product with more coming.

That's what i meant it isn't 2005 anymore. That board is for developers, to have something to create distros for in place of a better CPU that time will provide for.
But i don't have to tell you that.
The new DC-Roma board, i didn't know about that! Let's see what happens in the next few months, but i don't see RISC-V application processors that are really interesting (read: that can rival the higher end of ARM chips), right now.
Hope that you will catch up soon. And that you figure out an open and unified way for boot bios and stuff
 I'd rather have a RISC V than an ARM chip. I really got into the "open" movement in the last couple of years, i strongly believe it's the better things for us as consumers

[DiTBho]

Linus Torvalds Says Risc-V will make the same mistakes as arm and-x86, read here.

Dunno, I am still developing Xscale.
StrongArm contains many solutions used on the Alpha project by DEC.
And Xscale integrates solutions used by Intel on the i960.

Things were promising back then.
Then things went to hell.

Let's hope it doesn't happen again.

[SiliconWizard]

It sure will take time to reach the level of performance we have with x86-64 CPUs, but I don't think the "same mistakes" will be made.

The first reason is that it's an open "platform" with a lot of implementations out there now (and it's only the beginning) - x86 has never had much competition at all outside of Intel & AMD. That can't be compared. So the RISC-V CPUs are a lot more diverse and with a real drive to innovate rather than just extract the last drop they can extract on a flawed architecture.

The second reason I see is that the pressure is a lot lower, since users do not expect any kind of "compatibility" with older architectures, as was the case with the x86. So there's a lot more freedom to design things properly rather than just use band aids.

A third reason is that security has been a concern right from the start and designers benefit from decades of history from other architectures. As an interesting example, the target of choice for CHERI is RISC-V (it was historically MIPS, another "challenger", but they have deprecated it and turned to RISC-V entirely).

Now if major companies such as Intel or AMD entirely switched to RISC-V and binned the x86 architecture (who knows...), the story that would unfold could be very different. They would have only themselves to challenge, and it's pretty likely that they would make all the same mistakes again. So let's hope smaller challenger companies get there first and remain on the market for a while (this is the case for now), otherwise RISC-V could very well be doomed as well, at least in what is widely available for the average consumer. Just my 2 cents.

[brucehoult]

Linus Torvalds Says Risc-V will make the same mistakes as arm and-x86, read here.

Linus is a git and that's a completely meaningless statement.

Is he talking about instruction set? Instruction set evolution? CPU core design? CPU bugs? SoC design? BIOS and bootloader? Kernel drivers? Hardware backwards compatibility and feature discovery in applications? Price vs performance? Compilers and JITs? Porting software?

[DiTBho]

Debian 13.0 Ready To Introduce Formal RISC-V Support But Still Bound By Slow Hardware, read here.

[brucehoult]

Debian 13.0 Ready To Introduce Formal RISC-V Support But Still Bound By Slow Hardware, read here.

The news of a variety of RISC-V hardware with performance ranging from a bit more than Pi 5 / Rock 5 / Orange Pi 5 up to near Apple M1 or Zen 2 is really starting to firm up now, with board vendors such as Sipeed starting to tease new products. The avalanche might start before the end of this year, but for sure will be in full flow throughout 2026.

[JPortici]

Debian 13.0 Ready To Introduce Formal RISC-V Support But Still Bound By Slow Hardware, read here.

The news of a variety of RISC-V hardware with performance ranging from a bit more than Pi 5 / Rock 5 / Orange Pi 5 up to near Apple M1 or Zen 2 is really starting to firm up now, with board vendors such as Sipeed starting to tease new products. The avalanche might start before the end of this year, but for sure will be in full flow throughout 2026.

And that would be a dream, i guess we'll see

[DiTBho]

Preliminary Argon ONE UP specs:

Powered by a Raspberry Pi CM5, you can theoretically buy the module that  most meets your need.

All CM5 units have BCM2712 quad-core ARM Cortex-A76 processors, but there are models with as little as 2GB of LPDD4R-4267 memory or as much as 16GB of RAM. Other options include up to 64GB of EMMC storage, and optional support for WiFi 5 and Bluetooth 5.0.

14 inch, 1920 x 1200 pixel IPS LCD display

backlit keyboard and multitouch trackpad

M.2 2280 slot for user-replaceable PCIe 2.0 x1 storage

2 x USB 3.1 Gen 1 Type-A
1 x USB 2.0
2 x USB 2.0 Type-C
1 x USB-C for charging only
1 x HDMI (4K/60Hz)
1 x microSD card reader
1 x 3.5mm audio

[DiTBho]

Once again... Linus T. rejected the risc-v patches, judging them to be "garbage" ....

Too rush? well ...no doubt about.

He was partly right, in my opinion, because it wasn't clear which part of the code was RISC-V-only (arch-dependent) and which part was generic.

That's precisely what he also doesn't like with ARM and MIPS too...