> In some ways, the Framework Desktop is a curious machine. Desktop PCs are already very user-repairable! So why is Framework even bringing their talents to this domain? In the laptop realm, they're basically alone with that concept, but in the desktop space, it's rather crowded already. Yet it somehow still makes sense.
And even more curious, Framework Desktop is deliberately less repairable than their laptops. They soldered on the RAM. Which makes it a very strange entry for a brand marketing itself as the DIY dream manufacturer. They threw away their user-repairable mantra when they made the Desktop, it's less user repairable than most other desktops you could go out and buy today.
Dunno, nice, quiet, small machine, using standard parts (power supply, motherboard, etc).
If you want the high memory bandwidth get the strix halo, if not get any normal PC. Sure apple has the bandwidth as well, and also soldered memory as well.
If you want dimms and you want the memory bandwidth get a threadripper (4 channel), siena (6 channel), thread ripper pro (8 channel), or Epyc (12 channel). But be prepared to double your power (at least), quadruple your space, double your cost, and still not have a decent GPU.
> They threw away their user-repairable mantra when they made the Desktop
You forget the value proposition of Framework products is not only they allow you to bring your own hardware but they also promise to provide you with replacement parts and upgrades directly from the vendor.
In this case they could not make the RAM replaceable (it’s a limitation of the platform) but you can expect an upgrade board in about 2 years that’s actually going to be easy to install for much less cost than buying a new desktop computer.
HN admins: can the domain extractor be changed to say "world.hey.com/dhh" here instead of just the domain name? From what I see, Hey World is a blogging platform, similar to Medium but markdown and email based. And the username (blog name) is in the second part of the URL.
It fun to see that, in an era where most CEOs are all-in with AI both on a personal level and at their companies, dhh chose to rather take a deep dive into the world of linux and config files and indie computer brands.
Curious what will the long term impact of this be on the longtime viability of Basecamp and its sister/daughter brands.
CUDA isn't really used for new code. Its used for legacy codebases.
In the LLM world, you really only see CUDA being used with Triton and/or PyTorch consumers that haven't moved onto better pastures (mainly because they only know Python and aren't actually programmers).
That said, AMD can run most CUDA code through ROCm, and AMD officially supports Triton and PyTorch, so even the academics have a way out of Nvidia hell.
In practical generative AI workflows (LLMs), I think AMD Max+395 chips with unified memory is as good as Mac Studio or MacBook Pro configurations in handling big models locally and support fast inference speeds (However Top-end Apple silicon (M4 Max, Studio Ultra) can reach 546GB/s memory bandwidth, while the AMD unified memory system is around 256GB/s). I think for inference use either will work fine. For everything else I think CUDA ecosystem is a better bet (correct me if I'm wrong).
My impression is the same. To train anything you just need to have CUDA gpus. For inference I think AMD and Apple M chips are getting better and better.
All of Ollama and Stable Diffusion based stuff now works on my AMD cards. Maybe it’s different if you want to actually train things, but I have no issues running anything that fits in memory any more.
There is a project called SCALE that allows building CUDA code natively for AMD GPUs.
It is designed as a drop-in replacement for Nvidia CUDA, and it is free for personal and educational use.
There are still many things that need implementing, most important ones being cuDNN and CUDA Graph API, but in my opinion, the list of things that are supported now is already quite impressive (and keeps improving): https://github.com/spectral-compute/scale-validation/tree/ma...
> There's at least a little flexibility with the graphics card if you move the board into a different case—there's a single PCIe x4 slot on the board that you could put an external GPU into, though many PCIe x16 graphics cards will be bandwidth starved.
I like Framework and own one of their laptops. But the desktop seems more a triumph of gimmicky marketing than a desktop that's meaningfully different. And, it seems significantly overpriced.
I'm realizing that I may have misunderstood Framework's market. I thought it was tinkerers and environmentally conscious FOSS nerds like me, but I think there maybe be a huge enterprise segment whose employees in charging of purchasing are like me but answer to much more strict business needs than "Isn't it cool that it comes with a screwdriver in the box?" So for example the underpowered cpu in the fw12 makes no sense to me until I found out that it's also designed for mass purchases by schools and designed to be flung around by angsty teens. The desktop seems to be meant to be strapped to the underside of 40 identical cubicals in an office as much as it's meant to be apparently hauled around by people that want to have CSGO lan parties.
I cancelled my Framework Desktop order and ordered a HP Z2 Mini G1a instead, the goal being to replace my Mac Studio as I've had it with Apple's arrogance and lousy software quality. The HP is much smaller, has ECC RAM and 10G Ethernet. Significantly more expensive, however.
huhm the ars review makes it seem like this thing does worse on benchmarks than stuff like the mac but the test suite in DHH's example goes way faster. Bursty perf perhaps?
I was baffled by the comparison to the M4 Max. Does this mean that recent AMD chips will be performing at the same level, and what does that mean for on-device LLMs? .. or am I misunderstanding this whole ordeal?
A laptop CPU defeats the purpose. Get a 9800X3D for gaming will be waaaay faster or Threadripper for productivity or the 9950X3D chips with 16 cores/32 threads.
Why this laptop crap when you can get a nice PC case.
Then again he thinks the Fractal North was "bulky"? What?
I could do with someone explaining to me (sorry not very advanced knowledge in this area), how did Framework manage this performance at a lower price point? And why can't average Joe at PCPartPicker do something like this?
RDNA 3.5, which means you don't get Matrix Cores. Those are reserved for RDNA 4, which comes to laptop chips later this year. Desktop RDNA 4 only shipped in 2025.
For comparison, Nvidia brought Tensor Cores to consumer cards in 2022 with the 4000 series and Apple had simdgroup_matrix since 2020!
We are moving towards a world where this hardware is ubiquitous. It's uncertain what that means for non-ML workloads.
Is there a real alternative, where is not not needed Windows in any way, to the Remote Device Management baked into firmware as Apple does with its hardware ? This is biggest missing to bring Linux to enterprises.
What are more budget friendly options for similar workloads (running web app test suite in parallel)?
My test suite currently runs in ~6 seconds on 9700K. Would be nice to speed it up, but maybe not for $2000 :-) Last I checked 13700K or 13900K looked like the price/performance sweet spot, but perhaps there are better options?
There's been some theories floating around that the 128gb version could be the best value for on-premise LLM inference. The RAM is split between CPU and GPU at a user-configurable ratio.
So this might be the holy grail of "good enough GPU" and "over 100GB of VRAM" if the rest of the system can keep up.
58 comments
[ 3.0 ms ] story [ 75.0 ms ] threadAnd even more curious, Framework Desktop is deliberately less repairable than their laptops. They soldered on the RAM. Which makes it a very strange entry for a brand marketing itself as the DIY dream manufacturer. They threw away their user-repairable mantra when they made the Desktop, it's less user repairable than most other desktops you could go out and buy today.
If you want the high memory bandwidth get the strix halo, if not get any normal PC. Sure apple has the bandwidth as well, and also soldered memory as well.
If you want dimms and you want the memory bandwidth get a threadripper (4 channel), siena (6 channel), thread ripper pro (8 channel), or Epyc (12 channel). But be prepared to double your power (at least), quadruple your space, double your cost, and still not have a decent GPU.
You forget the value proposition of Framework products is not only they allow you to bring your own hardware but they also promise to provide you with replacement parts and upgrades directly from the vendor.
In this case they could not make the RAM replaceable (it’s a limitation of the platform) but you can expect an upgrade board in about 2 years that’s actually going to be easy to install for much less cost than buying a new desktop computer.
Curious what will the long term impact of this be on the longtime viability of Basecamp and its sister/daughter brands.
Max+ 395 specced with: 128GB of non-upgradeable LPDDR5x WD_BLACK SN850X NVMe - M.2 2280 - 8TB Noctua fan 3x + 3x extra USB-A & USB-C ports No OS option.
only $2,776.00!!!
In the LLM world, you really only see CUDA being used with Triton and/or PyTorch consumers that haven't moved onto better pastures (mainly because they only know Python and aren't actually programmers).
That said, AMD can run most CUDA code through ROCm, and AMD officially supports Triton and PyTorch, so even the academics have a way out of Nvidia hell.
You can find out more here: https://docs.scale-lang.com/stable/
There are still many things that need implementing, most important ones being cuDNN and CUDA Graph API, but in my opinion, the list of things that are supported now is already quite impressive (and keeps improving): https://github.com/spectral-compute/scale-validation/tree/ma...
Disclaimer: I am one of the developers of SCALE.
A very quick Google search would show that pretty much everything also runs on ROCm.
Torch runs on CUDA and ROCm. Llama.cpp runs on CUDA, ROCm, SYCL, Vulkan and others...
https://arstechnica.com/gadgets/2025/08/review-framework-des...
I suspect this one is very similar hardware and a slightly better deal if you give up the cool factor of Framework. Although I don't really know.
Anyone compared them head-to-head?
A laptop CPU defeats the purpose. Get a 9800X3D for gaming will be waaaay faster or Threadripper for productivity or the 9950X3D chips with 16 cores/32 threads.
Why this laptop crap when you can get a nice PC case.
Then again he thinks the Fractal North was "bulky"? What?
For comparison, Nvidia brought Tensor Cores to consumer cards in 2022 with the 4000 series and Apple had simdgroup_matrix since 2020!
We are moving towards a world where this hardware is ubiquitous. It's uncertain what that means for non-ML workloads.
That’d really drive compute.
My test suite currently runs in ~6 seconds on 9700K. Would be nice to speed it up, but maybe not for $2000 :-) Last I checked 13700K or 13900K looked like the price/performance sweet spot, but perhaps there are better options?
So this might be the holy grail of "good enough GPU" and "over 100GB of VRAM" if the rest of the system can keep up.