Introducing the Framework Desktop

Which we will hopefully see reflected in a longer warranty. But yeah like you I initially saw this through the lens of professional workstations instead of gaming/gimmick machines ($2k can take you pretty far on pcpartpicker). With that in mind I hope ROCm comes closer to parity with CUDA so I can pitch this to leadership for a few upcoming projects.

1 Like

Ah…that makes sense then. The limitation is on the processor itself.

Maybe we should wait for a similar offering from the desktop processor lineup with 50+TOPS, with expandable memory.

The challenge is making the memory both expandable and very high bandwidth.

For example the ASRock NUC BOX-255H has 96 TOPS total and should be compatible with the Crucial CT2K64G64C52CS5 to achieve 128 GB of unified memory (and that CPU supports more PCIe lanes, although that specific Mini PC I listed doesn’t expose many of the lanes). However that memory would only be 102.4 GB/s rather than the 256 GB/s that the Framework Desktop uses.

2 Likes

Currently. Yes, I understand the limitation. I’m looking into the future.

It will be solved. Otherwise, we’re looking into a future where servers are stuck with soldered-on memory. The alternative to that is to have dedicated NPUs on the bus away from the processor.

Update: Looking at the Xeon and Threadripper materials…the 256GB/s memory bandwidth is already achievable today(?) over 8 channels…it’s a matter of AMD not trickling that down yet? Maybe it’s a question of per channel peak bandwidth.

https://www.amd.com/content/dam/amd/en/documents/products/processors/ryzen-threadripper/amd-ryzen-threadripper-pro-workstation-raising-the-bar-white-paper.pdf

I’d really love to see that proof-of-concept eGPU. Really looked like a great idea, maybe with options for 16 GPU and other GPUs sans interposer.

Lack of pci express 16x is really disappointing. I am actually considering a desktop build in the near future but not having the ability to upgrade the GPU seems out of character for a framework upgradable, repairable product. Support for the framework 16 gpu would have been cool.

3 Likes

I think that for a lot of people, the fact that you can assign 96 gigabytes to the GPU means that it’s an unusually good system for running AI models… Even though that’s not what I’d want it for, that’s probably where they’ll get the most interest in the market.

Actually, their AI-370 boards are the most upgradable AI-370 boards out there. They’re one of only 2 makers that have AI-370 laptops that can take dimm memory, instead of having soldered ram. Because of that, if you are looking for upgradable anything, then that’s where you should focus, on their laptops, rather than their desktops.

Limitation of the CPU unfortunately. 4 sets of x4 lanes is all this has, if you put them all on the GPU, then there would be no way to add storage to the system. Honestly, most GPUs get by just fine with an x8, but an x4 might be a bit too little if it’s not PCIe Gen 5. If you’re going to want a GPU, it’d be best to just get a 9950x for now and put it in a normal mini-itx case, the main selling point of this is the built in GPU, and if you’re using an external GPU, then you’re just wasting half the cost of the system. I’m only buying this because I want a good, low power system that I can put at 45w, then play almost all of the current games without a power chugging discrete GPU. Given the recent power consumption of the Nvidia gpus, Chugging is an understatement. I’m not sure how to express just how gluttonously hungry those cards are.

Sure, you can do it with 8 channels. This is accomplishing it with 4 channels though. You wouldn’t be able to fit a threadripper with 8 channels of dimms into such a small case, and I’m not sure there’s any threadripper or epyc systems with the same degree of integrated GPU, outside of the very new MI300 Instinct ones, and those use HBM, even more tightly integrated memory than the LPDDR in this. The other thing is, latency is a thing, and the farther away the memory is from the processing, the higher the latency, the longer the turn-around times will be.

Heh, well, you can only use that x4 slot if you put it in a different case, the Frame.work case sadly has no provisions for actually using that slot. The only thing I’d use it for anyway would be if I wanted a full 10gbe card rather than the 5gbe port built in.

Sadly, the form factor for their laptop motherboards are drastically different from the mini-itx form factor that this desktop uses, but there is a $39 case that they sell to put an old laptop motherboard and use it as a tiny desktop. About the only thing that could make it more like their desktop case would be to add support for those little customizable tiles to those cases.

4 Likes

Is there any way to add a VESA mount to the back of the case? Without that I might end up needing to re-case it at some point in the future so that I can put it onto the back of my desk. Otherwise I’ll have to use a wall mount kit for it and that may block some airflow.

Because it’s the revolutionary tool of our generation. Running Flux and Qwen/Phi locally to have image, novel and code assist are some of the things I use it for.

And certanly GenANI is not the biggest contributor to CO2 emission. GenANI assited worflows use less energy than spending more time without GenANI assistence. My GPU uses 400W for 300s to make an high definition picture. It would take hours of photoshop to get similar results using more energy and time.

A use case I see for it for the DF-128GB is as a perfectly specced box to run 70B models locally and run local VS Code completion.

You are correct.

I was looking into it, the cheapest huge option is EPYC9124 with TWELVE DDR5-4800 channels. You get about 768 GB at 460 GB/s for around 4000€. That’s enough to run the full Deepseek-R1 671B

FW-128GB gives you 128GB at around 250GB/s for around 2400€, but you don’t have to deal with the headache of having server motherboards. It’s a compelling proposition. I think you can get two instances of Qwen 2.5 72B Q4 running, or one instance of Qwen 2.5 72B Q8.

1 Like

For anyone looking to add a video card to this, you probably shouldn’t buy it. For $1100ish, you can get a 9950x with 128gb of ram and a mini-itx board. If you’re going to just disable the one major advantage of the AI395+, then you’re wasting about $900 getting one of these. You’re also sacrificing CPU power by a little bit going to this over the 9950x, because that’s a full desktop chip you can push to 170w, instead of a laptop chip that tops out at 120w and has to share its power with the GPU.

On the other hand, if you want the best possible integrated GPU, and a system that runs almost everything while sipping power, this is the system for you. :smiley:

4 Likes

From what I understand, this standard was developed to counter the shortcomings of SO-DIMM. If it introduces other shortcomings instead, wouldn’t those turn it into a case of DOA?

EDIT: After doing some research, I got my answer, but sadly not from Framework directly. IMHO, they should’ve included that somewere. You can find it in LTT’s video at 07:25.

That’s a “chicken-and-egg-problem”.

If a 256bit memory bus is so difficult to do due to electrical transmission line problems, maybe the technology needs to move over to optic fiber connections instead. One can pack optic fibers closer together than electrical traces on the board and they are easier to route and can cross without interfering.
So maybe, with ram speeds are reaching their limits using electricity, it will force the market to move more to PCB level fiber links quicker than they might have done otherwise.

Strix Halo doesn’t support normal DDR5, only LPDDR5x

It’s exciting to see Framework put out other product lines and this is just version 1 - I’m sure down the line we will see a version of the Framework desktop with removeable ram. From the sounds of it, it’s really up to AMD to make it happen. For someone like me, who doesn’t have a desktop already but does a fair amount of ML and 3D rendering for scientific research purposes, who hasn’t done much gaming other than replay the Last of Us, who would get lost trying to figure out what parts to get on Newegg, this is a viable option. They are on Batch 5 on the shipments for all 3 models already so clearly the market is out there.

Given Framework’s track record thus far with my FW13 for the last 3 yrs, I have no reason to doubt that they won’t continue to produce high quality repairable and sustainable products in the future. Although it would be nice to see removeable ram, I’m not the type of person who would need to change my ram unless I’m also planning to change out my motherboard anyhow.

2 Likes

Im excited about this new desktop, but putting that aside I think some people are missing the bigger picture here, who cares about running games at 4k, and ray tracing, think about it your running AAA games at 1440p high or ultra setting on an IGPU, When the Ryzen G series IGPU’s came out it was pretty impressive, now not long after we have the max+ 395 this is great for the laptop community as you have the processing power but doesn’t sacrifice space for better cooling or battery size I think these chips will get better and more powerful and I hope the max+ 395 chip comes to the Framework 13 series laptops, although that might be a bit of a stretch :sweat_smile:

2 Likes

Sad to see so much hate in here… why do folks always have to pitch in with “I need to play Cyberpunk 2077 in 4K Ultra at over 200FPS so this product is useless” … no sweetie, it’s just not for you.

Here’s hoping that this product does really well, cuz tiny PCs are something we all should want in our lives. I’ve already got a few weird and wacky Mini ITX cases, so I’m struggling to justify a purchase here… but I’ve got a config saved in my account… I just need someones credit card and a couple of beers … … j/k

I also priced up a generic Chinese Mini-ITX case (5.8 litres was the smallest I could see on PC PartPicker) with a B550 mobo, Ryzen5 5700 CPU, 32GB of DDR4, a 4060, 600W PSU and it still came out to nearly £1,000 … it’ll likely be lower performance than the Framework desktop, doesn’t look anywhere near as cool and doesn’t support a company I believe we’re all rooting for? Oh, and that 4060 only has 8GB of VRAM, so will suck for AI tasks… which is what this box appears to be for.

7 Likes

Why is there an pcie x4 slot but no bracket slot on the back of the case? How are you supposed to take advantage of the slot if there is no output slot for an add in card?

I’m guessing they wanted to reduce the volume of the case as much as possible.

But keep in mind that the motherboard is also sold as a standalone product, which can then be inserted into a case which features brackets.

Speaking of which, there will be probably be 3D printed cases specially mimicking the Desktop, but with an additional bracket.

1 Like

After this product announcement I’ve been trying to figure out how it could be useful to me. I think I have. Today I own a Framework 16 and a tower with a 5950X and 3080. I also own a NAS in a Jonesbo N3 case with an intel 6 core CPU. The 5 gigabit networking and PCIex4 slot is almost exactly what I would need in a homelab NAS setup (10 gigabit Ethernet would’ve been amazing). I don’t use the 3080 in my desktop as much as I would’ve expected when I first got it so I could justify adding in a GPU to my Framework 16 for complete desktop use, and upgrading the mainboard in my NAS to this desktop mainboard (no case since I can’t fit 8 hard drives in it). Using this mainboard as a NAS sounds like overkill but I could see myself using the processing and RAM for video decoding as a NVR, object recognition, local LLM, and other Strix Halo purposes via containers running on TrueNAS OS. I have 80TB of hard drive space in this chassis so the management of ZFS using the extra system memory is easily justified for me. If I decide to upgrade my NAS this board will be a strong candidate for me.

2 Likes

Would be so cool if it had separate compute (Altho mainboard kinda is that) and separate GPU modules that you’d slot into the case. Like an actual lego-style PC. Some NUC kits did try that but they’re gone now :confused: