There's no other way to get 256 GB/s of memory bandwidth for this cheap, and that's quite valuable in many workloads. I'm curious to get one for compiling code too.

You can get similar bandwidth with server boards that cost 5-10x as much, or with a Mac Studio that costs 2.5x as much.

Note that this is what CAMM[1] memory is intended to solve, although it remains to be seen to what extent it catches on.

[1] https://en.m.wikipedia.org/wiki/CAMM_(memory_module)

According to Framework, CAMM / LPCAMM is simply not compatible with this line of AMD chips, do to signal integrity reasons.

CAMM will be fine on laptops and other smaller form factor devices for CPU-class memory speeds, but it does not have the bus width or lanes to match solutions like Strix Halo, Grace, Apple M-series -- the memory bandwidth being a large part of their appeal. Increasing the bus width on CAMM modules is going to compromise many of the other advantages.

The problem is that these are integrated shared-memory systems with a single RAM pool. That's nice for a lot of reasons, but GPUs need many more memory channels and larger bus widths than CPUs do in order to do work and remain fed at a reasonable power draw. It's an inherent design trade off. I don't see a CAMM style solution for GPU memory coming anytime soon except on the low end.

A Mac mini with an M4 Pro and 64GB of memory has the same bandwidth and costs £1,999, compared to £1,750 for the Framework Desktop when factoring in the minimum costs for storage, tiles, and necessary expansion cards.

True, but less RAM.

One thing to note on the more RAM: for the 128GB option, my understanding is that the GPU is limited to using only 96GB [1]. In contrast, on Macs, you can safely increase this to, for example, 116GB using `sysctl`.

[1] https://www.tomshardware.com/pc-components/cpus/amds-beastly...

On linux, the gpu can go up to 110 GB.

Apologies, I stand corrected. Do you have a reference for this? I'm genuinely curious why the 96GB "limit" is so frequently cited - I assumed it must be a hardware limitation.

It's mentioned in LTT video: https://youtu.be/-lErGZZgUbY?t=126

(video also features Framework's founder/CEO)

It can go higher actually, just that when I setup my test devices I had a "ought to be enough for everyone" moment when typing `options amdgpu gttsize=110000`. I guess this number spread too far, heh.

See also:

[1] https://en.wikipedia.org/wiki/Graphics_address_remapping_tab...

[2] https://www.kernel.org/doc/html/v4.19/gpu/amdgpu.html#:~:tex...

That's a Windows limitation. On Linux it's 110GB.

> You can get similar bandwidth with server boards

Could be wrong, but I don't think you can. The bandwidth limit, AFAIK, is a problem with the DDR5 spec. These soldered solutions can go faster specifically because they aren't DDR5.

Desktop platforms only have 2 memory channels, amd's latest Epyc servers have 12 channels per socket. Strix Halo has 4 channels.

Please don't misused channels. In DDR4 1 channel was 64 bits. In DDR5 1 channel is 32 bits. So a 128 bit wide DDR4 system had 2 channels, but a 128 bit wide DDR5 system has 4 channels.

The latest AMD server is 12 DIMMS wide, but has 24 channels of DDR5.

Hmm, I think a Threadripper 7965WX can get you there. Probably around 4-5k all in so I guess similar pricing to a Mac Studio.

Nvidia Grace is DDR5, hopper is HBM. Many servers like Intel and AMD's latest and greatest all use DDR5.

Or the NVidia Project DIGITS device at 1.5x the cost, but, also Q2 2025 instead of Q3.

But no published memory bandwidth.

512GB/s from insiders

Sounds promising, hope it's true, much like a mac studio with the m2 max. Otherwise a 128GB amd strix with 256GB/sec memory bandwidth for $2k looks good.

This isn't competing with normal desktops.

Better to think of it as a competitor to Mac Studio & Nvidia Digits, which are much less repairable by comparison. The soldered memory is an unfortunate reality of these "unified" memory systems.

The only way to get a traditional desktop with 96GB of VRAM is to spend upwards of $10K loading it up with 2-4 GPUs.

I just wish they could provide additional options:

1) have ram soldered but provide also extra 1 slot for LPCAMM/LPDIMM ram. Soldered RAM would be used for GPU but external slower ram still could be used for non GPU related stuff (software) that doesn't require high bandwidth/latency.

2) I think motherboard will have 4 pads for ram and if you buy less 1-3 pads will be just having pins without chip. If they would sell memory chips and provide some tooling/documentations for specialised/certified shop they could solder those. You could still upgrade just not by yourself.

The soldered ram was necessary for Strix Halo. There is a large group of people who really want Strix Halo, and are willing to pay for it. There is no reason they should have avoided making this product.

(The 32GB config is silly, though. With that little RAM, there is nothing it does better than a cheaper machine with a discrete GPU.)

There is a reason and I think my prior comment made it clear: When your declared purpose is to limit e-waste, making a new product that does not foster that goal risks alienating the people you won with your purpose description.

> The soldered ram was necessary for Strix Halo

In the LTT video the framework CEO explains that AMD wasn't able to make LPCAMM work because of signal integrity over the bus reasons.

But 2000 dollars for up to 110GB of VRAM in Linux makes this a VERY interesting little machine, so much that the framework website has a cloudflare queue right now...

32gb should be plenty for a little emulation station to sit beneath my TV.

A big product category that is developing right now is hobbyists running LLMs on their own computers - more likely desktops rather than laptops. I presume they want to build expertise and market in this category, and that is why they think the compromise is worth it.

There was no way to make it upgradable because of the architecture of the chip itself. They had no choice, but they are still delivering exceptional value in this design.