was there no possible way to support more than only 128GB ram?
I read that only 128GB LPDDR5x ram solodered onto with the highest availavble apu/cpu is what framework desktop offers (for now?)
was there no higher capacity possible for this CPU? I was hoping for more than 128GB these days. maybe its because its actually a mobile platform cpu though?
also none of this new kind of CAMM2 physical format that would have been pluggable by the userbase as self service?
thanks.
According to the specifications for the AMD Ryzen™ Al Max+ 395 CPU, the maximum supported memory is 128GB. A shame, really, as the Ryzen 7640U, 7840U, 7840HS, and 7940HS all support 256GB of RAM.
Seems AMD made compromises in some areas to add things like their better GPU, a NPU (Neural Processing Unit which is basically a really good multiplier from my understanding), and allow for sharing of the RAM with the GPU, especially in large amounts. Up to 96GB for the GPU in windows and up to 110GB in Linux.
From my understanding, the real market for this CPU is for AI applications, even over gamers, and especially for those looking for a desktop for basic daily use. For instance, if I were to get one of these Framework Desktops, it would be a complete waste of my money, since all I do is web browsing, sound editing, and things of a similar nature that could easily be done very well even on Framework’s lowest end Ryzen CPU.
It sounds like LPCAMM was not kept in mind for the APU design and there is no signal headroom to support it. So you are stuck with soldered chips. And I assume that’s the same reason you cant double them up to support 256GB.
Keep in mind though, the memory bus is twice as wide vs. 7840U et. al. Even vs. desktop chips.
Seriously makes me question this decision by the Framework team… Like, I get the perspective of making a repairable AI PC; but in all honesty, most people that even want to waste time on AI nonsense can probably build better systems already.
Disappointing to see them go this route. Would not be surprised if down the line they release another motherboard that is not AI centered (AND with modular RAM)
Don’t blame Framework, blame AMD on this one. They wanted that super high 8000megatransfer per second memory on this, as well as 4 channels. There’s no room for extra wire length when you’re going that fast, and LPCAMM memory only supports 2 channels, so even if they wanted to use slower memory, there wouldn’t be enough room for two lpcamm modules. I was going to get one of these from HP or Asus otherwise, them releasing this just lets me support a company I would have rather bought it from anyway.
Repairability is a solved thing on desktops for the most part, so there’s no real work for them to do in this space, unlike laptops. The fact that they’re giving us a bespoke mobile APU in a standard mITX form is bonus enough.
I mean, it is Framework who ultimately chose to partner with AMD and chose to sell an AI PC. And to not be crude, but I am pretty sure the AMD executives did not hold the Framework CEO at gunpoint or anything to make a Framework PC.
I guess it’s to be seen if people truly want to buy this specific product out of the new line they announced this week.
I certainly do, this is the product I’ve been waiting for from AMD for 10 years, ever since they released the Kaveri, then did a decade of half-hearted, barely above low-end GPUs on their APUs, they finally have something that’s not only midrange, but upper midrange in an APU. As someone who prefers tiny PCs and low power consumption desktop boxes, this is exactly what I’ve been waiting for AMD to produce. They showed they could produce high end APUs with the Xbox One, Xbox One X, Xbox Series X PS4, PS4 Pro, PS5, and PS5 Pro, but they’ve always withheld them from desktop users.
I mean, it is Framework who ultimately chose to partner with AMD and chose to sell an AI PC
well there is no alternative so who would they have partnered with instead?
there is no way. that is the maximum that AMD supports and CAMM2 doesn’t support 8533 mt/s, just 7533 currently.
Sounds like it would have had to have been even slower than 7533. I do wonder what the actual numbers were though.
Depending on your workload, tbh, even with the extra Strix Halo bandwidth more than 128GB might start to get bandwidth constrained. Bigger issue is being unable to buy at 32GB and upgrade later.
no, lpcamm works with 7533
How about quad channel lpcamm? Does that exist?
well technically yes, because rather than each dimm handling all 64 bits of bandwidth over 1 channel, it handles 2 channels of 32 bits each per dimm.
in theory the 8 channel setup youre suggesting here is possible, there shouldnt be any reason why it shouldn’t work
One of Apple M series’ biggest plus factor is their unified memory and the memory bandwidth. Its already got a niche community of people that are building local LLMs, plugging in multiple Mac Minis together, etc. I think it wrong to “blame” AMD on this, they’ve seen how well the mac minis are selling and are finally delivering an alternative.
In case anyone is interested:
GitHub - exo-explore/exo: Run your own AI cluster at home with everyday devices 📱💻 🖥️⌚ (25k stars)
One more thing - even if one is not keen on LLMs, the high bandwidth unified memory model actually finally makes it feasible to run games on an integrated GPUs without the memory bottlenecking. My personal view is that unified cpu/gpus like these are the future, at least for consumer electronics.
Yes, but I suspect the issue is that the APU memory controller does not have a clean enough signal path to allow LPCAMM to work even at 7533. Keep in mind how much trouble AMD had (vs. intel) with DDR5 on AM5. This is a different IO die but would not be surprised if same deficient controller design / firmware is used.
Or maybe AMD did not both trying it and FW did not want to risk it. I dont know. Just saying there is more to it than just LPCAMM rated speed.
ah. yeah, so in the LTT video nirav stated that they talked to amd about it and they put one of their top engineers on it IIRC and the signal just wasn’t clean enough. They didn’t say anything about memory speed but most OEM’s ship laptops with ram that has the max frequency that the flavor of ram theyre using (be it ddr4, lpddr4x, lpddr5x, ddr5 etc) can support
To add to this, in the LTT video Nirav said that camm2 is 128bit so they’d need 2 modules, and there was no way to route twice the number of traces while keeping signal integrity.
I think they could’ve made it work if they dropped the clock as much as needed at the cost of bandwidth.
The chip supports 128G max, so for 128G version it make sense to solder chips down; you can’t upgrade it. It’s only blocking repair / reuse of mem.
Yeah I think the framework team made the right call here.
The only real downside is not being able to buy the cheaper board and upgrade later.
Not sure if possible, but maybe they will make it possible to have a skilled electronics repair person solder on new (larger) chips later. That would be cool if they sold a bundle of the correct loose chips and a solder mask in the market place. Not ideal, but better than nothing.
most people that even want to waste time on AI nonsense can probably build better systems already.
As someone who wants to “waste time on AI nonsense”, yes I could build a better system but it would be both significantly larger and more expensive than what Framework is offering.
I personally don’t see the appeal of this system as a gaming rig, you can build a much better desktop for the same price
Maybe as a console alternative/small, living room gaming box? Of course, you could just build a Mini-ITX machine yourself for that purpose.