I’d prefer socketed RAM, but I’d cave to soldered / in-SoC RAM in some cases:
If RAM must be soldered, maybe one could have daughter boards with CPU and RAM on the daughter board and then all the usb ports etc. On a mainboard.
I.e. a bit like a raspberry pi. CM4
The eventual problem would be once pcie speeds increase, and then needing that soldered as well.
Or one finally gets affordable optical interconnects.
So… a post-Panther Lake perspective on this. Seems that all the SKUs with Arc iGPUs need LPDDR5. I think we’re pretty squarely hitting bandwidth limitations with iGPU performance, so there is a fork in the road for Framework. We could continue to use SO-DIMMs and be limited in iGPU performance and probably efficiency, or shift to LPDDR5.
I think the time to shift to LPCAMM2 is here. I would really like to see Panther Lake offered for the next FW13 and 16 refresh, but I would also like to use a SKU with the Arc B370 or B390. To me it seems like the only way to have Panther Lake, Strix Halo, etc. is to make this transition. It doesn’t look like AMD is going to move away from RDNA 3.5 in their iGPUs before Zen 6, but i can bet that RDNA5 based iGPUs will require LPDDR too.
I really don’t want to use soldered RAM, but it does look like SO-DIMMs may have reached the end of their life cycle, unless there’s a possibility to use clocked SO-DIMMs (i suspect they’re more power hungry though).
Of course, ideally we would be able to use SOCAMM2 in laptops and skip LPCAMM2, but unless the AI bubble bursts, i don’t think that’s going to happen any time soon, especially given the ongoing RAM hyperinflation crisis.
If Framework starts offering soldered RAM in more models than they already do, I’m convinced they’ve lost sight of their vision.
Remember back in 2020 this dumb TV blog post?
The product was perfect for us: a dumb TV that as an added bonus reduces e-waste by using panels that would otherwise be scrapped.
I don’t care if the industry is moving towards this or that kind of chip, supporting specific kinds of memory. Heck, the fact that the AI bubble is making DDR5 so scarce should have already seen Framework embracing older tech. Probably less money and glamour in that.
I’m a simple technologist. I don’t run AI models. I write code, tolerate the occasional container, and generally see computers as dumb tools that do things I can’t do in a notebook.
The product that would be perfect for me: a dumb computer that as an added bonus reduces e-waste by using old chips and memory formats that would otherwise be scrapped.
I was big time in favor of framework aggressively adopting lpcamm2 early, even at higher cost but the last 6 months have basically cooked that. Micron was the only company really even selling them to the consumer market, and now they’ve exited consumer altogether because they can make way more money selling every single thing they can make to AI hyperscalers.
I don’t know who you’re even going to source modules from at this point unless MS/Google et all decide they’ve finally built enough datacenters.
I’ve said this in other threads but if the only way you can do panther lake is soldered, do that. It’s way too good to ignore, honestly the first really good upgrade from Phoenix. If you’re willing to do soldered in a desktop of all things, at least it makes more sense in mobile. I want to see lpcamm2 and further camm evolution but right now “ai” is destroying consumer markets (along with some frankly way more important things) so I’ve done a complete 180 on whether I think *camm is going to be viable in the short to medium term.
You will get similar results with Pantherlake.
@earthling Intel sometimes allows for out of paper spec configs. There’s an Asus device out there with 358H using DDR5-7200. It will be slower, but it’s going to be faster than whatever Framework has right now. Also, Intel Xe3 uarch is pretty balanced. 7200 to 9600 is 33%, and the differences will be some fractions of that. Probably 10-15%.
Possible I’ve missed it, but I don’t think I’ve seen any market availability of csodimms over 6400MT, and using xmp is probably a nonstarter in a thin and light because those profiles drive the ram at 1.35v or more. There are already threads here about the fw13 chassis having issues with heat on jedec spec sodimms with sustained loads.
Higher clocked csodimms are probably going to face the same issue as *camm right now; early adopter tech in extreme market conditions.
12 Xe cores are still going to be massively faster even with 6400. Neither FW systems with AMD nor Intel has that support and the new uarch is much more efficient at using memory plus a better architecture.
Have you thought maybe there’s nothing over 6400 because there’s nothing that supports it yet?
https://www.asrockind.com/en-gb/NUC-358H
https://data-us.aaeon.com/DOWNLOAD/2014%20datasheet/Systems/UP%20Xtreme%20PTL%20Edge.pdf