I do want socketed lpddr as soon as possible but I am not willing to give up upgradeability and got with soldered lpddr.
Also current desktop arm processors still kind of suck, especially software wise which kind of defeats most of their potential efficiency advantages or chain you to an apple eco system (and in case you are fine with that just get one of those).
Socketed ram also keeps manufacturers honest, as much as I like framework I doubt they can resist the temptation the get some juicy margin on soldered ram like everyone else does forever.
There is also the advantage of having fewer different skus which is good for the manufacturer on initial sale and for the secondary market.
Battery life was one of the explicitly called out factors in the poll - I was merely pointing out that designing hierarchical memory doesn’t grant every advantage of both on package and socketed; not that it necessarily have to although I would and will argue that battery life is the primary driving force behind on package ram.
I would dispute this, at least in consumer space. As far as I know every on package consumer design has been a mobile first architecture. I would argue battery life has been at the core of Apple’s chip design philosophy, and it -certainly- was with lunar lake. Frankly, the difference between 8000MT/s RAM and 8400MT/s RAM is not all that significant.
I’ve “fixed” enough computers by removing/replacing a stick of RAM to have a very strong opinion on this.
If a company were to offer a five year mainboard repair/exchange+recycle program, the same or better, then I would gladly accept soldered on RAM.
I dislike how most manufacturers kneecap their low end offerings in order to increase sales of more expensive products. If my Surface Go were actually supported by Windows 11, I wouldn’t have sold it, doubly so if it had 16GB or memory.
I ordered the FW12 because I want a small device with a little kick, and significantly more memory than I need. I also ordered it because I know what my next laptop is going to be, and the one after that.
With the caveat that you are prepared to accept some performance degradation. That is the reason the FW desktop has gone with soldered RAM, the degradation using socketed RAM was too great.
As I said in the other half of my message you replied to the issue was the required trace layout for using LPCAMM2 modules with a 256-bit bus (which is what the Strix Halo chips in the Framework Desktop uses). With the 256-bit bus Framework said that the penalty would be close to 50%, however with a 128-bit bus it works much better.
For example Lenovo Thinkpad P1 Gen 7 supports LPCAMM2 and can run it at full bandwidth that the CPU inside is rated to support with soldered ram (and only 6.25% lower bandwidth than the Ryzen AI 300 series CPUs in the FWL13 are rated to support, although based on Micron’s announced roadmap they expect to improve LPCAMM2’s bandwidth by 13% this year and make up for that).
If more platforms with wider buses show up I hope they make a bigger lpcamm standard that can cover those, as far as I understood the main issue was needing multiple lpcamm modules which brings back a lot of the issues lpcamm was trying to solve vs sodimms in the first place.
Disclaimer: haven’t read any of the above comments. Just here to share my opinion.
My usecase: I’m a developer who is basically tied to a desk (and powersockets) 24/7.
I do very much like the upgradeability that socketed RAM provides. While I get that ARM would be cool, I don’t necessarily find battery life to be a big factor for me. If i’m working, it’s going to drain anyways within only a few hours (even on ARM, can say from experience). And like I said, I always have power sockets nearby.
However. If framework is going to keep doing (awesome) framework things, it might not be as big of a deal to have soldered RAM. I think that if you buy a framework and get a good amount of RAM (for me, I could get away with 16Gb, but I’d go with 32 if I were to buy one today), it should last you a long time before needing an upgrade. By the time that upgrade is needed, you’d probably want to replace the whole mainboard, so that’s not really an issue.
Obviously, having to add/replace RAM before you’re due for a mainboard upgrade wouldn’t be possible, which would suck. Hard.
That’s why I voted “I’d prefer socketed”. But I wouldn’t hate framework if they do end up going soldered at some point. As long as the upgradeability and repairability of the rest of the system doesn’t get compromised!
LPCAMM3 is expected to be 192 bits wide as part of a bigger push to make 192 bits the new normal (like how 128 bits was normal for decades).
Although I expect that when 192 bits is normal then products like Strix Halo will have moved on to 384 bits. So making 192 bits the new normal wont completely solve the problem.
Kinda doubt they would go down to 192 (at least unless they go with something like onboard hbm or something + slower memory). Would be nice if they figured out how to use multible lpcamm modules or the standard just expanded to making “narrow” and “wide” modules.
Also just add ecc by default already come on, it’s pretty much 98% already there they just gotta close the gap.
The poll is flawed. What if I “am fine with soldered LPDDR” but also “I’d prefer socketed DDR RAM” at the same time?
I think the real reason is because most soldered-on RAM (regardless of type) have a absurd price tag, tied to other configs that also have insane upcharge.
The other is capacity. 16 is not enough to be “maximum possible ever”. But 32 is quite good.
If I can choose between a $100 soldered LPDDR 32GB versus $100 socketed DDR 32GB, I dont think I would really care. Though, I have some edge cases where I did hit 32GB (Unity Studio, lol), so maybe I do want socketed.
Maybe a good follow up poll would be “would you be in favor of more rapidly adopting new memory form factors and technologies when introducing new platforms if they offer real performance advantages over what is currently in use, even if it means not having backwards compatibility with memory used in previous generations?”
The concrete example of this now would be lpcamm2 versus sodimms, but this question has been relevant before with the earlier Intel memory controllers which supported both ddr5 and ddr4. At the time, FW chose ddr4 citing the ability to carry over ram from previous motherboards, but that philosophy may not be in line with what a lot of people want.
I think of this more from the perspective of repair ability and reducing waste.
If everything is soldered onto the main board. If anything fails on that board it generally means throwing it in the trash and getting a new one.
If the RAM chip is in a socket that a user can replace. The RAM chip can fail and the user can replace it. Result, very little waste.
If FW were a little more open with their board view and schematics, I could potentially repair the main-board myself, without sending it to the skip. Thus less waste.
So, making the board view and schematics public would go some way to mitigate the soldered on RAM problem.
I recall Framework saying that it’s due to their mainboard design partner. Unfortunately, it seems they aren’t as enlightened about open sourcing their hard work as FW is. I assume Framework will have to grow larger as a company before that becomes an option for them.
To be quite honest I would prefer a socketed solution to a soldered one any day. Having soldered RAM as opposed to socketed RAM in future FW laptop boards would be a step backwards in upgrade ability and repairability, in my opinion. I understand that for Framework it could be more scalable if soldered RAM was offered, but in doing so we’d be adding to the e-waste problem by having to ditch otherwise useable boards in case the board becomes damaged or carries a faulty RAM chip that isn’t user-replaceable. There might be a price to pay scale wise, but sustainability wise I would rather dispose of a small memory stick than a whole motherboard. I hope that new solutions and upgrades for the FW 16 will keep socketed RAM; but if soldered-on RAM becomes the go-to standard, I will adapt.
Not to refute or undermine the importance of the poll in the established community to help define direction of development for the Framework however this poll might be subject to sampling bias it terms of convenience sampling and homophily(majority of people here might advocate for acess to repairable parts like yours truly and hence are here in the first place). It might serve the company better to include opinions of larger and more diverse population groups.
Just a suggestion.
To be fair, this poll seems to have been made by a member, not by Framework. I’m sure Framework takes into account a lot of factors before making big changes.
I think one reason I want Framework to stick with socketed RAM as long as possible, is because I like choice. And I don’t just mean the choice to swap out RAM, but the choice to have socketed RAM at all. I know there is higher performance to be gained by soldered RAM, but there is always going to be plenty of choices for devices on the bleeding edge of high performance. And plenty of choices for average machines with soldered RAM as well.
But there are very few options out there for laptops with maximum modularity. Framework is one of the few. If they end up going with soldered RAM simply because the majority of folks in the world are fine with it, or even prefer it for the slight performance gains, so be it. But since they are one of few, remaining options out there with socketed RAM in a laptop, I hope they stick with it for now.
I am in favour of socketed RAMs too and agree with you that it would be nice if this option continued to be available. And though this might be a poll by a member ( I was not aware of that) the companies decision might be affected by community opinion and hence be subject to bias which might further inadvertantly affect the consumers downstream, economically and support wise so i was just putting it out there.
Don’t worry about it, whenever I share community discussions, poll and survey results internally, I always point out that these are the community results and does not represent the entire customer base
)