iFixitThe @frameworkcomputer Desktop is a powerhouse that gives Apple’s Mac Studio a run for its money at half the price. But it has also created confusion over the soldered down APU and memory. Here’s why we think this device is different, a thread:
@frameworkcomputer Desktop is just a small desktop PC…right? It sure is, but so is the Mac Studio. It’s tempting to critique both on their non-replaceable parts, but the performance we see from these machines depends on those soldered connections.
@frameworkcomputer The immense memory bandwidth of these machines can’t (currently) be achieved on a socketable interface. Extra interference created by the connection points reduces the overall signal integrity.
@frameworkcomputer According to Framework, AMD genuinely tried, assigning a technical architect and running various studies. We do hope it’s possible soon—and we’re waiting with you.
@frameworkcomputer What we end up with is essentially a new device category: a powerful PC with emerging technologies that'll hopefully mature to a point where upgradeability can be achieved. But Framework decided to package it as a mini desktop w/standard form factors, sticking to its ethos.
They’ve committed to providing parts, manuals, and even schematics. Schematics! When was the last time we saw a desktop manufacturer release schematics?! Genuine question, we’d love to know!
They’ve committed to providing parts, manuals, and even schematics. Schematics! When was the last time we saw a desktop manufacturer release schematics?! Genuine question, we’d love to know!
@frameworkcomputer We challenge AMD—and NVIDIA, Micron, Samsung, SK Hynix, and everyone else in this space—to not give up on socketed memory. The future should be upgradeable!
The Great Llama 
@iFixit @frameworkcomputer
Shout that from the rooftops! I've upgraded almost every machine I've ever had.
Shout that from the rooftops! I've upgraded almost every machine I've ever had.
MiskatonicMisericordiaThey already said in the Linus Tech Tips this was a limitation of what was immediately possible, and they themselves want the same thing.
𝘋𝘪𝘳𝘬@MiskatonicMisericordia @iFixit @frameworkcomputer … but the one who pays the band decides what music they play.
alihan_banan@iFixit @frameworkcomputer CAMM memory and advanced nanoguard coating are the two things I wait for in the laptop space - the first one to implement both of them is gonna reap the benefits Toyota once did by becoming a synonym for a reliability.
I hope Lenovo will make a thinkpad p14s gen 6 with ryzen 395, camm memory and waterproof coating to make an ultimate laptop for me.
Scott Williams 🐧@alihan_banan @iFixit @frameworkcomputer I'm typing this on a P14S Gen5 and just gonna affirm that Lenovo has definitely found a sweet spot in this line. My old P14S was fairly easily serviceable as well, though required a little bit of patience with wrestling that back cover.
mxk@alihan_banan @iFixit @frameworkcomputer yes CAMM2 exists, but doesn't solve that soldered memory always will allow for higher speeds and higher bandwidth.
Compute has by far outgrown memory bandwidth.
There are good technological reasons why Apple will never support upgradable memory and for the PC world the only way to continue to support upgradable memory is, if PCs would support tiered memory with both high bandwidth soldered and lower bandwidth upgradable modules.
Compute has by far outgrown memory bandwidth.
There are good technological reasons why Apple will never support upgradable memory and for the PC world the only way to continue to support upgradable memory is, if PCs would support tiered memory with both high bandwidth soldered and lower bandwidth upgradable modules.
Andrew Zonenberg@mxk @alihan_banan @iFixit @frameworkcomputer Modern high speed connectors are very good. DRAM isn't even close to pushing the limits of them, although the DIMM form factor itself is definitely running out of steam. I look forward to seeing what happens with CAMM2.
I expect to see innovation on the DRAM PHY side of things too. GDDR6x uses PAM4 and I wouldn't be surprised if eventually (perhaps as part of a move to an area array form factor like CAMM with more pins) we see them go differential, maybe even using bonded serial lanes like PCIe rather than the classic parallel DRAM architecture (the last holdout in the PC space that hasn't gone to multilane serial yet).
Tangentially related I expect the PCIe card edge connector to eventually be replaced by a coaxial flyover cable using something like Samtec FireFly, keeping the card edge for initial enumeration and power only. Not sure what generation it'll happen but I'm sure it'll happen eventually.
@mxk @iFixit @frameworkcomputer the fact that compute has outgrown the memory bandwidth is true for the long time, isn't it? But 3D V-cache AMD uses in their CPUs seems to solve the bandwidth issue by introducing an additional buffer the same way L2, L3 caches did when they were added to CPUs. Infinity cache in RDNA2 did the same and finally Nvidia did the same thing in the Ada GPUs. Apple wins a lot due to the 32 MBs of L2 and large system cache compared to 8+4 MBs of snaps. Why not keep on it?
@alihan_banan @iFixit @frameworkcomputer CPU managed caching has its limitations, if we are talking about gigabytes of HBM, which is, what I expect as soldered on memory for future SOCs
@mxk @iFixit @frameworkcomputer idk if replacing DRAM with HBM would be good. But cache limitations are cruel. I remember Semianalysis has written that TSMC and most of computer scientists got stuck with SRAM scaling up - they can't get it smaller even with modern tech nodes and they fear that cache is going to be the main bottle neck in the transistor count growth.
But 3D V-cache and chipplet approach AMD seems like a good workaround to me.
@mxk @iFixit @frameworkcomputer sounds like another CISC vs RISC, but now with memory typing.
Or industry will make great in memory computation or just to move to quantum tech before this battle happens ¯\_( ͡° ͜ʖ ͡°)_/¯
Rasmus Lindegaard@iFixit @frameworkcomputer this was also very much the sentiment we saw from framework, which is why my respect for them has only increased.
I want there to be an alternative to the Mac studio. This gives us that, at an incredible price.
And framework expressed their desire to do upgradeable, and they've given us every reason to believe that they're going to deliver an upgradeable system as soon as it's feasible
Rexx Deane 🏳️🌈@iFixit @frameworkcomputer It's not just about upgrading. A tiny memory fault can render the entire machine useless.
woe2you@rexxdeane @iFixit @frameworkcomputer I wouldn't say useless. Component level repair is a bit more involved than just swapping out a DIMM but it's fixable with a microscope, a hot air station and a solder stencil, and Framework strikes me as the sort of company that'd be friendly to third party board repair rather than obstructive.
@woe2you @iFixit @frameworkcomputer it's not exactly DIY, or 'take it to the nearest pc shop' repairable, though.
@rexxdeane @iFixit @frameworkcomputer I get that going that deep into the weeds isn't for everyone, but there are varying bars for DIY. I have a microscope and a hot air station (under £200 for both), all I'd need is the stencil for the particular RAM chip.
Why do I have kit like that? Because it irked me that components worth literally pennies dying can kill expensive tech.
gittaca
Preston Maness ☭@iFixit @frameworkcomputer The device is definitely interesting. That being said, if the tagline selling point is "heavy-duty AI compute," then I'd argue that Framework doesn't know its own customers. But I suppose the market will answer that question in due time.
Ed@aspensmonster @iFixit @frameworkcomputer I mean they seem to be selling a lot of them so apparently people do want it. Not something I need but really happy for them if it keeps upgrades for my laptop coming.
Thibaultmol 🌈 🔜@WHY2025@aspensmonster @iFixit @frameworkcomputer "their customers" isn't necessarily a thing though.
If these are new customers, that's fine too
mmu_man@iFixit @frameworkcomputer and it's sad, as providing schematics has been the norm for 100 years. Tube radio cabinets often had them printed inside back then…
Cassidy James 
@iFixit @frameworkcomputer probably the @system76 Thelio desktops? https://github.com/system76/thelio
System76 
Jagermo@iFixit
I have a notebook that would still work, but because the ram is soldered on and has a defect, its not reliable anymore. It angers me, to be honest. Soldiered ram is the wrong way to go.
@frameworkcomputer
I have a notebook that would still work, but because the ram is soldered on and has a defect, its not reliable anymore. It angers me, to be honest. Soldiered ram is the wrong way to go.
@frameworkcomputer
arturo2bodegas@iFixit @frameworkcomputer Trying really hard to deflect from the fact that the modular computing company is making a desktop system that's less modular than literally 99% of existing desktop computers...
Dashtop
Matt Godden@iFixit @frameworkcomputer In real terms though, are we actually talking a difference in performance that shows up outside of benchmarking, vs using socketed memory?
Ölbaum
ToranpettoI love upgradeable electronics.
But to reduce power consumption we are now at the point that we have to reduce the lateral distance data is traveling. Inside a processor feature shrink did this now between dies is the bottleneck. Memory on package helps but on the cost of upgradeability.
More attention should take to the software side to allow upgrades to the recent version and still keep old hardware.
My Linux still runs on a 6 year old NUC with 5 watts idle.
Steven Reed@toranpetto @iFixit @frameworkcomputer my whacky idea is to have a socket for a RAM disk, something loads faster than SSD, even if it's not as fast as the on-board / on-die RAM. Plenty of workloads could be tweaked to take advantage of it, I'm sure. If nothing else, the memory allocators in modern languages are poor enough that swap is always a benefit, even better if it can be on an ultra-fast disk.
FoxVK@srtcd424 @toranpetto @iFixit @frameworkcomputer SImilar thing came to my mind too. To have high performace soldered ram + slower socekted ram. Of course it would need to have additinal logic guessing which data need to be where. And another question is if there is market for it. Except few hunderds of us.
@FoxVK @toranpetto @iFixit @frameworkcomputer I wonder if the NUMA / fake NUMA stuff in Linux could be made to handle something like that? Have the slow memory on a virtual NUMA node that doesn't actually belong to a CPU?
p
Anthony@iFixit @frameworkcomputer I'm glad Framework is getting more attention, because I genuinely love what they've been doing and want to see them succeed. But to me, this is just antithesis to the whole point. We already have phenomenal flexibility with the ITX form factor, and they're introducing something almost completely un-upgradable? It would at least make more sense with their laptop mainboard which could be later swapped into a desktop chassis, but you don't even get that here.
ProfessorBoop@iFixit @frameworkcomputer Are future framework laptops going to remove the RAM sockets? Why not make a new case that uses the existing laptop boards and then upgrade the boards? To me, this desktop is testing to see if people are willing to tolerate laptop mainboards no longer having socketed RAM. Don't release a new product if it doesn't have the same repairability! Isn't repairability the point of buying Framework? Do schematics even have a practical use if they continue to lock things down?
4bz@ProfessorBoop @iFixit @frameworkcomputer
The schematics in theory could let you upgrade the RAM if you're good with a soldering iron no?
Never done that myself so idk if this is a legit option for these boards or not