Poll on soldered vs socketed RAM

Note:
This poll concerns the future of framework laptop mainboards (12, 13, & 16). I’m not concerned about the framework desktop.

Goal:
I’d like to understand how the framework community feels about a framework mainboard with soldered LPDDR RAM.

Some context for the poll:
With ARM chips becoming more popular, and soldered LPDDR RAM being necessary to squeeze the efficiency out of ARM chips, I want to get a pulse check on how open the framework community would be to a soldered RAM laptop mainboard.

The tradeoff:
CON - With soldered RAM, you lose a granular step in upgradeability. Instead of upgrading RAM whenever you see a sale on the latest speeds, you need to wait until you are ready to upgrade the entire mainboard. Additionally, we lose the “bring-your-own” option during the buying process, which leads to potential price gouging (framework has said they will avoid doing this for the framework desktop that already uses soldered RAM)

PRO - What you gain in return is the ability to use ARM processors. ARM processors have a much better trade-off between power and efficiency. Meaning longer battery life and cooler systems with the same level of performance or better (*when paired with properly calibrated firmware). This would allow the framework laptops to catch up to the others, leading the industry in battery efficiency. Which is of particular importance to the framework 12, where we would likely see an increase in processing power and battery life working within the thermal constraints of a smaller chassis

So, how do you feel about soldered LPDDR RAM vs socketed DDR RAM

  • I really want soldered LPDDR RAM (and ARM CHIPS!!)
  • I’m fine with soldered LPDDR RAM
  • I’d prefer socketed DDR RAM
  • I only want socketed DDR RAM
0 voters

PS: Please let me know if I got anything wrong here!

2 Likes

moving this to general topics as this is about 3 separate products, curious to read the discussion here :slight_smile:

3 Likes

appreciate it, Destroya!

I’m fine with soldered RAM if there’s a major performance benefit to it, such as the unified memory on the new AMD Strix Halo chips. We’re reaching a point in time where the need to increase RAM is leveling off, at least for my personal use. Unless software engineers really hate end users for some reason, I think it will be a very long time before 32 GB of RAM will be a bottleneck to people.

Soldered RAM on 16 GB or 8 GB systems is a harder sell though

5 Likes

100% agree. I think a disclaimer would be useful, and ideally, the base model starts at 32gb (maybe 16gb for the FW12)

those who shall not be named offering base models with 8gb soldered, feels very dishonest to end users (tho they have finally started putting 16gb in their base models :slightly_smiling_face:)

1 Like

Yup! Give me meaningful support for important RAM-dependent use cases (looking at you AMD/ROCm) and the “do I really need 128/256GB RAM?” dilemma melts away. I’d just go for max possible RAM for the benefit of running more capable local models.

1 Like

Soldered ram is not synonymous with LPDDR. LPDDR can be made replaceable using LPCAMM2 modules.

With Strix Halo the 256-bit bus meant that the trace layout and signal integrity required to make LPCAMM2 work wasn’t feasible, however current ARM chips use a 128-bit bus and are almost the exact use case that LPCAMM2 is intended for.

12 Likes

If prefer socketed RAM, but I can accept soldered RAM on the smaller, low power models if it’s plenty and inexpensive. Like with soldered ram I wouldn’t do lower than 64 GB because I can’t upgrade it. With socketed RAM, 32GB would be more than enough for me at the moment. So the soldered RAM for me would need to be half the price (of the Sticks from Amazon) or provide other huge benefits to be worth it for me.

1 Like

I went with “I only want socketed DDR RAM”. For a couple reasons.

I’m not really on board with arm for PC use, yet. Because, one, we can not accept anything but completely open bootloaders, from the factory, across all chips for PC use, no exceptions, no unlocking required. Hard line. We can not accept even needing to press one button to unlock. As it will lead to making it more difficult, introducing requirements and / or forfeiting warranty. As has been proven with mobile phones and tablets, they will lock it down. If given a way to progressively get there, they will use it. If your bootloader is locked then you are not permitted to use it how you wish, and you do not own the device you paid for.

Additionally (click to show)

If qualcomm is the only option, either with the best performance or for any other reason, then it is not worth it. They drop support for their chips as soon as they can, preventing manufactures from being able to offer OS upgrades. And due to their many other bad, arguably illegal practices, our prices are higher, our choices less. Generally, I trust qualcomm as far as I can throw them.

I don’t want to lose software compatibility when moving to ARM. I don’t know how good emulation or translation is at this time, or what level of performance hit there is. I suspect it isn’t good enough yet. Particularly care about Linux, via free solutions.

No less importantly, moving to less upgradable and repairable PCs should be resisted strongly. It certainly may be required to accept soldered ram at some point in order to have halfway acceptable performance, but I don’t think we are there yet. We have LPCAMM2, which we haven’t even begun to utilize. If people accept soldered ram too easily, then there isn’t the need for companies to utilize LPCAMM2 right now. And there certainly isn’t need to invest time and money to ensure an upgradable, repairable solution remains available for as long as possible in the future. Only by strongly resisting soldered ram is there the need for them to do this. If we don’t demand it, it will be taken away long before it’s truly necessary.

So, for these reasons I feel Framework should only use socketed RAM in their laptops, for as long as is possible. And we should encourage them to do so.

7 Likes

All those MacBooks and Microsoft Surfaces with 8 GB of RAM which will forever be just trash… shudders

My problem with the built-in RAM is the margin that manufacturers build into it and get away with. It’s built-in monopoly and you get charged for it thusly.

Right now 64 GB DDR5 RAM kit (32x2) is 143€ on Amazon, and 360€ in Frameworks configurator for Laptop 16. I don’t mind Framework giving an option to people to overpay out of convenience or ignorance, but for me, this is a significant amount of money to be saved.

4 Likes

Another small factor for why I will always prefer socketed RAM when possible, is that when soldered, it’s just one more component that will render the board all but trash if it fails, vs. being able to pop a new one in. I know, soldered RAM can be changed in some instances. But, for instance, on a FW 12 board? On something like that, would it be financially feasible to pay someone (or have the equipment yourself) to solder on a new RAM chip? No, you’re going to pitch the board and get a new one.

I know that RAM tends to be pretty reliable these days, and maybe I’m just being silly. But I prefer socketed RAM for repairability, in addition to the reasons mentioned by others above.

5 Likes

The biggest problem with soldered ram imo isn’t the lack of upgradability or repairability, it’s bundling - companies don’t want to maintain a billion skus so large ram capacity becomes tied to the highest tier of CPU. Not only does the buy in cost for large ram quantity go way up as a result, but many companies will also use that to disguise ludicrous margins on RAM chips.

That being said, the advantages to soldered RAM are not limited to arm. Look at what Intel did with lunar lake - they were able to match arm idle/low load power consumption in part because of on package ram (this goes one step further than just soldered, but same idea, and this is what apple does as well).

That being said, before lpcamm2 became a standard, I actually would have been in favor of FW moving to soldered RAM. The reality is we’re leaving igpu performance on the table because we’re stuck with 5600 MT/s sodimms. Now that lpcamm2 is a thing, I’d much rather framework pivot into that; it’s way easier to swallow giving up like 12% of potential memory bandwidth vs 50%.

The caveat there is that even with lpcamm2 you probably won’t match low load power draw of soldered, let alone on package, but my bias is to not care about battery life so much or I probably wouldn’t own a FW to begin with.

E: I do think fw should be more aggressive about moving to new ram types and breaking backwards compatibility with models from previous generations when there’s a legitimate performance reason to do so. Imo the strix point main board should have been the first on 7500 MT/s lpcamm2. Let the 890m breathe a little.

Why not both? It’s just a memory (information accessibility) hierarchy architecture. Sure, it’s not something for Framework to tackle, but let’s not rule out that it’s a matter of design not laws of physics.

Give us L4 (on chip) and L5 (on board).

People bought into Apple’s crap thinking that you can only have one or the other.

People have been using L1,L2,L3 cache on-die all this time, they don’t worry about cache dying / repair. (All of them is a form of volatile memory)

2 Likes

Because if your purpose for on package ram is to realize battery life advantages, also having ram on the board as well probably undoes that.

If you’re trying to realize performance advantages, that approach has been used; see accelerators with on package hbm combined with much larger quantities of dram on board.

It’s heinously expensive, and frankly probably not worth the cost and complexity if you’re talking about on package dram combined with on board dram. Even on package lppdr5x is only running at about 8400MT/s.

I wouldn’t expect to see this done in consumer space any time soon.

E: if you’re talking about something that acts more traditionally like an l4 cache that happens occasionally like back when broadwell had a small cache for igpu but that’s not going to be in ram like capacities - small and fast is the name of the game there.

Let’s not forget the desktop use case. It shouldn’t be a one-size-fits-all solution. Leave the choice to consumers.

The on package solution can be faster…they just want to sell people the idea that “We’ve moved RAM to the chip”, it’s a narrative thing.

Degree of complexity is a matter of civilization progression. Try time-travel back 50 years and tell them that we now have 1000 nit+ smartphones with AI, and multi-TB storage in personal laptops. Today’s norm is yesterday’s complexity.

Agree with this.

My point is, it’s not one or the other. It’s just that the choice was made for you, leading you to think these are mutually exclusive options.

…hell, cache used to be on the motherboard. (e.g. 386 L2 cache, it was a thing as late as 486 until Pentium)

Nowadays, we have more L3 cache (e.g. on 9950x) than recommended system memory in the Windows NT4 days (32MB)

I don’t think they’re mutually exclusive options; there are real world examples where multiple hierarchies of system memory are in use right now, including on package; they all cost 100k USD+. If there’s a substantial advantage to doing it and it’s in a market segment where margins are high enough, of course it will be done.

My point is that nobody will make specifically on package DRAM + on board DRAM because the benefit is not worth the cost - especially if it means giving up one of the main reasons on package system memory is happening at all - lower idle consumption. As far as I know, every consumer design with on package system memory has been a mobile first SOC/SIP; Apple’s A/M series, and intel’s Lunar Lake. One of the main reasons they do it at all is because they think people will pay for battery life.

And in fact, someone at Intel gave an interview stating they won’t do on package again after Lunar Lake because OEM/ODMs didn’t like the increase in BOM.

Talking about changes over 20 years is pointless; in 20 years we may not be using silicon or DRAM at all. I won’t speculate what computing looks like then. But right now a memory controller supporting multiple hierarchies of DRAM plus all the ancillary costs to enable DRAM signal integrity on PCB all to enable user replaceable ram with LPDDRNX performance to first level system memory won’t be viewed as economically viable by any of the major chipmakers, and they’re probably right in that assessment.

More importantly, framework can’t consider using a chip that doesn’t exist. Unless they intend to spin up a semiconductor design team and spend tens of millions buying wafer allocation from TSMC, Samsung or Intel, the only options they can consider are what existing chipmakers are willing to sell them, and they are absolutely not big enough to get semi-custom solutions like Apple used to get from Intel.

Then we’re on the same page.

Point taken. But that’s only because you’re operating under the pre-condition that both operate with similar performance…and therefore no extra benefits. For example, if they put DDR7 on package, and DDR5 on board…that changes the balance of things.

Your point of view is tied-in with current memory chip technology. My point of view is tied to memory / [data] storage accessibility hierarchy. A strategy that’s been around humans for tens of thousands of years in the physical world before even being leveraged into silicon. The fabrication technology / mechanism is not my point here.

You think? They kind of can (in the future)…if/when Framework gets big enough to be clued in on chips / product roadmap from Intel / AMD / Qualcomm…etc…for ‘consideration’.

Nature of the poll is about “understand how the framework community feels about a framework mainboard with soldered LPDDR RAM.” in conjunction with “future of framework laptop mainboards”.

Anyway, we don’t have to impose how we ‘feel’ onto each other.

Also, just want to point out that when it comes to upgradable components in terms of adding more chips, that’s pretty much almost always going to require more power. It’s inherent to the upgrade action.

So, if that’s the argument…then you’ve already framed the answer into the argument that no one should upgrade any components.

Incorrect, I’m not arguing that components shouldn’t be upgraded at all; if you read my previous, original post you’ll see that I favor using lpcamm over soldered ram, even at the cost of battery life. I am probably more willing to trade away battery life than most, even.

I was merely pointing out that your hypothetical non-existent consumer SOC/SIP with both soldered and socketed RAM to get the best of both worlds doesn’t actually achieve the all the benefits of both, especially given one of the primary stated objectives of using soldered RAM in the original poll question was better battery life.

With current chip design you cannot really sidestep the tension between upgradability and idle consumption - how much of one you’re willing to sacrifice for the other depends on your specific priorities, which is actually the basic premise of this entire poll - are you willing to sacrifice upgradability for battery life, performance, and more uarch options.

Where are you getting this need for / mention of “all the benefits of both” from?

If you design the chip with mandatory memory on the chip, and allow for optional on-board memory, then it’s up to the consumer to design if they want to add any on-board memory, with trade-off in battery life…for a larger memory pool, but slightly slower (but still faster than SSD).

Plus, energy savings, to my knowledge was not the primary driver for memory on the package. It just happens to be the aftermath, after performance and signal integrity, but I could be wrong. It’s not clear how much energy saving is a consequence of the memory being on the package.

We, by nature of upgrading memory, already gives implied consent use more power. e.g. If you could upgrade / swap one M4 chip to another with higher on-package memory…you would consumer more power.

So, if I were to go with this non-existent chip:
Day 1: Fully on-package memory only.
Day 2000: Fully on-package memory, plus additional memory on the board.

This would consume more power, at my choosing.

I could use it as a desk unit by then…who knows. The point is, I don’t want to buy a feature PC (as if it were a feature phone). Also, similarly, I don’t want SSD to be soldered on. I want to have upgradable storage as well. Also secondary SSD as well. It’s all along the same line…of information / data storage and accessibility to the processor. (just a matter of the medium being volitile or not, and how fast, how close to the processor)

People upgrade their storage, have battery consumption impact. Add secondary storage, have battery consumption impact. Upgrade the OS / software, have battery consumption impact. Upgrade display panel from IPS to OLED, you guess it, battery consumption impact. Don’t hide behind “It’s energy savings” to remove upgradeability ‘option’ away from the choice of each end-user.

The choice to upgrade or not should be with the consumer. So, really, I much rather to see a trend / direction where “Sure, here’s on-package memory”, but also with “We also allow for on-board memory if you need that”.

I’m done with this topic for now. It’s wearing me down.