[tadfisher]

Soldered LPDDR5x has more bandwidth, lower latency, and consumes less power than the equivalent DDR5 SODIMM. It can also be clocked higher due to signal integrity constraints, and you really want higher clocks with a Ryzen SOC.

If you plan on maxing-out the memory, the only advantage of SODIMMs I see are in repairability. That said, I have verified exactly one bad stick of RAM in over 3 decades of computing, so I personally think the downside of soldered RAM is overblown.

[LeonenTheDK]

For me it goes beyond repair, if I sell the device later and the next user wants more RAM, they have the option to do so. Plus bringing ram forward between mainboard upgrades (save for transitions between DDR4, DDR5, and whatever else may come in the future).

I remember now too Dell was working on a new memory module for laptops called CAMM, I think it was aiming to bridge the gap between soldered and SODIMMs, but I'm not sure where its development currently is, nor the real world differences.

[freedomben]

Yes, I have carried RAM forward to new machines twice now. It's a huge benefit to me and worth the tradeoffs.

[dmm]

> the only advantage of SODIMMs I see are in repairability.

Another advantage is that you don't have to buy ram from the OEM. I can buy 64GiB of ddr5 for <$150. Apple wants $400 for 64GiB.

[linguae]

Non-soldered RAM also means that I can also take advantage of the general trend of declining prices instead of having to pay top dollar for maximum configurations upfront. For example, 32 GB is plenty for a laptop for me in 2023, but my needs may change in 2025 or 2026, especially as memory requirements tend to increase over time. Thus I could get 32 GB of RAM today and then upgrade to 64 GB in a few years, when it would probably be cheaper. I’ve done this many times before: my 2006 MacBook shipped with 512 MB of RAM and a 60 GB hard drive. Three years later I upgraded to 2 GB of RAM and a 320 GB hard drive, and the price I paid for these upgrades was much lower in 2009 than it would have been in 2006, especially if I had to get the upgrades from Apple. Sadly being able to perform gradual upgrades hasn’t been an option for Apple laptops in a long time, and this practice is spreading to PCs.

[volkandkaya]

One of the main reasons I didn't buy a new Macbook pro.

[Leimi]

yeah I get the benefits. And yeah the "repair" argument I agree, good luck finding a bad stick :)

The real frustrating thing as for now in the real world is, there is an extremely low number of laptops with soldered ram that offers 64 GB. And the few that do, charge an absurd amount of money for it.

With socketed ram, I can:

- buy the cheapest built-in config of a laptop

- then buy the RAM I currently need on my own, often saving a few hundreds bucks just doing that

- then, in a few years, buy some new RAM again, when I need it, if I need it, instead of having to buy a whole new laptop.

That's how I went with thinkpads during 15 years. Now I have to pay 500$ more to be a bit future proof. If the manufacturer offers it. Double that if you want a mac.

So, still today, I'm 100% taking socketed ram instead of soldered one.

[thomastjeffery]

I can't imagine actually using 64GB of RAM on a laptop.

How many concurrent VMs would I ever want to run anyway?

If I ever go over 16GB, and actually notice it swapping on an NVMe drive... why not just remote to a real workstation?

[Nullabillity]

I can't imagine living with a workflow that regularly requires me to remote into servers to get shit done. Life is miserable enough as it is.

[thomastjeffery]

  With a little setup it's no more effort than logging in. Less, even.

[pacifika]

Muscle memory closing apps by keyboard shortcut means the remote session is closed. Also multi monitor support is not as good in my experience

[thomastjeffery]

> Also multi monitor support is not as good in my experience

Are we still talking about a laptop?

[Leimi]

I guess we use our machines in different ways :)

And I have a few use cases where 32 Gb is limiting. So I don't want to buy a brand new machine stuck at something that is, sometimes, already not ideal. And well, the usual next step is 64 Gb.

[sireat]

I routinely use 64GB on one of my laptops.

Load a "largish" dataframe you will wish for even more RAM.

Even on smallish dataframes with say 20-30GB usage it can be convenient to have a copy near by.

[thomastjeffery]

I have no clue what you are talking about when you say "dataframe".

[RulerOf]

16 gigs for the Docker VM

24 for chrome

8 for electron

10 for OS caching

4 for shared video memory

...and 2 that sit unused to let you know that you don't need any more to get the job done.

[ksec]

Modular LPDDR Memory Becomes A Reality: Samsung Introduces LPCAMM Memory Modules [1], fixes most of the problem. The only thing that soldered LPDDR ram offers is saved motherboard space. Which really isn't as much of a problem in modern laptop when SoC and GPU are so well integrated.

"Samsung says that LPDDR5X LPCAMMs only occupy 40% of the space of a DDR5 SO-DIMM, and improve power efficiency by up to 70%, roughly in line with the general benefit of LPDDR5X over DDR5."

[1] https://www.anandtech.com/show/21069/modular-lpddr-becomes-a...

[justinclift]

Hopefully something like CAMM gets traction, as it seems like an improvement over SO-DIMM:

https://www.extremetech.com/computing/342342-jedec-says-dell...

[sva_]

> That said, I have verified exactly one bad stick of RAM in over 3 decades of computing, so I personally think the downside of soldered RAM is overblown.

Man I must've been unlucky. I had a dead DDR stick around 10 years back and LPDDR (soldered on a ThinkPad x1) die on me around 6 years ago. Both of them died very early on though and were covered by warranty.

[m-p-3]

I had to send some bad RAM under warranty as well. Another upside with socketed RAM is that I was able to buy a small stick of RAM and get my machine going (although with constrained memory) until I received the new pair, which was a minor inconvenience.

Having soldered RAM would have meant shipping the entire system back, then find a spare system and the trouble of setting up all my softwares until then.

[volkandkaya]

" has more bandwidth, lower latency, and consumes less power than the equivalent "

Any data points for the above?

Is it worth the ability to not replace them?

If so look forward to a marketplace of Framework laptops where you can easily sell your old model, and buy a new one.

[IshKebab]

Surely the main benefit is cost? RAM size (and disk size in Apple's case) are used for price differentiation. If you can replace them yourself they can't charge you through the nose for bigger sizes.

On a technical level I think you're right.

[tjoff]

... and the 10x price difference, the waste and crippled hardware, the pathetic pricing options you get. Soldered ram has no place nor reason to exist in sad reality we live in.

[huijzer]

> Soldered ram has no place nor reason to exist in sad reality we live in.

Doesn’t it have performance benefits since it can be located closer to the processors?

[prmoustache]

But is that performance benefit worth the non replacability/upgradability.

Looks to me like laptops are fast enough these days and have been for more than a decade. I have 3 laptops, one from 2011, one from 2017 and one from 2019.

Even the oldest one is fine for everything. If I had specific tasks that needed more comoutation speed it would be make more sense to use a remote server than replace it.

[huijzer]

That is not my experience at all. My Intel laptop bought new in 2023 took 40 minutes to compile LLVM, would get extremely hot and loud, and the screen would hang multiple times during the process. This laptop’s fan would even spin when watching YouTube. My current M1 chip takes only 20 minutes without fan or freezing screens. I’ve looked it up and the M1 chip requires only half the clock cycles to get data from memory and I suspect that does play a role in the dramatic difference in performance and energy efficiency.

[prmoustache]

I am not a gentoo user.

[tjoff]

Sure, but that is one very small consolation.

[topspin]

Soldering has benefits, but RAM is likely the most important thing that will need upgrading in extended service of a system.

[thomastjeffery]

> I personally think the downside of soldered RAM is overblown.

It is now, but only because it wasn't before. Narratives are slower to change than their subjects, especially in tech.

When soldered RAM started to take off, it was usually 1-2GB, which, even at the time, was a painful compromise. Even a lightweight Linux distro running a browser (i.e ChromeOS) will feel the limits of 4GB.

Now, most laptops have at least 8GB, which is good enough for most. 16GB is plenty unless you have some specialized workload that actually uses more, like compiling a large codebase or video editing.

64GB on a laptop is absurd. Whatever workload you have that needs that much memory should almost definitely run on a remote server anyway.

[alluro2]

8GB is hardly good enough for most, with proliferation of Electron apps everywhere and browsers being such memory hogs.

I am easily utilizing all of my 32GB just normally running a browser, email client, Slack, database mgmt tool, a couple of light servers (e.g. NodeJS, Redis for local dev), VS Code, Figma and a Git client. I would definitely get 64GB for my next work laptop, and I don't consider myself special in my needs in any way as a developer, nor would anything significantly change if I used remote servers.

So your generalization might be a bit off.

[prmoustache]

I have yet to see a use case where I need and can't replace a shitty electron app with a native one or its web equivalent running in a browser tab or window.

I have a laptop with 32GB, the only use case where I was maxing out the memory was when I was running a 6 nodes kubernetes cluster in vagrant virtual boxes.

And that was kinda stupid because I could have run those VMs in a remote hypervisor.

[thomastjeffery]

> I am easily utilizing all of my 32G

But if you only had 16GB, would you notice?

Browsers will happily use the RAM you give them, because they are caching everything they touch. With any relatively new NVMe SSD, that cache is probably close enough.