Hacker News Reader

Comment by wpwpwpw | Hacker News Reader

If a CPU is socket-able, so should memory chips be.

It won’t ever be able run as fast as a soldered system.

Have you installed a server CPU?

It’s really easy to fuckup and lose a few channels of memory due to the contact being bad. Right now I’ve got a 3647 Xeon phi cpu that’s refusing to train dimm a1 for _reasons_

That’s not an experience Apple wants any user of their products to have.

Here’s an example BGA socket: https://www.ironwoodelectronics.com/products/bga-sockets/

Not something that’s going in a tiny laptop chassis.

People running a single desktop machine are way better served by being able to upgrade RAM modules than worrying about single contacts being bad between the RAM stick and motherboard.

Yes? I wasn’t making a claim that it was better to solder everything for everyone. I’m saying the overlap between most Apple users and those people is low.

Give definitive evidence that soldered is faster, in my experience with decent contact this is not true at all. I think the general confusion is around typical ram sticks with a controller onboard and much fewer io broke out compared to the raw ram without controller which is often what you get soldered, with more io broken out.

The bga socket you chose is more for test or industrial hardware, versus desktop cpu sockets which are much slimmer and consumer friendly.

Really disingenuous, imo. It's absolutely possible for Apple to make these chips replaceable, using the heat spreader as the retaining plate.

Sure, they are. Buy a Epyc or Xeon, burns 200+ watts, has 8 to 12 channels, which requires 8 or 12 dimms, which are barely fitting in 19" racks (next gen are moving to 21").

Or you could get a m3 max, run the memory at twice the speed, still have a 512 bit wide memory bus, and have a 10+ hour battery life. Presumably similar with the m4 max, rumors claim later today (Wednesday).

How much do you want that socket?

Apple puts the memory, CPU, and GPU all on the same chip. This generates less waste as you only need a single package and socket, and uses less energy during operation.

And many desktops do that today, but like everything it has tradeoffs, such as peak bandwidth and power usage. DDR sockets inherently make this sacrifice, integrated designs will always have wider buses, higher bandwidth, etc. That's also why you don't get sockets for your GPU memory, either. It's a design tradeoff.

-> It won’t ever be able run as fast as a soldered system.

Yeah, just take a look at PCIe 5 and it's 512GB/s of bandwidth.

-> Have you installed a server CPU?

Yeah, and none of the problems you mentioned.

-> That’s not an experience Apple wants any user of their products to have.

Yeah, just look at the older macs with upgradable components and the easyness you had replacing them... So, instead of making it easier, let's just remove it altogether.

PCIe is a serial interface, not parallel like modern DRAM interfaces. They're completely different at a hardware level, the electrical design constraints are completely different, the latency characteristics are completely different. I think you are just throwing words and numbers out and don't really know what they mean at all.

Reply on news.ycombinator.com