[cubefox]

I wonder whether (without soldering) replaceable GDDR RAM on graphics cards would really result in such a large speed penalty. System memory often isn't soldered to the main board either, and this doesn't seem to have large performance penalties.

[Telaneo]

The RAM in your desktop could run at much higher frequencies if it were soldered. The RAM the M5 Macbooks use runs at 9600 mbps. The base spec for DDR5 is 4800, and I rarely see people running at much higher than 6000. There's probably someone who's gotten a 9600 mbps overclock running on their desktop, but it's unatainable for a normal consumer outside of machines with soldered RAM.

[cubefox]

This doesn't explain why non-soldered RAM isn't okay for graphics cards when it is still considered perfectly acceptable for desktop computers.

[Telaneo]

Because GPUs have had it soldered since the beginning, while desktops have had it slotted since (close to) the beginning. There are some exceptions to that rule, but it mostly holds.

Not to mention that the bar of acceptibility is moving. We had RAM in slots in laptops for a long while, but it's (not so) slowly going away. You can still find laptops with SO-DIMMs, but it's not a given anymore.

[Nursie]

A longer time ago, sound cards used to have expandable RAM too.

Go back far enough (486) and cache was off-chip on simms too!

[inigyou]

Yep. Soldered is just the best manufacturing process (besides on-package). Moving things close together simply delivers better outcomes and that will always trump openness concerns.

[theodric]

I would imagine that something like LPCAMM2 for GDDR is entirely possible and will never, ever be seen because people already accept replacing an entire $10000 GPU to get another 32GB VRAM on the next generation, so why give up that guaranteed revenue stream?

[gosub100]

Because you can't sell more graphics cards if consumers get more use out of them.

[jve]

Perhaps it impacts size, manufacturing (automation) and money considerations. I mean now you suddenly need a slot of golden plated pins, you suddenly need a module with said pins and a memory chip soldered. You're introducing so much steps and complexity (thus expense) there for net negative.

I know that people like to argue "but I can't upgrade RAM". Yes. But as we see here not impossible. Your car also needs some effort to change some parts and perhaps even more effort compared to resoldering some chips at a shop that knows how-to.

[cubefox]

> Perhaps it impacts size, manufacturing (automation) and money considerations. I mean now you suddenly need a slot of golden plated pins, you suddenly need a module with said pins and a memory chip soldered. You're introducing so much steps and complexity (thus expense) there for net negative.

How is this relevant when the same point applies to system RAM?

[butvacuum]

Desktop DDR rarely exceedes 64bits(you have to combine more than one dimm). Since RAM is a parallel bus that means 64 conductors that have to be impedance and length controlled. GPUs use >192bit busses, with 512bit in the high performance GPUs. So that's now 512 conductors.

512 conductors that have to be exactly the same length. Make them longer you'll say- now you just lost your high memory clocks to propigation delay and parasitics.

Seeing the problem?