Poll on soldered vs socketed RAM

I know we used to have that, but that was back when we measured memory in bits and Kb rather than GB. Small upgrades made a huge difference back then! I just think if I’d be able to upgrade my vRAM on my old GTX 1080 I could probably stretch it a few more years. :laughing:

Also, there’s a much lesser incentive to socket the GPU’s VRAM. On desktop computers, the GPU itself is socketed and it comes with its own VRM and VRAM. For each different GPU model, there are differences of power and VRAM requirements so it’s better to install them as part of the GPU rather than on the mainboard. On laptop computers more than half of them favor portability so no dGPU hence no GDDR

LPCAMM2 seems like the best of both worlds here. You get the lower power draw and higher memory bandwidth without permanently giving up upgradeability. A few people here already mentioned it, and I think it deserves more attention. Hopefully Framework leans into LPCAMM2 more aggressively going forward rather than jumping straight to fully soldered RAM (desktop). Repairability is the whole reason I chose Framework in the first place.

2 Likes

Gaming consoles use GDDR with their APU, sacrificing CPU performance because of latency, but boosting the iGPU performance.

It’s really disappointing the PC market has APUs with DDR controllers, but no APU with GDDR controllers.

Even more amazing, would be an APU with 128b of DDR controllers, and 128b of GDDR controllers. It makes the memory bus much harder to sync but it would allow to maximize CPU and uGPU performance, it would be a killer offering.

Another amazing product would be an APU with a wider DDR controller. Server CPUs those days get up to 16 channels of 64 bit of DDR5 memory (1024 bits). Regular CPUs have just 2 channels of 64 bit of DDR5 for 128 bit.

The AI Max 395 and some AI boxes have 4 channels at 256 bit bus of DDR5, it’s a product that demands having two 128b LPCAMM2 modules.

Even better you can imagine a APU with four 128b SOCAMM2 modules at 512b bus width to really push the bandwidth up.

In short, memory configuration has lots of room of improvements, and we are mostly stuck to dual channel DDR since forever.

Screw gddr, just go full hbm2 at that point XD

Would love to see high speed ARM chips and maybe even RISC open hardware CPU availability eventually