Samsung and SK Hynix rumored to boost AI accelerator performance with the advent of the HBM4 DRAM standard

[Redaktion]

HBM4 could offer doubled 2048-bit bus width and a memory bandwidth of up to 2.3 TB/s if Samsung and SK Hynix manage to maintain the same number of memory stacks from the current HBM3E configurations.

https://www.notebookcheck.net/Samsung-and-SK-Hynix-rumored-to-boost-AI-accelerator-performance-with-the-advent-of-the-HBM4-DRAM-standard.749048.0.html

[NikoB]

Regular PCs/laptops have been required to have at least 200GB/s for several years now, but what do we see? Here it's 2.3Tb/s, and x86 is a shameful 50-80GB/s...

8-16 cores are already suffocating with such slow RAM. It's not for nothing that Intel/AMD developers are shamefully adding a crutch in the form of the L3 cache more and more, not being able to cope with the main task - increasing the RAM bandwidth significantly, at least to the level of 200GB/s in mainstream PCs/laptops, not to mention about 400GB/s, like a 512-bit memory controller in the Apple M2 Max 2022...