Shared memory is different to unified memory, AMD’s got an implementation of the later with their “Ryzen AI MAX+” (ugh) systems, does quite well in benchmarks.
It also doesn’t hurt that Apple puts the RAM on the SoC and gives it a truckload of bandwidth. DDR5 is about 70GB/s, meanwhile the M4 Max is around 540GB/s.
I didn’t know AMD had managed to switch over to unified memory too. Managing that while remaining x86 compatible is quite an achievement!
I think the next big thing will be when storage becomes as fast as ram and they unify that too, getting rid of separate RAM. Working with data directly in place could have massive efficiency boosts. But the industry has been trying to get it that fast for many years and still not succeeded. And once they do, separate SSDs wouldn’t be possible, at least not as a primary storage, so it wont be an advance that makes sense for every use case.
Yeah “universal memory” is the holy grail, seemingly as hard to find as it as well.
The articles on Wikipedia about the related tech is great, it’ll mention something like “Developers expect commercialisation to happen relatively soon” and then link to an article from 2004, or research papers from the 1980s.
Shared memory is different to unified memory, AMD’s got an implementation of the later with their “Ryzen AI MAX+” (ugh) systems, does quite well in benchmarks.
It also doesn’t hurt that Apple puts the RAM on the SoC and gives it a truckload of bandwidth. DDR5 is about 70GB/s, meanwhile the M4 Max is around 540GB/s.
I didn’t know AMD had managed to switch over to unified memory too. Managing that while remaining x86 compatible is quite an achievement!
I think the next big thing will be when storage becomes as fast as ram and they unify that too, getting rid of separate RAM. Working with data directly in place could have massive efficiency boosts. But the industry has been trying to get it that fast for many years and still not succeeded. And once they do, separate SSDs wouldn’t be possible, at least not as a primary storage, so it wont be an advance that makes sense for every use case.
Yeah “universal memory” is the holy grail, seemingly as hard to find as it as well.
The articles on Wikipedia about the related tech is great, it’ll mention something like “Developers expect commercialisation to happen relatively soon” and then link to an article from 2004, or research papers from the 1980s.