[klardotsh]

Nearly all modern SOCs have built in RAM now. Apple Silicon does it, AMD Strix Halo and beyond do it, Intel Lunar Lake does it, most ARM SOCs from vendors other than Apple do it…

Now, unified memory shared freely between CPU and GPU would be cool, like Apple and AMD SH have, if that’s what you meant.

[wtallis]

AMD Strix Halo does not have on-package RAM. What makes it stand out from other x86 SoCs is that it has more memory channels, for a total of a 256-bit wide bus compared to 128-bit wide for all other recent consumer x86 processors.

Qualcomm's laptop chips thus far have also not had on-package RAM. They have announced that the top model from their upcoming Snapdragon X2 family will have a 192-bit wide memory bus, but the rest will still have a 128-bit memory bus.

Intel Lunar Lake did have on-package RAM, running at 8533 MT/s. This new Panther Lake family from Intel will run at 9600 MT/s for some of the configurations, with off-package RAM. All still with a 128-bit memory bus.

[notenlish]

Isnt Strix Halo and apple's m series a bit different. Iirc you need to choose how much ram will be allocated to the igpu, whereas on mac it is all handled dynamically.

[Tsiklon]

On the Mac it's all dynamically handled.

With Strix Halo there's two ways of going about it; either set how much memory you want allocated to GPU in BIOS (Less desirable), or set the memory allocation to the GPU to 512MB in the BIOS, and the driver will do it all dynamically much like on a Mac.

[einsteinx2]

You can also change it manually on macOS [0], though yes by default it’s automatic but has a much lower limit (I think 2/3 is total RAM or something).

[0]: https://github.com/ggml-org/llama.cpp/discussions/2182#discu...

[adgjlsfhk1]

strix also does it dynamically, just with a limit (which is generally set to ~75% of your total RAM)