i know very little about this topic so maybe this is completely retarded, idk. it just seems weird that even integrated CPU-GPUs logically separate the two components so you have to transmit data back and forth with e.g. glBufferData
bonifartius 𒂼𒄄
@velartrill gpus are heavily parallelized simd architectures. they are more like batch jobs of old times. you load a bunch of data, press start and then retrieve the results.
@velartrill i think it just doesn't fit the bill. i guess you could add it as a submodule. it just isn't useful for many things except graphics or number crunching. essentially it's just a bunch of alus.
futhark-lang.org abstracts over cpus and gpus, it looks nice but i don't have any use for it currently.
i'd really like to work on something heavy parallelizeable, it's currently only used for things like KI and cryptomining because nobody (i know of) really bothers to design new algorithms.
@velartrill @bonifartius My understanding is that there's a large overhead in loading data back and forth between the main RAM and the GPU RAM. So would the shared memory of an integrated GPU eliminate that?
They're more compact which makes them a better choice for portable devices, but that's about it.
@jleedev sure, but i don't know how feasible it would be to put several GB of RAM onto the CPU die.
Not necessarily, CPUs and GPUs are well suited to different memory configurations.
