3 ms·
> "Fundamentally, we've built our architecture around this shared memory model and that optimization, and so it's not entirely clear to me how you'd bring in an
by smoldesu 3y ago
> "Fundamentally, we've built our architecture around this shared memory model and that optimization, and so it's not entirely clear to me how you'd bring in another GPU and do so in a way that is optimized for our systems,"
So... why expose PCI lanes in the first place, then? "Optimization" isn't really an excuse, as long as you can saturate the PCI bus it should run smoothly. Some support would be better than nothing, especially if you're brave enough to pay $7,000 for a Mac Pro.
This whole thing is kinda baffling. Surely they haven't gotten rid of the multi-GPU address space support, they still support Intel/AMD systems after all.
- pavlov 3y agoI think quite a few people would be happy to be able to put an Nvidia card in one of those Mac Pro slots and use it through CUDA only, as a compute/ML accelerator. But Apple doesn't want to allow that either because they see CUDA as a competitor to their own APIs. So they do this whole spiel about the indivisible beauty of their own shared memory architecture.
- pyrolistical 3y agoI don’t get how Apple is going to block nvidia from releasing cuda only drivers?
- fzzzy 3y agoNvidia hasn't released a Mac driver in years due to Apple's dislike for them, so they effectively already did block them.
- shrubble 3y agoWhat I don't get is, why the lowered RAM limit? Wouldn't being able to run a 512GB model on the GPU be a big selling point?
- jitl 3y agoI think it’s a symptom of the same system design - Apple needs the RAM on the chiplet or as close as possible, because it’s shared with the GPU and wiring the GPU up to standard DIMMs all the way out on a motherboard would yield unacceptable latency. These are essentially phone CPUs that got super super beefy - they’re limited by what they can fit in their chiplet/system-on-chip. There’s no allowance for heterogeneous hardware access like a second tier of RAM via a motherboard northbridge or something. I think they’ll get to those higher RAM densities by M4 or M5 once they have more die space to spend on memory controllers.
- villgax 3y agoFor M5 maybe