eGPU’s, in the way that your average person might google “eGPU” and consume the first few entries to understand what they do and how they work… THAT is not possible on Apple Silicon, though. eGPU’s, the way they currently work on Intel portables is not possible on Apple Silicon. Has never been possible, and would never be possible because Apple would never make the changes required for it to be possible.
See, this is an example of where people like Hector (and now, myself) get into trouble. You've seemingly integrated an untrue idea into your worldview - that Apple Silicon hardware cannot talk to an eGPU at all - and that makes it very difficult to explain anything to you.
Found a thread I was following where (if that was him) he was asked if eGPU’s were possible, like, in the same way they are used connected to Windows laptops. It appears his problem is (well, WAS) that he’s very unequivocally Yes on “possible” and then gets wishy washy around whether or not it would be implemented. Even when someone provided a very clear use case that he knew was NOT true for Apple Silicon (he’d already listed reasons why it wasn’t true for Apple Silicon
) he could not bring himself to type that that very narrow use case would not be possible. So, there was
something else there beyond just poor communication and he’s probably better off away from the project.
Apple Silicon hardware can in fact talk to a an eGPU. Discrete GPUs are just PCIe devices, and AS support for PCIe is fully generic. It can map any PCIe device's collection of BARs (base address registers, descriptors for the card's memory-mapped IO resources) into Apple Silicon's physical memory address space.
The one thing Apple Silicon
can't do with these MMIO resources (that x86 PCs can) is allow its CPUs to perform misaligned accesses to PCIe device memory. If you don't know what a misaligned access is, it's reading or writing an object of size N bytes (where N is a power of 2) when the address of the object in memory is not divisible by N.
For nearly all types of PCIe device, that's fine, you never needed misaligned accesses anyways. But when trying to talk to AMD and NVidia GPUs, misaligned accesses become important. Mind you, they do not
need to be important. You can always structure data in GPU memory as fully aligned, and the GPU can work with that. However, there's a very large library of software out there written against AMD and Nvidia GPUs (both drivers and application software) which assumes it can treat mapped GPU memory just like regular RAM, implying that misaligned access is available and costs next to nothing. That's where the problems come from.
This is why, when you ask someone like Hector "is it possible?", he is going to say
yes. That really is the only honest answer. Effectively, it is "just" a software problem: if you could fix every bit of software that ever touches PCIe GPU memory to always align every access, it would all work great and performance would be good.
So sure, you can find plenty of things Hector wrote that you see as "wishy washy" and evasive, but are really just him trying to explain that while it's technically possible, it's an enormous task for a small team. Worse, it would touch lots of projects they'd have difficulty getting to go along with it (including but not limited to closed-source software). Or sometimes he talked about workarounds he could imagine which don't require patching everything in sight, but they'd have a performance cost that probably wouldn't be acceptable.
