Isn't X usually the original source these days?

Chicken/egg. NVidia tooling is lacking surely in part because the hardware wasn’t usable on macOS until now. Now that it’s usable that might change.

I misunderstood eGPU for virtual GPU. But I was wrong it means external GPU.

The government doesn’t care? They’re a minority of the market? The vast majority of their computers didn’t have slots to put Nvidia GPUs in, and now none of them do?

Apple doesn’t have a monopoly in any market they are in.

Evidence that NVIDIA has even been trying? My understanding is that Apple didn’t allow 3rd parties to write graphics drivers past 10.13, but they could’ve done a non-graphics driver like this.

The same way Google evaded regulatory scrutiny for refusing to allow a YouTube client for Windows Phone?

Isn't all you have to do disable SIP?

https://docs.tinygrad.org/tinygpu/ are their docs, and https://github.com/tinygrad/tinygrad/tree/4d36366717aa9f1735... is the actual (user space) driver.

My read of everything is that they are using Docker for NVIDIA GPUs for the sake of "how do you compile code to target the GPU"; for AMD they're just compiling their own LLVM with the appropriate target on macOS.

I mean when it comes time to output the image from the GPU, I don't want to add a hundred milliseconds of network latency...

You can buy a cheap GPU enclosure for about 100$ off ali express.

Takes a standard PSU. However, Mac Minis don't have occulink. So you might be a bit limited by whatever USB C can do.

Now if Intel can get there Arc drivers in order we'll see some real budget fun.

https://www.newegg.com/intel-arc-pro-b70-32gb-graphics-card/...

32 GB of VRAM for 1000$. Plus a 500$ Mac Mini.

Maybe I’m lacking imagination. But how will a GPU with small-ish but fast VRAM and great compute, augment a Mac with large but slow VRAM and weak compute? The interconnect isn’t powerful enough to change layers on the GPU rapidly, I guess?

“Lying around”. I’ve got an unopened 5090 in a box that I know will suffer the same fate, so I’m sending it back. So privileged to have the money to impulse buy a 5090 and yet no time to actually do anything with it.

Doesn't Apple support the major standard device categories: NVMe, XHCI, AHCI, and such, like most operating systems do? The challenges are all for hardware that needs a vendor-specific driver instead of conforming to a standard driver interface (which doesn't always exist). Lots of those can be supported with userspace drivers, which can be supplied by third parties instead of needing to be written by Apple.

> Why does Apple need to make the drivers in a walled garden?

Isn't that the whole point of the walled garden, that they approve things? How could they aim and realize a walled garden without making things like that have to pass through them?

> Why does Apple need to make the drivers in a walled garden?

For the same reason that Microsoft requires Windows driver signing?

Drivers run with root permissions.

Macs and PCs are fundamentally different. Their architectures have always been distinct though the Intel Mac era has somewhat blurred the line.

Modern Mac is Macintosh descendants and by contrast PC is IBM PC descendants (their real name is technically PC-clone but because IBM PC don’t exist anymore the clone part have been scrapped).

And with Apple silicon Mac the two is again very different, for example Mac don’t use NVMe, they use just nand (their controller part is integrated in the SoC) and they don’t use UEFI or BIOS, but a combination of Boot ROM, LLB and iBoot

> Why does Apple need to make the drivers in a walled garden?

Because third party drivers usually are utter dogshit. That's how Apple managed to get double the battery life time even in the Intel era over comparable Windows based offerings.

Don't purchase? I don't own any Apple devices, everything works fine.

Well, to be fair, the whole shebang is from a completely different company, that have their own ML library and such, so that isn't that surprising. Although I agree that some CUDA shim or similar would be a lot more interesting, still getting to the place of running inference and training with your very own library is pretty dope already.

As a sizable share of the market is going to want to use this for local LLMs, I do not think this is that misleading.

GPUs can do graphics too?

Graphics was not what came to mind when I saw the headline.

This driver doesn't support CUDA.

tinygrad does have pretty neat support for sharding things across various devices relatively easy, that'd help. I'm guessing you'd hit the bandwidth ceiling transferring stuff back and forth though instead.

I can speak to my own experience, YMMV

I hooked up a Radeon RX 9060 XT to my Feodra KDE laptop (Yoga Pro 7 14ASP9) using a Razer Core X Chroma (40Gbps), and the performance when using the eGPU was very similar to using the Radeon 880M built into the laptop's Ryzen 9 365 APU.

So at least with my setup, performance is not great at all.

On paper, TB4 is capable of pushing 5GB/s, which is somewhere between 4x and 8x of PCIe 3.0, while a 16x PCIe 4.0 link can do ~31.5GB/s.

For numbers about all PCIe generations and lane counts, see the "History and revisions" section here: https://en.wikipedia.org/wiki/PCI_Express

Edit to add: the performance I measured is in gaming workloads, not compute

Well, for starters, PCIe 5.0 x16 would do something like about 60 GB/s each way, while Thunderbolt 4 does 4 GB/s each way, TB 5 does 8 GB/s each way. If you don't actually hit the bandwidth limits, it obviously matters less. Whether you'd notice a large difference would depends heavily on the type of workload.

It carries pcie, but only at x4. Thunderbolt 4 is pcie gen 3 and Thunderbolt 5 is pcie gen 4.

They... do? Or rather, they built a system where they don't need to; macs happily run Linux on bare metal or VMs. (Whether Linux supports Apple hardware well is another matter)

Apple got out of the server game long before they adopted aarch64, so that's a trillion worth of server hardware they never would have sold anyway. And probably not actually a trillion.