28 comments

[ 5.3 ms ] story [ 68.1 ms ] thread
Won’t bother asking for Laptop/Optimus support
Only school Optimus with a MUX chip used to work just fine on my T420. For a long time the newer Optimus streaming through the (Intel) iGPU wasn't supported transparently, but you could use indirect OpenGL using environment variables. I once used this for an application that required OpenGL extensions lacking in the open source Intel driver at the time. It's annoying and fragile. If you can let go of the battery runtime always running the dedicated (Nvidia) GPU is preferable.
Been using Ryzen + Nvidia on KDA plasma for the past few months and it works quite good; generally, it works better on X.org, tough Wayland isn't that far behind either.
It has been updated frequently in the past: https://www.freshports.org/x11/nvidia-driver

(scroll down for commit history)

edit: Has Nvidia been building a FreeBSD driver for the past 20 years, or did they initially port a *nix driver from a different OS to FreeBSD?

nVidia uses a common core for all their drivers so FreeBSD code is essentially wrapping that.

This is is where the Linux griping actually come from - they don't rewrite their driver when kernel decides to rethink their graphics stack.

They're even moving most of the kernel-space driver code to the chip on-board the GPU now (see GSP) and having the kernel driver talk to that.
Pretty soon NUCs and RPis with external GPUs will be common.

I look forward to never fitting those fiddly power LED/switch wires ever again.

Tons of nostalgia for having been there, done that but that’s all it is; fiddling with nostalgia. It’s a hardware dark pattern.

They should better have released a NetBSD driver - people would be able to port it to all the alternative OSes then.
How would that make a difference?
As far as I know, the NetBSD driver odel is so good anyone can port a NetBSD driver to any other OS. This means it is optimal to make NetBSD drivers so you effectively cover all the OSes by letting enthusiasts and backer companies to port wherever the need is, even to exotic OSes like Haiku.
This is a binary release. There's no source that can be ported by hobbyists.
True test of LLM based coding would be to enable Nvidia cards via eGPU on Apple Silicon using this or Asahi Linux work for ML/DL tasks. Would allow for separate upgrade paths for GPU compute & main device instead of having to keep another PC just for local experiments
Why wouldn't the eGPU work on Apple hardware? It's just thunderbolt. You'll be running the same setup as a laptop with an Nvidia GPU built in as long as Thunderbolt works correctly.

If Asahi Linux gets Thunderbolt support to work (or if Apple releases a Linux driver for their hardware, which will never happen I suppose), eGPU support should arrive almost instantly.

If you want to run ML workloads in this manner, any existing laptop or tablet with Thunderbolt will do, though.

Only Intel is support officially - https://support.apple.com/en-us/HT208544 - Unclear why.
Exactly why I'm perplexed when people are able to install USB drivers for MacBook Air's to be able to drive 2 displays one over DP & the other over USB somehow after a driver software is installed
There's a product called DisplayLink that's used with certain USB monitors and USB docks. It compresses the video displayed on the screen in a userland daemon.

It requires Synaptic's software and it does work as a solution for devices that don't have enough I/O capacity to support multiple DisplayPort screens.

They use that at work on some of the freeze seating places.

It's quote buggy and resource hungry on macos and didn't play nice at all on my Ubuntu.

I guess you're meaning this from that page?

    To use an eGPU, a Mac with an Intel processor is required.
People have previously mentioned that for Apple Silicon based macOS, since the "GPU" is effectively part of the cpu they've gone with a shared memory model (for the OS).

If you try plugging in an external gpu to an Apple Silicon mac, it would need to understand "separate graphics memory". That's how macOS runs for x86 arch, but it's not the memory model used by Apple Silicon arch. Thus, failure.

That's my understanding of things from half remembered readings anyway. :)

(comment deleted)
Even if Apple did implement more eGPU support, their relationship with Nvidia has soured so badly that I don't expect their drivers to ever work with macOS. AMD cards may work because Apple has used their GPUs before, but even on their new ARM chips Apple has decided to forego support for AMD GPUs.
Doesn’t Intel own a bunch of necessary patents for Thunderbolt?
Where's CUDA on FreeBSD, Nvidia?!
If Nvidia ever wants to see their GPUs on next generation PlayStation consoles, an updated FreeBSD display driver is a good stating point.