News for hackers. Stuff that matters.
Re:I rage-converted my RTX4090 into an eGPU by bee_rider on Friday June 07, 2024 @09:14AM (#40608423)
I wonder what applications he has in mind. Maybe GPGPU/ML training kind of stuff? Rampant speculation:

Cable extenders are known for messing with signal integrity. Flipping bits over long wires takes a while. ML training seems (I could definitely be wrong here) less latency sensitive than classic GPU applications, like gaming. But it does want the bandwidth. Could there be room for an active cable with repeaters? Does anyone make a cable with, like, little flip-flops in it? Can an extender act as a little buffer or pipeline or something like that?

Cunningham’s law: activate!

Re:I rage-converted my RTX4090 into an eGPU by tanelpoder on Friday June 07, 2024 @09:32AM (#40608538)
Just random LLM use (no training) and some cuDF experiments so far. So far I haven't noticed any issues (and I tend to investigate any unexplained glitches).
Re:I rage-converted my RTX4090 into an eGPU by bee_rider on Friday June 07, 2024 @11:11AM (#40609462)
I should have noticed the url/username match, haha.

Neat project! I’ve done a couple (basic, consumer, no modifications of the case) mini itx builds and always wished I could get the GPU out of the case. Stick the hot thing out in the world, where it has infinite airflow, seems like the most obvious thing to do, but it really isn’t catered to well enough…

Re:I rage-converted my RTX4090 into an eGPU by lreeves on Friday June 07, 2024 @09:24AM (#40608485)
Like the workstation cards the 4090s support an ECC mode for their memory if that's important.
Re:I rage-converted my RTX4090 into an eGPU by jauntywundrkind on Friday June 07, 2024 @10:44AM (#40609230)
The PCI AER (advanced error reporting) can detect issues quite well & report them. The author mentiona that they sort of expected problems, but no, AER hasn't reported any.

Links have to be very smart & capable of self training in modern PCIe to perform at the speeds they do: they have link equalization. So there's incredible diagnostics builtin.

As for active cables, Marvell just announced Alpine P, which is a PCIe6 retimer that capable of working in active electrical cables (and also active optical cables!). I don't know if they buffer data or not - if so it would be a small amount of data I'd expect - but they do receive & retransmit. https://www.servethehome.com/marvell-extending-pcie-gen6-rea...