• 0 Posts
  • 238 Comments
Joined 2 years ago
cake
Cake day: April 13th, 2024

help-circle

  • She answered that in her blog post that the Phoronix article links to:

    Which GPUs does this work with? Is it only AMD GPUs?

    Whether or not your GPU can benefit from it depends on the kernel driver - more specifically, whether it sets up the dmem cgroup controller.

    amdgpu and xe both have support for the dmem cgroup controller already. In theory, Intel GPUs running the xe kernel driver should benefit as well, although I’m not sure anyone tested this yet.

    For nouveau, I have sent a patch for dmem cgroup support to the mailing lists. This patch is also included in my development branch, so if you use my AUR package it should work. In other cases, you will need to wait for the patch to be picked up by your distribution, or apply it yourself.

    The proprietary NVIDIA kernel modules do not support dmem cgroups yet, so this won’t work there.







  • I’ve been thinking more about this: It seems unlikely they will ever provide a significant portion of intercontinental traffic, even if there is some latency benefit. The fundamental issue is one of bandwidth. You can stuff fiber optics full of data in ways most lay people wouldn’t believe. Using different frequencies you can put many data channels parallel. 88 x 200 Gigabit/s per fiber is no issue at all with components we bought 5-10 years ago already for our use case on land, and spectral efficiency is still getting better. A typical subsea cable will have around 8 fiber pairs, so 8 x 88 x 200 G in both directions for one cable is probably normal.

    The intersatellite links on Starlink satellites are reportedly also at 200G, and there are three of them on there, but intended to go to different neighbors I think. I’m not familiar with the specifics of free space optics, but I expect both wavelength division multiplexing and space division multiplexing will be much harder.

    So only applications where latency is really critical will probably be able to buy into that limited intercontinental bandwith. High frequency trading maybe. And then diplomacy and military, where the tapping resistance plays a bigger role than the latency.



  • Geostationary satelites at 38000±2000km distance take around 125ms up and the same back down for about 250ms for you to reach a server, maybe you were thinking of that. My uncle has that on his farm in Australia. It’s bad, you can’t have a good IP based phone call because the delay is too long, people keep starting to talk at the same time.

    Starlink flies in low earth orbit 450 - 500 km, so maybe up to 1500 km distance from you if its at the edge of reachability, for a worst case. That’s 5 ms, or 10 ms from you to the server. At those low distances the buffering overhead of their system will dominate like with Wifi. I don’t know how it works for Starlink specifically.

    I think the real killer issue for starlink and realtime tasks is the constantly changing latency, and the handoff between satellites.



  • Holy shit that’s low. So to get even just 8 weeks you need to have worked there for 5 years. And this from a rich company.

    I work for a small company of 150 people. I’ve got 3 months of notice period in my contract, though it’s going both ways. This protection started immediately after my trial period ended, which was also 3 months long. That’s considered normal here in Switzerland for office jobs. The minimum by law after your first year of employment would be 2 months.

    Just for reference, in the European context the Swiss labour laws are considered quite weak. Market liberalism is comparatively strong here.