3 ms·
On the other hand, I wish eGPUs were more popular/standard. Having a separate enclosure is nice when the GPU is the main outlier to a modern system. The rest o
by solardev 3y ago
On the other hand, I wish eGPUs were more popular/standard. Having a separate enclosure is nice when the GPU is the main outlier to a modern system.
The rest of the PC can run off, what, 200 watts, and fit on a nano motherboard with basic cooling? Then comes the big ass space heater GPU that draws most of the power and disrupts airflow for everything else. Why not have it in its own little cage, almost like a game console? It can get better cooling that way, its own power supply that maximizes output instead of efficiency, and easy upgrades. Connect it when you wanna game or do AI stuff, leave it completely off the rest of the time.
If it had its own little web enabled controller, maybe you could even rent it out for processing workloads in the clouds.
- GordonS 3y agoWhat's the current state of eGPUs these days? I had a quick look, and wow, there aren't many eGPU enclosures going about, and what there are, are expensive. I always saw them as intended for laptops rather than desktops - are there any downsides or performance penalties compared to slotting directly into a PCIe slot?
- mschuster91 3y ago> are there any downsides or performance penalties compared to slotting directly into a PCIe slot? Latency, for one. All signals must at least pass through two Thunderbolt/USB4 chipsets on their way to and from the GPU, and at least for anything but high profile AMD CPUs Thunderbolt is connected to the southbridge instead of northbridge (where the x16 GPU slot is).
- solardev 3y agoAre there none that still use pcie, just through a cable instead of a slot? That's too bad, if so :(
- mschuster91 3y agoPCIe can only go to like 14 inches in trace length, and that's already challenging enough to do for risers and extender cables without compromising signal integrity. [1] https://www.protoexpress.com/blog/pci-express/ https://www.protoexpress.com/blog/pci-express/
- solardev 3y agoI see. Thanks for the explanation! (Not a hardware guy. Didn't realize there physical/electrical limits to these.)
- oakwhiz 3y agoThere is a way to solve this with optical but the cost is still high and it's not terribly compact. Usually you can only get 4 lanes directly into one optic. Most implementations I have seen rely on a 2-port PCIe packet switch chip, and they rely on SRIS/SRNS clocking and hotplug/"surprise removal" support is provided by that switch (don't want a crash if you yank the cable)
- j16sdiz 3y agootoh, many risers are out of spec and still work
- mschuster91 3y agoThere's a difference between "appears to work" and "works at advertised speeds". PCIe is surprisingly tolerant but it will downgrade the link speed during initial training.
- oakwhiz 3y agoI haven't looked at the eGPU space in a while, but the convenience of being able to move a very expensive GPU between machines without taking apart the machines or having excess power is definitely an interesting idea especially now that ML models are so popular.