Hacker News Reader

Comment by bilegeek | Hacker News Reader

It's still a problem if you want to run previous gen hardware. For instance, if I have a Gen4 motherboard and pass through a second GPU to a VM - so the second GPU at Gen3 x8 plus the host GPU at Gen4 x8 - I'm only using a quarter instead of half the throughput the hardware actually supports, because Gen3 is half the speed of Gen4 per lane.

Or even think about Gen5, doubling Gen4 speeds: With a PCIe switch, I could get essentially 3/4 performance out of a Gen5 host GPU, and full performance out of a Gen3 second, instead of 1/2 and 1/4 respectively.

View on HN

Do GPUs even saturate PCIe3 x16 or x8 connections, let alone PCIe4 or 5, that we need to be worried about bottlenecks?

The only thing I'm aware of that will actually saturate a PCIe connection is NVME storage.

For gaming, not even close. I've often had a GPU monitor running, charted bus utilization...and it's barely a few percent no matter what game I'm running, except for a very brief spike during texture and model load-in at the start of a round/level.

Specifically for VM's, something like Looking Glass (shared framebuffer so you don't need a dedicated monitor) needs the extra bandwidth to get the video signal through without tanking performance, especially at higher resolutions and refresh rates.

Aside from that, I agree it's not too big an issue for 95% of computer users, but it's something to keep in mind.

[1]https://looking-glass.io/

To quantify this: a 4k 60Hz uncompressed video stream needs approximately one lane of PCIe gen4 (though only in one direction, while PCIe has the bandwidth in both directions). So any realistically obtainable resolution and refresh rate will not be a problem for sharing the framebuffer on a GPU with PCIe gen4 x16, but lower-end GPUs often have just x8 and there are even some mobile GPUs with x4 being packaged onto desktop graphics cards now.

If I remember correctly, running a 3080 10GB in PCIe 4.0 yields >5% performance average in gaming benchmarks compared to running it in PCIe 3.0.

You could probably argue that you should just settle and not take the 5%, but when you are buying in at such a price point, you are likely concerned with 5% differences. There are also higher end SKUs such as the 3090Ti where that performance gap is likely larger.

I've had significant frame rate drops with PCIe3 x8 compared to PCIe3 x16 (Easy 20 - 40% drop, depending on the game)

This is precisely why I am not going to buy into this year's Ryzen motherboards, how am I supposed to attach all of my PCIe 2.0 LSI HBA as well as my 2 GPUs, and NVMe? My options are buying 2 machines (one for storage, one for compute), sacrificing speed somewhere, or buying into either Threadripper Pro OEM or EPYC ($$$).

The bandwidth is there by an order of magnitude, but not the lanes to actually use it.

Reply on news.ycombinator.com