The Initial Performance Of NVIDIA's R515 Open-Source Linux GPU Kernel Driver

[jeoshua]

Your surprised that a new, self-proclaimed "alpha" GPU driver that isn't a line for line copy of the original driver is typically 2%-3% slower than the original driver with years of optimizations and bug fixes?!?

Yes. It's not like they just coded this up from a black box implementation from first principles. All they really had to do was release their code as open source and it would have the same performance. This shows that they didn't do that, and the fact its so close in performance shows that it WAS from the same stock code, but they changed it in a way that has hurt performance.

[ryao]

That was much better than I was expecting with power management still to be implemented.

Since the driver is production ready for datacenter applications, no power management probably means that it runs at maximum clock speed all of the time, rather than the status quo with nouveau where no power management means that it runs at the minimum clock speed all of the time.

[ryao]

Yes. It's not like they just coded this up from a black box implementation from first principles. All they really had to do was release their code as open source and it would have the same performance. This shows that they didn't do that, and the fact its so close in performance shows that it WAS from the same stock code, but they changed it in a way that has hurt performance.

They have a new driver architecture that leverages the GSP. The proprietary driver supports that too. You need to configure the NVreg_EnableGpuFirmware kernel module parameter to turn it on. Had Michael tested that as a third option, we would know whether it was the switch to the GSP or the creation of the open source driver that caused the performance difference.

Also, it is possible that the lack of power management is responsible for the discrepancies. Since the GPU cannot downclock, it is putting out far more heat, so the temperature dependent peak clock speed likely suffers. They plan to fix this in the future.

Edit: Another possibility is that the issue is that the open source kernel module does not yet have resizeable BAR support:

https://github.com/NVIDIA/open-gpu-kernel-modules/pull/3

[sleax]

Good chance that some of the performance differences between the open and proprietary drivers are due to the removal of application specific "hacks".

It's not a coincidence proprietary drivers updates coincide with major game releases (not to say this is the only reason).

Benchmarks are everything in this market space. Little shortcut here, bit of extra culling there...

If so, unlikely the open drivers will ever quite match up to the closed source ones.

[STiAT]

Hmh, I'll switch anyway if Fedora decides to provide those drivers.

[gnarlin]

Ah makes sense, cheers. A little disappointing but at least there's hope for Nouveau.

One might say that there's a..........Nouveau hope?
I'll see myself out.

[Jabberwocky]

The performance doesn't look too bad for such an early release. I was expecting much worse results.

No. Just no.

https://blogs.gnome.org/uraeus/2022/...ant-for-linux/

The article talks about sharing a kernel module between proprietary Nvidia and mesa. It does not explain how the code is going to be mainlined with firmware that masks a huge part of the HAL. Nvidia like to lie and cut corners[1], this approach seems to fit their modus operandi.

The burden is on Redhat to make the code work but Nvidia is not going to give anyone full access to the hardware. Redhat now needs to use it's influence to try and get the code merged with maintainers not being happy on the one side and Nvidia strong arming Redhat to get the job done on the other side.

IMO C8292's statement holds.

1. Nvidia lied about how much money they made by selling gaming GPUs to miners. Nvidia got fined 1.5 hours with of revenue for lying to everyone which is just a small operating expense at this stage: https://www.guru3d.com/news-story/nv...o-revenue.html

[nordkamp]

One might say that there's a..........Nouveau hope?
I'll see myself out.

HA
Never apologise for a good joke my friend!

[mangeek]

...cards that now sit in the awkward spot between nouveau and this new driver.

Yep. I stupidly bought an MX250-based laptop several years ago and waited its entire lifecycle for viable accelerated 3D. It's a shame that this thing is basically going to live and die without seeing the NVIDIA GPU used for daily driving (the fractional scaling performance on X11 is so awful that it's not viable).

It would be nice if NVIDIA threw Maxwell and Pascal users a bone and provided a mechanism to load the proprietary firmware blobs from their packages in via module params in the open source driver, even if it was unsupported. I'd be happy with 'Nouveau' running on a GPU that had the ability to reclock.

[pete910]

Why not test with mesa ? Is that not the open driver?

Does the Mesa stack need support adding ?