NVLink Performance in DaVinci Resolve 17.0

Introduction

NVLink is a very interesting technology from NVIDIA that allows GPUs to be able to directly communicate with each other at speed up to 112 gigabytes per second. This direct connection allows for data to pass between cards much quicker than it can over the PCIe bus, which can enable greater multi-GPU performance and even features such as VRAM pooling which could, in theory, effectively doubles the amount of VRAM available to the application.

However, this technology is not a magic bullet for performance. Not only does the application you are using have to support NVLink itself, but it also has to have a workflow that benefits from the faster interconnect between the GPUs. In some cases, it can dramatically improve performance, while in others, the performance benefits are minimal.

In the recently launched 17.0 version of DaVinci Resolve Studio, Blackmagic has listed in the update notes:

Support for NVLink high speed interconnect with supported NVIDIA GPUs.

Beyond this one notation, however, there is no additional information that we could find about this feature – even the comprehensive New Features Guide has absolutely no mention about NVLink. So, to see if there is any performance benefit to using NVLink in its current state, we decided to test NVLink using our PugetBench for DaVinci Resolve benchmark.

If you would like to skip over our test setup and benchmark sections, feel free to jump right to the Conclusion.

Test Setup

Listed below are the specifications of the systems we will be using for our testing:

*All the latest drivers, OS updates, BIOS, and firmware applied as of March 2nd, 2021

In past versions of DaVinci Resolve Studio, there were major issues with multi-GPU systems whenever SLI/NVLink was enabled. Because of this, we did not want to simply compare a dual GPU setup with NVLink/SLI on or off, but also experiment a bit with a few other configurations in case we continued to see issues with SLI enabled. To do this, we tested the following combinations with the NVIDIA RTX A6000:

• Single GPU
• Dual GPU – no NVLink bridge, SLI disabled
• Dual GPU – NVLink bridge installed, SLI disabled
• Dual GPU – NVLink bridge installed, SLI enabled

These combinations will let us get a good look at whether we still need to enable SLI, and how the performance compares to a single GPU and a dual GPU configuration where NVLink is not present.

For the testing itself, we are using our PugetBench for DaVinci Resolve benchmark, which has recently been updated to include support for DaVinci Resolve 17.0.

Raw Benchmark Results

While our benchmark presents various scores based on the performance of each test, we also like to provide the individual results for you to examine. If there is a specific task that is a hindrance to your workflow, examining the raw results for that task is going to be much more applicable than the scores that our benchmark calculated.

Feel free to skip to the next sections for our analysis of these results to get a wider view of how each configuration performs.

DaVinci Resolve NVLink Performance with NVIDIA RTX A6000 48GB

Starting with the Extended Overall Score, the performance with NVLink is not looking particularly good. It is still faster than a single RTX A6000 GPU but falls behind the two non-NVLink dual GPU configurations by about 6%.

However, there are some very interesting nuances we noticed when looking at the results in closer detail:

Looking at the sub-scores from our benchmark, we get a look at how well NVLink works (or doesn't) in different situations.

Performance with GPU effects like noise reduction and OpenFX are where we expected NVLink to give the biggest performance gains, but in reality, the performance with two RTX A6000 GPUs is virtually identical whether NVLink was enabled or not. So, while NVLink doesn't seem to help much with these effects, it also doesn't hurt performance either.

Exporting relatively simple 4K and 8K timelines with various codecs is where NVLink starts to show some performance issues. In our 4K tests, two RTX A6000s with NVLink are only able to match the performance of a single RTX A6000, and fall behind the other dual GPU configurations by about 5%. Performance with 8K timelines, however, is roughly equal between all the dual GPU configurations, so there is neither a gain nor penalty to using NVLink in that situation.

Fusion is where NVLink has the most significant issues. In these tests, the RTX A6000s with NVLink ended up performing about 15% slower than a single RTX A6000, or 20% slower than dual RTX A6000 cards that did not use NVLink. There was a single test where NVLink improved performance (Phone Composite) by a decent amount, but that was more than offset by multiple other tests where it was significantly worse.

How well does NVLink perform in DaVinci Resolve Studio 17?

While NVLink is now officially supported by DaVinci Resolve, it does not work particularly well at the moment. In almost every test we ran, having NVLink enabled was at best equal to having it disabled, but in some situations, it actually carried a significant performance penalty that lowered performance to the point that even a single GPU was measurably faster.

To be fair, we don't know what Blackmagic's intentions are for NVLink, and it may be that the official support they added in version 17.0 is only the first step of many. Unfortunately, we won't know until additional information is released, but at the moment, you are better off leaving NVLink disabled if your system supports it.

Hopefully, we will get more information about Blackmagic's intentions for NVLink in DaVinci Resolve. It seems unlikely that they would invest the time and effort into adding a feature that works so poorly, so we remain optimistic that NVLink will eventually be a benefit for Resolve users. If we learn anything, we will update this article with more information, or publish a follow-up if we get an update to Resolve that improves NVLink performance.