Nvidia Resizable BAR Examined, Benchmarked

Today we are testing the newly supported Resizable BAR function on Nvidia GeForce GPUs by benchmarking an Asus ROG Strix RTX 3080 OC graphics card. With the release of the AMD Radeon RX 6000 series, the customizable bar became part of the GPU performance talk. AMD was promoting a new feature called Smart Access Memory, or SAM, but in reality AMD had just renamed a PCI Express feature called Resizable BAR that hadn't been used in a while.

The way AMD advertised and implemented Resizable BAR on Radeon 6000 GPUs had to be paired with a Ryzen 5000 series CPU and a newer 500 series motherboard. AMD made it sound like this was a new optimization they had developed that would be exclusive to their products, but it isn't. The claim was that in a number of games with SAM enabled, your AMD CPU / GPU combo was able to achieve a 10% increase in performance or more, which caused a stir but was short lived.

Of course, it's free service and not something we come across a lot when testing new features. Enabling SAM will increase the Radeon RX 6800's performance by 19% in Assassin's Creed Valhalla, 18% in Hitman 2, 14% in Borderlands 3, and 13% in Godfall. These are some impressive wins, and you are not going to sacrifice anything to achieve them.

However, a deeper dive into SAM found a number of cases where the performance was degraded or, in most other cases, made no difference at all. In the end, when we tested a wide variety of games, we found that SAM averaged a small 3% increase at 1440p.

Net income is still welcome, and it's possible that BAR size could play a more consistent role in the future as developers create games with this ability. As a result, Nvidia has stepped on the customizable BAR train and is now supported by GeForce RTX 30 GPUs.

There are three things you'll need to enable the customizable bar on your RTX 30 series GPU:

1] A supporting VBIOS is required. Anyone who has bought an RTX 3060 already has the required VBIOS, but RTX 3060 Ti, 3070, 3080, and 3090 owners will need to update.

For this article, we updated the VBIOS on the ROG Strix RTX 3080 OC and the process was simple. We navigated to the Asus support website, went to Drivers & Tools> BIOS & Firmware, and downloaded the latest BIOS. We ran the tool and it took only a few moments to finish. After a restart, the VBIOS update was carried out.

2] You need a compatible motherboard that supports the BAR resizable, including the BIOS version. In our case, the MSI X570 Unify has already been updated with a Ryzen 9 5950X because we previously used it to test the Radeon RX 6800.

3] The latest GeForce Game Ready driver. The March 30th release added support for resizable BAR. However, at this point, only use the latest driver and you are done.

Once these steps are completed, you can verify that the resizable BAR is working by opening the Nvidia control panel and going to the system information section. In the detail area you will find "BAR with changeable size". If it's working properly, the status will simply be "Yes".

What we found interesting about Nvidia's implementation is a claim they made in their announcement …

“In practice, the performance benefits of Resizable BAR can vary significantly from game to game. In our tests, we found that some titles benefit from a few percent up to 12 percent. However, there are also titles where performance degrades. As a result, Nvidia will pre-test titles and use game profiles to enable Resizable BAR only if it has a positive impact on performance. That way, you don't have to worry about bugs or slowdowns in performance, and you don't have to rely on the community to rate each title and find out if Resizable BAR is beneficial for the games you play. "

That sounds pretty good. Basically, Nvidia says that you can get all the benefits of a customizable bar with no flaws. With the Radeon RX 6800, we were able to determine a performance regression of up to 7% for certain titles. If so, it is a win for Nvidia and the customizable BAR implementation is much better than AMD's.

At the time of writing, Nvidia's whitelist of supported games is pretty limited, consisting of only 17 games. Of those 17 games, we tested 12, along with 8 others that are not currently officially supported. So this could be very interesting. So it's time to get into the benchmark results. We test with the Ryzen 9 5950X on the MSI X570 Unify motherboard with 32 GB DDR4-3200 CL14 dual-rank and dual-channel memory.

Benchmarks

Starting with Assassin's Creed Valhalla, we find that performance increases by 8% at 1080p and 9% at 1440p. Not nearly as big as the gains we saw from the Radeon GPUs in this title, but still ~ 8-9% is nothing to make fun of. Unfortunately, when we need the most performance boost at 4K, we are basically seeing the same level of performance. Overall still a good result in this title.

The frame rates in Forza Horizon 4 are already excessively high with the RTX 3080. Assuming these margins grossly affect the product stack, this is a good result.

At 1080p we see an 11% increase in performance, 9% at 1440p and 6% at 4K which was nice to see.

We see a slight increase in performance in Horizon Zero Dawn with the customizable BAR enabled, only 3% at 1080p, 5% at 1440p, and 4% at 4K. Not exactly something to get upset about, but it's again a little free performance boost. Who will complain about this?

The gains in Borderlands 3 are also small: a 5% increase in 1080p, 4% in 1440p, and only 3% in 4K.

We're seeing a relatively steep 7% increase in Cyberpunk 2077 at 1080p, but nothing at 1440p and 4K.

There have been a number of titles like Star Wars Jedi: Fallen Order where the performance remained virtually unchanged. If we are to believe Nvidia's claims, we could look at a situation where the resizable BAR has been disabled to avoid performance regression as this is not a whitelisted game, but more on that soon.

Wolfenstein Youngblood isn't whitelisted either, and we don't see any change in performance here at virtually identical frame rates with and without a customizable bar.

Shadow of the Tomb Raider is another game that was not whitelisted on Nvidia. Again, we don't see any change in performance with BAR enabled in resizing, although we're not sure if Nvidia still had the opportunity to test every single game.

Far Cry New Dawn is not whitelisted either, but here we see a performance regression with BAR size enabled at 1080p and 1440p. So what's going on here?

Shouldn't the Sizable BAR be deactivated in this case? The performance regression isn't great, but the 4% drop in frame rate at 1080p and 1440p was highly repeatable so it clearly doesn't disable the feature here. But let's get to an even more problematic set of results …

Watch Dogs Legion is a whitelisted game that is said to benefit from a customizable BAR and therefore has a performance improvement with BAR enabled. However, the opposite is the case if there is a significant 10% derating at 1080p and a 3% derating from 1440p, or a massive 16% derating from a low result of 1%.

We were able to consistently reproduce these results every time we saw a sharp drop in 1% low performance with BAR activated in Watch Dogs Legion.

We are not alone with these results either. Other tech media have also experienced the same poor performance with the scalable BAR enabled on Nvidia GPUs, and we've seen similar reports from users on forums and reddit.

This is pretty strange as Nvidia claims a 9% performance increase in Watch Dogs Legion with an RTX 3080 at 1440p with the Ryzen 9 5950X on an X570 motherboard with 32GB of RAM. It is possible that changes have been made since then that have adversely affected performance. It's hard to tell as the results were published about two weeks ago.

The strange results didn't stop there. Death Stranding also saw a degradation in 1080p and 1440p with the modifiable BAR enabled even though it was a whitelisted title. We saw an increase of up to 9% in 4K, which was great, but the implementation here is barely flawless, with a reduction of up to 9% in 1080p and 8% in 1440p.

Performance summary

Here's a look at the average performance of the 20 games tested. We'll jump to a per-game breakdown in a moment, but I wanted to show this graph first as this is typical of how GPUs are rated for performance and value.

We don't see any improvement for the RTX 3080 at 1080p, a 1% gain at 1440p when we round up, and a 2% gain at 4K. This is less than the average 3% gain for the Radeon GPUs. So when you factor in gaming performance as a whole, the customizable BAR's support changes practically nothing to change the picture.

If you look at the margins seen in each of the 20 games tested at 1080p, and as you can see, there are drops in performance. Nvidia claims that they use game profiles to enable Resizable BAR only when it has a positive impact on performance, but based on our findings, it doesn't. Division 2, for example, is not a game on the whitelist. Therefore, the game profile should automatically disable the customizable bar, which will not change any performance. However, we noticed the same problem with multiple titles.

Even at 1440p, we still see a number of cases where performance goes beyond the margin of error and these results were highly repeatable.

Without the drop in performance, the technology would only increase performance by 2%, but even without the drawbacks, there would be no reason not to enable it by default.

The customizable bar is a little more effective at 4K. The 1% to 2% drops seen in Far Cry New Dawn and Dirt 5 are insignificant, while some others benefit from them, if only marginally.

You might find this last graphic interesting. In the games that we can directly compare to our Radeon RX 6800 reviews, which are 18 of the 20 games we just tested, the GeForce RTX 3080 and RX 6800 compare each other in this way.

While Nvidia saw a ~ 1% improvement on average, AMD saw a 7% improvement with the modifiable BAR enabled in titles like Assassin's Creed Valhalla, Hitman 2, Borderlands 3, Godfall, and a few others. Based on these results, AMD appears to have better implemented support for customizable BAR.

What we learned

It's great to see Nvidia adding support for scalable BAR on the latest generation GeForce 30-series GPUs. It's just a bit of a shame it didn't work out as well as we hoped it would. Likewise, this would be an amazing feature for AMD to implement on more / all products if possible and this would mean we are one step closer to testing with resizable BAR enabled by default.

Note that in order to enable / disable the resizable BAR with a Radeon or GeForce graphics card, you must restart the system, enter the BIOS, and then turn it on or off. So that's not exactly a practical solution, and we would argue that it means that the performance upgrades are no longer free, but at the expense of your time and energy.

It was also interesting to see that the RTX 3080 saw a performance improvement of 9% at 1440p at best on AC Valhalla and Forza Horizon 4, while the RX 6800 saw up to a 19% increase in performance on Assassin's Creed Valhalla Forza Horizon 4 was untested as it hadn't left the terrible Windows Store for Steam yet.

As of today, a customizable bar seems more beneficial for Radeon users who happen to have the full Ryzen 5000 combo. Note, however, that this will not always result in a positive performance. For GeForce owners, it is only worth enabling the size of the bar for 2 or 3 games. So it's up to you whether you care or not, but at least the option is now and it could improve over time.

Purchasing links:
  • Nvidia GeForce RTX 3070 on Amazon
  • Nvidia GeForce RTX 3080 on Amazon
  • AMD Radeon RX 6800 on Amazon
  • AMD Radeon RX 6800 XT on Amazon
  • AMD Ryzen 9 5950X on Amazon
  • AMD Ryzen 9 5900X on Amazon
  • AMD Ryzen 5 5600X on Amazon

Leave a Reply

Your email address will not be published. Required fields are marked *