DLSS 1.0 vs 2.0 in 3DMark

neeyik

Posts: 1,367   +1,486
Staff member
Just been playing about with the recently updated DLSS test in 3DMark, to compare performance differences in the DLSS 1.0 and 2.0 modes now offered. Test system uses a GeForce RTX 2080 Super.

DLSS 1.0 = average fps results (without DLSS, with DLSS, % increase)
1080p = 52.09 | 72.07 | +38.3%
1440p = 32.28 | 46.32 | +43.5%
4K = 12.90 | 28.40 | +120.2%

The DLSS test uses the Port Royale ray tracing benchmark, which uses TAA to tidy up the final image. When using DLSS 1.0, the engine renders the scene at lower resolutions, then uses DLSS to upscale it to the desired resolution. The very final frame is then rescaled again, to match the monitor's resolution. The DLSS internal resolutions are 1440 × 810 for 1080p, 1920 × 1080 for 1440p, and 2560 × 1440 for 4K.

The overall results are very impressive. I only have the Advanced Edition of 3DMark, so I can't produce frame-to-frame comparisons, but 1080p DLSS 1.0 looks very nice indeed.

DLSS 2.0 Performance mode = average fps results (without DLSS, with DLSS, % increase)
1080p = 51.88 | 110.38 | +112.1%
1440p = 32.48 | 76.30 | +136.4%
4K = 14.85 | 39.41 | +165.4%
8K = did not run | 1.67

This mode performs much better than DLSS 1.0 but categorically looks a lot worse - rather fuzzy, with unpleasant hatching effects in some of the shadows. There's an Ultra Performance mode available and I only ran it for 8K, which came out at 15.34 fps on average. Massive improvement but really didn't look great at all.

DLSS 2.0 Quality mode = average fps results (without DLSS, with DLSS, % increase)
1080p = 51.96 | 85.60 | +64.7%
1440p = 32.27 | 54.20 | +68.0%
4K = 12.99 | 27.70 | +113.2%
8K = did not run | 0.92

This mode looks a lot better than the Performance one, and 1440p is particularly pleasing to watch. Interestingly, the 4K non-DLSS tests were repeatedly glitchy - lots of the traced shadows kept jumping about in place.

Sticking with the 1080p results, the results in terms of performance go in this order: DLSS 2.0 Perf > DLSS 2.0 Qual > DLSS 1.0. However, in terms of visual quality, the order is completely reversed: 1080p DLSS 1.0 looks absolutely pristine. I'd say it's objectively better than the 1440p DLSS 2.0 Quality mode.

3dmark_dlss10.jpg

It's a shame that DLSS 1.0 isn't going to feature in new games anymore, but hopefully Nvidia will continue to tweak the generalised neural network in 2.0.
 
Last edited:

neeyik

Posts: 1,367   +1,486
Staff member
  • Thread Starter Thread Starter
  • #3
Thank you @Kshipper - I was a little surprised by the 4K gains but I can only assume that the ray tracing used in the benchmark is very shader bound at normal 4K, so the internal reduction to 1440p must help enormously.