This content is a functional academic exercise, which means that we will sometimes stand out from real world examples and instead examine how the card behaves and the scale at the basic level. This means that you need to see scenarios that can not be played, such as 8K resolution, to see if AMD improves scaling over NVIDIA; theoretically, given what AMD told us about performance, it would be necessary to confirm this. Anti-aliasing has several samples per pixel, often 2x, 4x, 8x or 16x, and has a very similar effect on increasing resolution. Hit the same part of the pipeline.
If AMD's memory bandwidth really benefits in games, it is assumed that this would happen by reducing or turning the delta performance between Radeon VII and RTX 2080 as an enhancement of resolution and anti-aliasing, with the deterioration or rising grid that gives priority to RTX reduces resolution. We have already shown how Radeon VII can sometimes fill the gap in higher resolutions when we published our review, and talk about how the gap, as in F1 2018, would expand to 1080p. Now is the time to see if it turns or closes at high resolutions; However, do not forget that the obvious poor side is to enter the territory where the content can not be reproduced in the image, so we call it an academic exercise.
Superposition AA Testing
Starting with shaders set to "low", the textures are set to "high", changes in resolution at 4K, 5K and 8K showed that NVIDIA reached a 27.4% preference from the baseline and moved to 27, 9% guiding error – 5K, then 28.1% lead to 8K. It does not really change here from 4K to 8K, at least with regard to scaling, although these resolutions obviously decrease as they grow.
The next test was performed with "Extreme" shaders and found that AMD actually acquired this scenario. AMD averaged at 1566 points, and NVIDIA at 1504 points. We tested this test 4 times and averaged it, and found that this delta is beyond the standard deviation.
When we learned that there were major differences in Extreme shaders, we tested tests from 1080p to 4K. At 1080p, the RTX 2080 NVIDIA performed approximately 25.8% better than the Radeon VII card at a 100% baseline. At 1440p, the advantage of the RTX 2080 rapidly dropped to 16.1%. The results of 4K are too close and they create a nasty card, but the RTX 2080 ends a really better result than the Radeon VII, with Radeon VII having some advantages. Now, in terms of the "real" frame rate, the meter worked at approximately 11-12FPS AVG. This can not be played, clearly and this advantage is not really realized because no one would really use these settings, but it's interesting to see how two cards can turn positions in different test scenarios.
Testing GTA V AA
GTA V has many options to prevent overlap, so we will run both MSAAs from 2x to 8x and then enable MSAA with eight reflections. This will start with the AVG FPS table only in different settings, which best illustrates how the gap changes with each test.
With 4K and very high and ultra settings for everyone, 2x MSAA places us at 58.8FPS AVG for 2080 and 51.4FPS AVG for Radeon VII, but the more important indicator is that the 2080 has a 14.4% advantage over the Radeon VII card. . With 4x MSAA, management does not change much and reduces to 14.1%, which is actually within the framework of the error. At 8x MSAA we begin to see some movement. AVG FPS is 27.5 on Radeon VII and 31FPS on RTX 2080, with an RTX 2080 improvement ranging from 13%, compared to the original 14.4%. Until then, adding 8x MSAA reflections anti-aliasing, we see the difference changes to a 10.8% advantage for 2080 FE. Looking at a common graph, we can see that AMD is beginning to recover some of the losses at the far end of the scale, except that we are on non-playable territory for other reasons.
For the reference, with the amount of test passes that we performed, the standard deviation from the beginning to the beginning was about 0.3FPS AVG. These results are very accurate.
Fire resistance test AA
Firestrike remains one of the best tools for this type of synthetic load. For this purpose we will separately review GT1 and GT2. The GT1 heavily fills the GPU with polygons and tessellation, implements shadow and lighting crunching, and uses computational shaders for post-processing and particle physics. GT2 is heavily loaded on computers and greatly increases the number of processed pixels per frame, but reduces the load of tessellation by more than 50%. GT2 should therefore work better on Radeon VII than GT1, relatively.
Starting with GT2, we see that the Firestrike Ultra supplies AMD's NVIDIA settings at AMD 86.7% of the base 100% capability of Radeon VII. So far, this is unique for this criterion. At 2x MSAA, this difference was reduced to 90.7% of the baseline capacity. 4x MSAA brings us to 94.6% of the baseline, 8x MSAA finally allows the RTX 2080 to exceed AMD and has 104.7% of the baseline capacity.
It makes sense that AMD is generally the most competitive in GT2, where computer load and memory are loaded. As to why NVIDIA will begin to move closer to MSAA by 8x, we will work with the 3DMark team and others to try to better understand the specific behavior. Our estimate is that this may be related to NVIDIA memory compression, or perhaps by detecting edges in the anti-aliasing with this specific software execution or with some other advantage at the implementation level.
For GT1, where tessellations and polygons are more difficult, we can see that both cards start operating the same in accordance with the Firestrike Ultra settings, with NVIDIA gaining 16% for 2xMSAA, 21% for 4x MSAA and 27% for 8x MSAA.
At 8K with 8x MSAA, we achieve genuinely useless levels of framerate, but finally we see that Radeon VII pulls far ahead of RTX 2080. This was what was needed to fully exhaust the memory of the RTX 2080 at which point it began to choke as downloaded memory.
Far Cry 5 HD texture + AA testing
Far Cry 5 is next. We tested this with SMAA, TAA and HD Textures for various scenarios.
In the first test on 4K and using HD Textures with SMAA we measured NVIDIA to 62.8FPS AVG and AMD at 60.6FPS AVG, which means that NVIDIA is about 3.5% on. Similarly, they were also allocated 1% of the lowest value, and also 0.1% of the lowest value.
In order to test the scaling with different settings, we also performed a test with reduced VRAM consumption, otherwise the same settings. We can now switch to this card. Results that disable HD textures were functionally identical to HD texture results. It is not surprising that we are within the limits of the error, since the textures do not actually affect performance if VRAM does not become a limitation. On both devices, the quantitative experience is unchanged from the previous one. Qualitatively, the output is the same on each device. Also, NVIDIA is faster at about 3.3% faster.
Instead, instead of TAA, the performance is slightly reduced for each device, but NVIDIA still maintains a 3.3% advantage in AVG FPS. HD Textures and anti-aliasing changes in Far Cry 5 are not enough to reduce or modify the delta performance from card to card.
Conclusion: Not exactly what we expected
Our hypothesis was not quite correct. Looks like we came in a few cases, but Firestrike threw us off the road. The 3DMark Group and NVIDIA engineers have been invited to try to better understand this performance scenario. One of the cases where Radeon VII saw a tremendous rise was when we used VRAM to RTX 2080, which was done with Firestrike Ultra at 8K resolution and 8x MSAA. This, of course, is completely impracticable, and therefore we re-emphasize the text that this is really just an exercise in the research, but not one that is particularly practical. At 8K / 8xMSAA, we found that Radeon VII operated approximately 4.7-4.8FPS AVG by 2080. ~ 1-1.7FPS AVG. This is a "big" percentage improvement and is because we have exhausted the framebuffer to 2080, but obviously it can not be played on any device. Regardless, this is an interesting piece of information and could help you understand better the other card characteristics in future tests.
Editorial, testing: Steve Burke
Video: Andrew Coleman