What could be thicker than the fastest video card of the moment? Exactly: twice the fastest video card of the moment. Armed with two GeForce RTX 3090 Founders Editions and Nvidia’s new NVLink bridge, in short, almost EUR 3,200 in GPU hardware, we are investigating what performance this monster combination is capable of.
SLI in short
The term SLI nowadays stands for a scalable link interface and was first used by GPU manufacturer 3dfx in the late 1990s. Nvidia acquired the brand name through acquisition and breathed new life into it, for a technique to combine the computing power of multiple video cards. In recent years, this has usually been done on the basis of alternate frame rendering, in which the video cards render an image alternately. An inherent disadvantage of this is that the time between images (the frame time) can vary because not every frame is ready at the same time. You perceive that as ‘stuttering’.
With the arrival of DirectX 12, everything would be different – and above all better. This API contains a technique called explicit multi-adapter, which allows a game to control multiple video cards itself. It is even possible to mix completely different video cards from AMD and Nvidia. Most of the DX12 games, however, do not apply this technique, while Nvidia officially pulled the plug this summer from the ‘old’ way with SLI profiles in the driver.
The SLI bridge that you used to get with every nice motherboard has not been used for a long time. After a high bandwidth version of that bridge, Nvidia switched the RTX 2000 series to NVLink, a much faster interconnect that originated in the data center world. The same technology is used for the RTX 3000 series, but due to the changed connectors, you still need a new bridge.
The controlled death of SLI
In the past, you could put just about any video card in SLI. AMD’s counterpart CrossFire was even more flexible on that point: you could even mix different GPUs. If we look at the last four GeForce generations, it is striking that support for SLI has been phased out further and further. In the past, for example, you could still easily combine two GTX 960s with each other, with the 10-series you needed at least one GTX 1070 for this and with the previous generation it was only possible from the RTX 2080. In the new series, only the RTX 3090 offers support. for SLI.
|
RTX 3000 |
RTX 2000 |
GTX 1000 |
GTX 900 |
** 90 / Titan |
Yes |
Yes |
Yes |
Yes |
** 80 |
No |
Yes |
Yes |
Yes |
** 70 |
No |
No |
Yes |
Yes |
** 60 |
No |
No |
No |
Yes |
SLI with the RTX 3090
If you continue the line, the next generation will end the story for SLI. Life was not made easy for you as an SLI enthusiast: over the past three generations, functionality has become increasingly limited while the required hardware has increased in price. The NVLink bridge required for the RTX 3000 series uses different connectors from the previous generation for the third consecutive generation. Moreover, it costs no less than 85 euros.
For this test, we built a different test system than in our regular GPU reviews. Where PCI Express 4.0 adds little according to our tests , with the mainstream platforms from AMD and Intel, the GPU bandwidth is again cut in half. This is due to the limited number of PCIe lanes. To prevent this and to minimize the expected CPU bottleneck with so much GPU power, we performed these tests on a HEDT platform, about which more on the next page.
Test justification
In order to minimize the bottleneck of the rest of the test system for the SLI configuration of RTX 3090s, we put together an overclocked system based on Intel’s X299 platform. In this way, both video cards have sixteen lanes of PCI Express 3.0 bandwidth, while a possible CPU bottleneck will occur less quickly than with our standard 3900XT test system. Intel’s midrange platform was not an option. Due to the lack of sufficient lanes and PCIe 4.0, the bandwidth per video card would then be reduced by half.
As the processor, we installed an Intel Core i9 10980XE with eighteen cores, overclocked to 4.5GHz, on an ASUS ROG Rampage VI Apex motherboard. G.Skill delivered four 32GB modules from its Trident Z Neo range for a total of 128GB at a speed of 3200mt / s. We run the OS on a Samsung 970 Evo 250GB ssd, the games are on a Kioxia Exceria + 2TB. Furthermore, the system consists of a Corsair AX1600i power supply and an extensive water cooling set from Alphacool. The latest version of Windows 10 Pro 64bit is of course installed as the operating system.
Test system |
Processor |
Intel Core i9 10980XE @ 4.5GHz |
Motherboard |
ASUS ROG Rampage VI Apex |
Random-access memory |
G.Skill Trident Z Neo 128GB DDR4-3200 CL16-16-16-36 |
SSD |
Samsung 970 Evo 250GB (OS) + Kioxia Exceria + 2TB (games) |
Nutrition |
Corsair AX1600i |
Cooling |
Alphacool Eisbaer Extreme Liquid Core 280 + Eisblock XPX Aurora |
Operating system |
Windows 10 Pro 64bit May 2020 Update (2004) |
Drivers and measurement method
For the game benchmarks in this review, we used the GeForce 456.55 driver. We performed the ‘creative’ benchmarks with version 456.38 of Nvidia’s Studio driver.
Using PresentMon, we measure the performance in each game tested, from which we calculate both the average frame rates (fps) and the frame times of the 99th and 99.9th percentiles and report the latter two in milliseconds. Please note that this is a software measurement. In an SLI setup, not all calculated frames are necessarily displayed on the monitor.
In the graphs on the following pages you will always initially find the average frame rates or the average number of images per second that a video card can calculate. The frame times do not give a picture of the average frame rate, but of the outliers in the negative sense that can ensure that a game does not feel smooth despite a good average.
The time it takes to render images within a 3d game and thus within our benchmark varies from frame to frame. Our frametime measurement stores the render times of all individual frames. Then we discard the 1 percent slowest frames. The highest render time of the remaining 99 percent of the frames, or the slowest frame, is the 99th percentile frame time.
At the request of some readers, we also added the 99.9th percentile values. So we only leave out the 0.1 percent slowest frames for this. In theory this is even more precise, but in practice incidental causes and measurement errors sometimes throw a spanner in the works. For now, we’ve listed them in the review, so keep that in mind when reviewing these results.
Games and creative workloads
Since not all modern games support SLI or multi-GPU in general, we have made a different selection for this test than in our regular GPU reviews. After all, testing SLI in a game that we know does not support SLI is not the most useful activity.
In addition, we took the opportunity to run some ‘creative’ benchmarks, in the field of video editing and rendering. Several users therefore asked in response to our RTX 3090 review. To give an idea of the performance in video editing software, we render a project with GPU acceleration enabled in both Adobe Premiere Pro and DaVinci Resolve. We also calculate a Blender scene in two modes: standard GPU acceleration for which the Cuda cores are switched on and Optix, where the ray-tracing cores are also used. Finally, we render three scenes in OctaneBench, OctaneRender’s benchmark tool.
Power consumption
Given the limited time in which we had two RTX 3090 FEs, we did not perform extensive current measurements. Total system consumption during the heaviest workloads was 880 watts, but the peaks were probably much higher. Initially we had set up a system with a 1200W power supply, but it failed during the benchmarking of Shadow of the Tomb Raider.
Benchmarks: 3DMark and games