SLI & mGPU
BTR has always been interested in SLI, and the last review we posted was in January, 2018 – GTX 1070 Ti SLI with 50 games. We concluded:
“SLI scaling is good performance-wise in mostly older games and where the devs specifically support SLI in newer DX11 and in DX12 games. When GTX 1070 Ti SLI scales well, it easily surpasses a single GTX 1080 Ti or TITAN Xp in performance. . . . SLI scaling in the newest games – and especially with DX12 – is going to depend on the developers’ support for each game [but] recent drivers may break SLI scaling that once worked, and even a new game patch may affect SLI game performance adversely.”
Fast forward 2-1/2 years to today. There are still very few mGPU games, and NVIDIA has relegated SLI to legacy. They will not be adding any new SLI profiles, and the only Ampere card that supports it is the RTX 3090 using a new NVLink bridge – for benchmarking – to set world records in synthetic tests like 3DMark. NVIDIA has this to say about SLI support “transitioning”, quoted in part:
“NVIDIA will no longer be adding new SLI driver profiles on RTX 20 Series and earlier GPUs starting on January 1st, 2021. Instead, we will focus efforts on supporting developers to implement SLI natively inside the games. We believe this will provide the best performance for SLI users.
Existing SLI driver profiles will continue to be tested and maintained for SLI-ready RTX 20 Series and earlier GPUs. For GeForce RTX 3090 and future SLI-capable GPUs, SLI will only be supported when implemented natively within the game.
[Natively supported] DirectX 12 titles include Shadow of the Tomb Raider, Civilization VI, Sniper Elite 4, Gears of War 4, Ashes of the Singularity: Escalation, Strange Brigade, Rise of the Tomb Raider, Zombie Army 4: Dead War, Hitman, Deus Ex: Mankind Divided, Battlefield 1, and Halo Wars 2.
[Natively supported] Vulkan titles include Red Dead Redemption 2, Quake 2 RTX, Ashes of the Singularity: Escalation, Strange Brigade, and Zombie Army 4: Dead War.
… Many creative and other non-gaming applications support multi-GPU performance scaling without the use of SLI driver profiles. These apps will continue to work across all currently supported GPUs as it does today.”
It looks bleak for SLI’s future and dev-supported mGPU titles are beyond rare. So we tested our 40-game benching suite and identified just nine games that scaled well with RTX 2080 Ti. Of the baker’s dozen games that NVIDIA lists and that we have, Civilization VI using the ‘Gathering Storm’ expansion benchmark did not scale, and Red Dead Redemption 2 crashed when we tried to use it. The other games on their list are old and run great on any modern GPU negating any reason to use SLI anyway, except perhaps for extreme supersampling.
We didn’t bother listing the performance of games that barely scale, scale negatively, or exhibit issues when SLI is enabled. That list is very long. Of course, there are still SLI enthusiasts who tweak their games with NVIDIA Inspector and roll back to old drivers to indulge their hobby – but we use the latest drivers without tediously trying workarounds that may or may not be successful.
Build-Gaming-Computers.com was able to identify a total of 57 games (out of many thousands) that scaled well or partially with SLI in 2020, and they have concluded that overall it isn’t worth the trouble or the expense of maintaining two finicky cards using an extra-large PSU. However, here are nine relatively modern games that we tested that show SLI scaling without a lot of microstutter or other issues associated with them.
SLI Gaming Summary Charts
Here are the summary charts of 9 games and 3 synthetic tests that scale with mGPU or SLI. The highest settings are always chosen and the settings are listed on the chart. The benches were run at 2560×1440 and at 3840×2160. The first column represents the performance of a single RTX 2080 Ti, the second represents two RTX 1080 Tis, and the third column gives RTX 3090 results. ‘X’ means the game was not tested.
Most results show average framerates, and higher is better. Minimum framerates are next to the averages in italics and in a slightly smaller font. Destiny 2, benched with OCAT show average framerates, but the minimums are expressed by frametimes (99th-percentile) in ms where lower is better.
Although SLI scaling is good with these nine games at 4K, there are some issues with 256×1440 and framerate caps. We would prefer to play these nine games with a RTX 3090 that has no issues with microstutter. However, synthetic benches look pretty good.
Even though the RTX 2080 Ti has been surpassed by both the RTX 3080 and the RTX 3090, our PC scored in the top 1% of all PCs using Fire Strike Ultra.
If you are a professional overclocker and/or want to set a world record, we would suggest buying two RTX 3090s for that purpose instead of using any two Turing video cards.
We cannot recommend SLI to any gamer unless they have a very large library of old(er) games that they revisit and play regularly and who don’t mind the issues associated with tweaking and maintaining SLI profiles using old drivers. Then there is the added inconvenience of disabling SLI each time most modern games are played. Besides, there are the additional issues of heat and noise coupled with using two powerful cards with a large PSU, not to mention the expense of buying a second card, and the higher cooling power bills associated with using SLI during the warm months of Summer.
So let’s look at Creative applications next to see if 2 x RTX 2080 Tis are a viable option versus the RTX 3090 starting with the Blender benchmark.
Blender 2.90 Benchmark
Blender is a very popular open source 3D content creation suite. It supports every aspect of 3D development with a complete range of tools for professional 3D creation. We will look at Blender rendering later in this review, but here are the official benchmark results.
For the following results, lower is better as the benchmark renders a scene multiple times and gives the results in minutes and seconds. First up, two RTX 2080 Ti’s using the RTX TITAN NVLink bridge with CUDA.
There is no difference with SLI enabled or disabled. Here is the chart comparing the performance of a single RTX 2080 Ti with two versus a RTX 3090.
Performance is worse using the second RTX 2080 Ti as the benchmark is not optimized for a second video card. However, we will try to render a large scene in Blender as we show later.
Next we look at the OctaneBench.
OctaneBench allows you to benchmark your GPU using OctaneRender. The hardware and software requirements to run OctaneBench are the same as for OctaneRender Standalone and we shall also use OctaneRender for a specific rendering test later, under “Professional Apps”.
We have a win for 2 linked RTX 2080 Ti’s scaling. Here is the summary chart.
Next, we move on to AIDA64 GPGPU synthetic benchmarks that are built to scale with mGPU.
AIDA64’s benchmark code methods are written in Assembly language, and they are generally optimized for popular AMD, Intel, NVIDIA and VIA processors by utilizing their appropriate instruction set extensions. We use the Engineer’s full version of AIDA64 courtesy of FinalWire. AIDA64 is free to to try and use for 30 days. CPU results are also shown for comparison with the video cards’ GPGPU benchmarks.
First the results with a pair of RTX 2080 Tis.
Now the RTX 3090:
Again the pair of linked RTX 2080 Tis are faster at almost all of AIDA64’s GPGPU benchmarks including the RTX 3090. So let’s look at Sandra 2020 which which is also optimized for mGPU.
To see where the CPU, GPU, and motherboard performance results differ, there is no better comprehensive tool than SiSoft’s Sandra 2020. SiSoftware SANDRA (the System ANalyser, Diagnostic and Reporting Assistant) is a excellent information & diagnostic utility in a complete package. It is able to provide all the information about your hardware, software, and other devices for diagnosis and for benchmarking.
There are several versions of Sandra including a free version of Sandra Lite that anyone can download and use. Sandra 2020 R10 is the latest version, and we are using the full engineer suite courtesy of SiSoft. Sandra 2020 features continuous multiple monthly incremental improvements over earlier versions of Sandra. It will benchmark and analyze all of the important PC subsystems and even rank your PC while giving recommendations for improvement.
We ran Sandra’s extensive GPGPU benchmarks and charted the results summarizing them below. The performance results of the RTX 2080 Ti are compared with the performance results of the RTX 3090, and versus the two linked RTX 2080 Tis.
In Sandra synthetic GPGPU benchmarks which are optimized for mGPU, the linked RTX 2080 Tis are faster than the RTX 3090 and they generally scale well over a single Ti. Next we move on to SPECworkstation 3 GPU benchmarks.
SPECworkstation3 (3.0.4) Benchmarks
All the SPECworkstation 3 benchmarks are based on professional applications, most of which are in the CAD/CAM or media and entertainment fields. All of these benchmarks are free except to vendors of computer-related products and/or services.
The most comprehensive workstation benchmark is SPECworkstation 3. It’s a free-standing benchmark which does not require ancillary software. It measures GPU, CPU, storage and all other major aspects of workstation performance based on actual applications and representative workloads. We only tested the GPU-related workstation performance. We did not use SPECviewperf 13 since SPECviewperf 2020 is coming out in mid-October.
Here are the SPECworkstation3 results for two linked RTX 2080 Tis. Higher is better since we are comparing scores.
The RTX 3090 was unable to complete two benches, probably because of a conflict with Ampere’s new drivers. But there is no scaling whatsoever, or negative scaling for the NVLinked RTX 2080 Tis. So we questioned the people who are responsible for maintaining the SPECworkstation benchmarks:
Q: I am comparing its SPECworkstation results with 2 X RTX 2080 Ti that are connected using a RTX Titan NVLink HB Bridge. Are using two GPUs in this manner supported by the benchmark?
A: The short answer is “no”, it will not produce the desired scaling effect if you bridge the two cards. The longer answer has more to do with your expectations and that the benchmark does not explicitly do anything to preclude multi-GPU scenarios from improving support, but it does not have any code that explicitly enables it.
The graphics portions of SPECworkstation come from SPECviewperf which, in turn, is based on recordings of real-world applications. The creation of a rendering context to draw 3D scenes is done in a way that tries to very closely mimic the real-world application and thus, if the real-world application would benefit from multiple GPUs, so might the viewsets that comprise the benchmark.
The GPU compute portions of SPECworkstation run on only a single GPU. We are working toward multi-GPU support in the next major version but it’s not in there now.
So mGPU scaling may depends on if a benchmark is optimized for it or not. However, let’s next look at some professional applications where a large memory buffer makes a big performance improvement over having a smaller one.
Creative Applications with Large Memory Workloads
Rendering large models, detailed scenes, and high-resolution textures require powerful GPUs with a lot of vRAM. Render artists using the highest quality renders, require high capacity GPU memory which allows them to create more detailed final frame renders without needing to reduce the quality of their final output, or to split scenes into multiple renders which take a lot of extra time. Until now, no GeForce has been equipped with 24GB of vRAM while the RTX 2080 Ti offers 11GB. Let’s look at three pro apps that can use much more than 11GB and also test render times. First up is OTOY OctaneRender.
OctaneRender is the world’s first spectrally correct GPU render engine with built-in RTX ray tracing GPU hardware acceleration. The RTX 3090 allows large scenes to fit completely into the 24 GBs of GPU memory so out-of-core rendering is not necessary, providing faster than rendering times using out-of-core data for GPUs with lesser memory capacity. We tested the RTX 3090/24GB against the bridged RTX 2080 Tis.
Following NVIDIA’s very specific instructions, we rendered a very large detailed image. Looking closely, we see that out-of-core data was not needed since the entire render fit into the 24GB vRAM buffer, and the large image provided only took 45 seconds to render.
We tested the NVLinked RTX 2080 Tis, and it took much longer at 2 minutes and 27 seconds because it requires much slower out-of-core memory. The 11GB vRAM of the RTX 2080 Tis are evidently not pooled for this render.
However, a pair of RTX 2080 Tis are faster than a single card and the results are summarized in the chart below.
So for rendering, it appears that two linked RTX 2080 Tis are faster than one in OTOY rendering. Let’s look at Blender next.
Blender is a popular free open source 3D creation suite that supports modeling, rigging, animation, simulation, rendering, compositing, motion tracking, video editing, and the 2D animation pipeline. NVIDIA’s OptiX accelerated rendering in Blender Cycles are used to accelerate final frame rendering and interactive ray-traced rendering in the viewport to give creators real-time feedback without the need to perform time-consuming test renders. The 24 GB framebuffer on the RTX 3090 allows it to perform final frame together with interactive renders that may fail due to a smaller vRAM framebuffer on the RTX 3080 or issues with linked RTX 2080 Tis.
This large render took 31.24 seconds using the RTX 3090 but it caused an error when we tried fitting the scene into the linked RTX 2080 Ti’s framebuffer and it could not complete the render as shown below.
However, it did render with a single RTX 2080 Ti, and here is the summary chart.
So in regard to mGPU and Blender rendering, it appears that “it depends”.
Finally, we looked at Blackmagic Design DaVinci Resolve and 8K Redcode RAW projects.
Blackmagic Design DaVinci Resolve | 8K Redcode RAW projects
Blackmagic’s DaVinci Resolve combines professional 8K editing, color correction, visual effects and audio post production into one software package. With 8K projects featuring 8K REDCODE Raw (R3D), files will use most of the memory available on a RTX 3080 which result in out of memory errors particularly when intensive effects are added. Indeed, the RTX 3090/24GB was able to perform a very intensive LFB project quickly using an 8K R3D RED CAMERA clip on an 8K timeline with a temporal noise reduction processing effect applied. In contrast, the RTX 3080 and a pair of RTX 2080 Tis just generated error messages which means that we would have to workaround – taking a lot of extra time and effort. There is really no quantitative benchmark here.
Older single cards – the RTX 2080 Ti and the TITAN Xp – can run many of these workloads with various degrees of success without errors, but they are much slower than the RTX 3090.
After seeing the totality of these benches, creative users will probably prefer to upgrade their existing systems with a new RTX 3090 based on the performance increases and the associated increases in productivity that they require. The question to buy the RTX 3090 or a second RTX 2080 Ti should probably be based on the workflow and requirements of each user as well as their budget. Time is money depending on how these apps are used. If a professional needs a lot of framebuffer, the RTX 3090 is a logical choice. Hopefully the benchmarks that we ran may help you decide.
Let’s head to our conclusion.