Although multi-GPU technologies for gaming, like SLI and CrossFire, have been dead for many years now, multi-GPU is still clearly useful for other applications, as seen in a demo in FluidX3D. Under the handle ProjectPhysX, FluidX3D developer Dr. Moritz Lehmann has demonstrated a dual-GPU setup combining Intel's A770 with one of Nvidia's Titan Xp, which is approaching its seventh birthday. Though this is a strange combination at best, the results show that the two GPUs combined pack a punch.
For a multi-GPU demo, it was surprisingly simple. Dr. Lehmann used Acer's Predator A770 16GB and an Nvidia Titan Xp, each of which simulated and rendered half of the simulation. While DX12 and Vulkan are (or were) the highest-profile APIs to perform multi-GPU workloads, FluidX3D actually runs in OpenCL, which incidentally is developed by Khronos Group, the same developers behind Vulkan.
Precise performance figures and data weren't offered, but Dr. Moritz says the dual-GPU setup outperformed each individual GPU by roughly 70%, which makes sense as the A770 and Titan Xp perform about the same in FluidX3D per the software's scoreboard. It took an hour and 13 minutes to compute the simulation and then around 14 minutes to render it, which means either card on its own would take around two whole hours just to simulate.
While this combination may seem like it was chosen simply for the sake of humor, there was actually some good reasoning for pairing the A770 with the Titan Xp. As the developer says, it makes very little sense to pair a very powerful GPU with a much weaker one, and at least for FluidX3D, it's ideal to have similar memory capacity and bandwidth. With the A770 running its 16GB at 560GB/s and the Titan Xp running its 12GB at 548GB/s, the match actually makes some sense.
With a 70% performance boost, it might seem hard to believe the gaming industry passed up on multi-GPU technology, a sentiment that was echoed by many commentators on the FluidX3D demo. After all, DX12 and Vulkan have great support for the technology, GPU-to-GPU linking technology is more robust than ever, and the latest versions of PCIe are very fast.
In reply to those comments, Dr. Moritz gave his analysis and pointed out a few issues with multi-GPU setups for gaming. The biggest issue would be the cost that goes into developing multi-GPU solutions for games, which used to fall on the shoulders of Nvidia and AMD but shifted to game developers with the arrival of DX12 and Vulkan, which have powerful multi-GPU features but require manual tuning to work effectively. However, game developers don't profit if they implement multi-GPU, which has always been extremely niche, even among PC enthusiasts.
The industry instead took a different direction to get more performance, focusing more on single-GPU setups by making even bigger flagships, which are now "so hilariously large that you can't even fit a single one in a normal PC case, let alone 2" according to Dr. Moritz. Considering many GPUs like the RTX 4090 Founders Edition take up three slots, it's hard to disagree. Today, multi-GPU is thriving in data centers, supercomputers, and AI-focused systems, where the cost of implementing support is more than worth it.
