Intel’s Arc A770 has been paired with NVIDIA’s Titan Xp, a $1200 behemoth from the Pascal days when GPUs were more affordable. These two graphics cards were coupled for a FluidX3D run by ProjectPhysX and the results were posted. The dual-GPU configuration took 73 minutes to compute the simulation and 14 minutes to render it. This implies that each GPU would take close to two and a half hours for the same workload.
FluidX3D leverages OpenCL, an open-souce GPGPU API used by hundreds of applications to accelerate their workloads. It’s Vulkan’s predecessor, but hasn’t lost much popularity in the scientific community. If multi-GPU setups can yield such gains, you might wonder why SLI and XFX were abandoned after being used for so long.
In the last half decade, temporal antialiasing and upscaling have become a vital component of several PC games. This includes DLSS, FSR, and TAA. These filters require data from older frames that is accumulated and superimposed on the current frame to improve visual fidelity. Unfortunately, SLI is mostly based on AFR (Alternate Frame Rendering) where individual GPUs render consecutive frames separately. Using this rendering technology with temporal filtering induces a lag that nullifies most SLI gains.
Of course, there are other reasons why SLI/XFX was abandoned such as the minute user count and the extra workforce required to implement it separately in every game. But, from what I recall, the wide scale adoption of temporal filters was the final nail in the coffin.