For the last fourteen years, AMD and Nvidia have both supported multi-GPU configurations. Initially, these types of systems were limited to two GPUs using either Nvidia’s SLI technology (the acronym dates back to 3dfx) or AMD’s Crossfire. When Windows Vista launched, the supported number of video cards grew to encompass three-way or even four-way configurations.
That era, it seems, is coming to an end. When Nvidia launched its Pascal family, it made it clear SLI would only be supported on high-end cards and would be limited to just two GPUs, even then. AMD has been offering Crossfire on lower-end cards than Nvidia does, so there was some hope that the company would continue to push three-way or four-way configurations. AMD has now confirmed this isn’t happening, in an email exchange with PC World’s Brad Chacos.
“We have delivered two-way mGPU support in games,” and AMD representative told PCWorld. “Three- and four-way configurations will continue being supported in compute and professional applications.”
That’s the effective end of three-and-four-way graphics cards, and frankly, it’s been a long time coming. Multi-GPU configurations with more than two cards have never worked particularly well outside synthetic benchmarks. The more graphics cards you have in a system, the harder it is to keep the cards synchronized and rendering properly. In DX9 or DX11, data is duplicated across each card. If you have four GTX 980s or Radeon Fury’s, you don’t have 16GB of VRAM — you have 4x4GB of VRAM. There’s an enormous amount of data that must be synchronized across the GPU cluster, which is one reason why AMD and Nvidia both historically provided additional connectors to increase GPU-to-GPU bandwidth.
While this slide came from Nvidia, it illustrates the larger point — breaking workloads down across multiple graphics cards is very difficult.
But if you’ve paid attention to any of the discussions of microstutter in mGPU configurations, you know that this doesn’t always work as smoothly as you’d like, even in two-way mGPU configurations. Nvidia’s multi-GPU implementations have typically been more robust than AMD’s, but using multiple GPUs has always been somewhat hit-or-miss. Some games don’t support mGPU at all. Some support the option, but implement it so badly you’re better off not using SLI or Crossfire at all in that title. Rather than delivering a guaranteed 1.5x – 2x speed-up, SLI and Crossfire are more like conditional boosts. In the right game, with the right API, and the right driver, you see huge gains.
Part of the problem is there’s an intrinsic catch-22 built into the concept. Gamers use multi-GPU configurations because they want to run games as quickly as possible. But the faster the game is, the less time the GPUs have to render their frames. As detail level rises, the amount of pressure on the GPU’s VRAM rises as well. The more GPUs you use, the harder it is for the system to keep them synchronized. The conditions that make people want to use multi-GPU configurations, in other words, also make it more difficult to use the capability. And because running 3-4 GPUs doesn’t give you 3-4x the RAM, you weren’t necessarily future-proofing a system in this configuration.
DirectX 12 presents a theoretical solution to some of these issues, since it can give developers the option to split the rendering workload and distribute it as they see fit. We discussed this when Ashes of the Singularity demonstrated Nvidia and AMD GPUs working side-by-side in the same rig. To our knowledge, however, no other game supports this functionality, and Nvidia surely wants to keep it that way.
Overall, dropping support for three-and-four-way systems may make enthusiasts unhappy over their 3DMark scores, but it’s not going to impact the overall market. Only a fraction of gamers deploy multi-GPUs, and only a fraction of that fraction try to push above two cards.