For a while, the obvious answer to "how do I make my gaming PC faster?" appeared to be wonderfully excessive: buy another graphics card.
Nvidia called its system SLI. AMD called its competing system CrossFire. Both promised to split rendering work across multiple GPUs, and benchmark charts could make the future look gloriously simple. One expensive card was fast; two could be much faster; three was the sort of thing hardware sites tested because somebody had to find the edge.
The catch was that games are not spreadsheets. Two GPUs did not automatically mean twice the experience, and the gap between a high average frame rate and a smooth game eventually became impossible to ignore.
The benchmark dream was real enough to be convincing
Period testing shows why multi-GPU had a serious audience rather than being pure vanity. Tom's Hardware's 2011 three-GPU study found substantial scaling in several games, especially at higher resolutions where the graphics workload was large enough to keep multiple cards busy.
The same testing also showed the limits immediately. A third card could add very little at lower resolutions, and the extra boards increased power use and complexity.
Still, supported games could look spectacular on paper. PC Gamer's 2018 *Far Cry 5* testing found CrossFire scaling as high as 98% at 4K for one tested configuration, while two GPUs were the only setups in that test approaching consistent 4K/60 at maximum settings.
That result came with a giant asterisk. PC Gamer also called multi-GPU support a "crapshoot," noting that plenty of games either did not support it or received support weeks or months after launch. Even inside one well-behaved game, different SLI configurations scaled by very different amounts.
The dream required the game, driver, CPU, resolution and profile to cooperate.
Micro-stutter exposed the difference between FPS and smoothness
Average frame rate was especially good at hiding one of multi-GPU's ugliest problems: micro-stutter.
With alternate-frame rendering, different GPUs can produce consecutive frames. If those frames arrive at uneven intervals, an average counter can look healthy while motion still feels irregular.
Tom's Hardware investigated that problem directly in 2011 and concluded it was real, not just forum folklore. Its testing also showed that power consumption climbed as GPUs were added, making the price of scaling more than the cost of the second card.
AMD spent years improving frame pacing. PC Gamer covered Catalyst 13.8 beta drivers in 2013 because they specifically targeted CrossFire micro-stutter on supported configurations. The fix itself proved the larger point: multi-GPU performance depended on software coordination as much as raw silicon.
Developers had another burden. A game or driver profile had to know how to divide work safely, and new rendering techniques did not always fit the assumptions that older SLI and CrossFire systems were built around.
The industry eventually stopped pretending support was automatic
By 2020, Nvidia was preparing to stop adding new SLI profiles to its Game Ready drivers and shift explicit multi-GPU responsibility toward developers using APIs such as DirectX 12 and Vulkan.
PC Gamer described the change as SLI being put on life support. That was dramatic wording, but the underlying transition was real: instead of Nvidia maintaining a large catalog of per-game driver profiles, developers would need to implement multi-GPU support themselves if they wanted it.
The consumer market had already been moving that way. Powerful single GPUs kept improving, while the cost, heat, power draw and inconsistent game support of two-card rigs became harder to justify.
CrossFire followed the same broad decline as an everyday gaming recommendation. Multi-GPU ideas did not disappear from computing — modern accelerators, rendering farms and multi-die packages can use multiple pieces of silicon in very different ways — but that is not the old promise that two consumer graphics cards in one PC will transparently behave like one faster gaming GPU.
SLI and CrossFire were not failures because they never worked. The interesting thing is that they did work, sometimes brilliantly.
They faded because "sometimes brilliantly" is a rough foundation for an expensive upgrade. Once a single-card setup became the simpler route to consistent frame pacing and broad compatibility, the future stopped looking like a motherboard full of GPUs and started looking, again, like one very fast card.
Reporting notes