Why Multi-GPU Gaming Setups Died and What Replaced Them

Multi-GPU gaming was never killed in one announcement. The interconnect was removed one product generation at a time, and by the RTX 40 series there was no longer anywhere to plug a bridge. In 2009 and 2010, reviewers were benchmarking three Radeon HD 5870s in CrossFire, the setup AMD called TriFire. What follows is what happened to that idea, and what now sits in the slot the second and third cards used to occupy.

SLI and CrossFire offered a simple deal: bridge identical GPUs and get more frames. In How-To Geek's retrospective, Monica J. White describes what that deal looked like in practice: "One game might scale beautifully and make the enormous investment feel worthwhile, while another barely used the second card at all."

The sources for this piece do not include the published three-card and four-card scaling figures from 2009 and 2010 reviews, so we will not make up a percentage for them. The documented pattern is enough. Scaling depended on the game. Multi-card rigs were also prone to microstutter, which means uneven frame delivery hiding behind a healthy average FPS. An average frame rate describes throughput. Microstutter is a frame-time problem, and an average does not show it.

How-To Geek's timeline is specific to Nvidia. Three-way and four-way configurations died out in the Pascal era. The RTX 20 series swapped the SLI bridge for NVLink, but only on the most expensive cards. The RTX 3090 was the last consumer GeForce card with NVLink. Neither the RTX 40 nor the RTX 50 series, including the RTX 5090, has any multi-GPU interconnect.

The software side went first. Nvidia stopped issuing new SLI driver profiles for RTX 20-series and older cards, which moved the entire job onto game developers. DirectX 12 and Vulkan do allow explicit multi-GPU rendering. Almost nobody implements it, because, as White puts it, "the extra effort just wasn't worth it for such a small audience." Our sources do not date AMD's retirement of CrossFire, so that side of the history stays open here.

Nvidia's replacement for a second GPU is frame generation on the first one. According to Nvidia, DLSS 4.5 is supported in 400+ games and apps and adds a 6X Dynamic Multi Frame Generation mode. DLSS 5 launched September 3, 2026 in NBA 2K27. Nvidia says early development builds ran on dual RTX 5090s, and that a "5X performance gain in just six months" brought it down to one card. That is a vendor claim, and nobody outside Nvidia can check it without the early builds.

The headline numbers need the same caution. Club386 notes that Nvidia's 594 fps figure for the RTX 5090 at 1440p stacks DLSS Quality on top of 6x Multi Frame Generation. The FPS Review's roundup, citing DSOGaming, has the 5090 falling from 180 to 134 FPS at native 1080p with ray tracing once neural rendering is on. That is roughly a 26 percent drop. The same roundup also reports a general loss of 50 to 60 percent. Those two figures do not agree, and settling it would take per-game frame-time captures that nobody has published yet.

The load also lands on one card now. TheFPSReview reports that some DLSS 5 test configurations went from 480W to 720W. Power that used to be split across two slots now runs through a single 12V-2x6 connector.

A ReShade add-on called the Neural Coprocessor, or MGPU Bridge, gives DLSS 5 its own dedicated GPU and moves frames between the two cards over PCIe. It was reported by Time.news, a low-credibility source. The reported gains are real against single-card DLSS 5, but the setup is still slower than simply turning DLSS 5 off. It is the dedicated PhysX card again, not SLI.

Buy the single best card your budget allows and spend the rest on the power supply. DLSS 5 required an RTX 50-series card at launch. Tech Insider reports that RTX 40 support has been promised for later in fall 2026. Pricing is the catch. As of August 2026, RTX 50 cards were selling 15 to 119 percent above MSRP, and the $1,999 RTX 5090 was reselling for up to $4,381. The money that once bought a second card now mostly covers the markup on the first.