For roughly a decade, from the mid-2000s to the mid-2010s, the fastest graphics card you could buy often wasn't a single GPU at all. Both NVIDIA and AMD regularly took their flagship chip, put two of them on one board, and sold the result as the ultimate halo product. Cards like the GeForce GTX 690 and Radeon R9 295X2 were engineering showpieces: huge, power-hungry, expensive, and capable of performance no single chip could match at the time. Today the format is essentially extinct in gaming, but the idea behind it never really went away.

A dual-GPU card is essentially two complete graphics cards sharing one circuit board. Each GPU has its own memory and power delivery. A PCIe switch chip on the board splits the slot's connection between the two GPUs, and a link between them does the same job as an external SLI or CrossFire bridge. To the computer, it's not one big GPU. The system sees two separate GPUs, and the driver has to split the work between them. The most common method was Alternate Frame Rendering (AFR), where GPU one draws frame 1, GPU two draws frame 2, and so on. When it worked, frame rates could approach double those of a single GPU. When it didn't, the second GPU sat idle. - Memory doesn't add up. Each GPU needs its own copy of the textures and scene data, so the memory is mirrored, not pooled. A card advertised with 8 GB (like the R9 295X2) effectively behaves like a 4 GB card. Marketing usually quoted the total anyway. - Microstutter. AFR can deliver frames at uneven intervals. A game might report 60 fps but feel less smooth than a steady 45 fps on a single GPU. This became a major topic in reviews around 2012–2013 once frame-time testing became common. - Driver dependency. Scaling depended on game-specific SLI/CrossFire profiles. New games often launched without support, and some never got it. - Heat and power. Two flagship chips on one board meant extreme cooling demands. The R9 295X2 shipped with a factory liquid cooler and a rated board power around 500 W.

NVIDIA GeForce 7950 GX2

2006 | 2 x G71 | 2 x 512 MB | GDDR3

NVIDIA GeForce 9800 GX2

2008 | 2 x G92 | 2 x 512 MB | GDDR3

NVIDIA GTX 295

2009 | 2 x GT200B | 2 x 896 MB | GDDR3

NVIDIA GTX 590

2011 | 2 x GF110 | 2 x 1.5 GB | GDDR5

NVIDIA GTX 690

2012 | 2 x GK104 | 2 x 2 GB | GDDR5

NVIDIA GTX TITAN Z

2014 | 2 x GK110B | 2 x 6 GB | GDDR5

AMD Radeon HD 3870 X2

2008 | 2 x R680 | 2 x 512 MB | GDDR3

AMD Radeon HD 4850 X2

2008 | 2 x R700 | 2 x 512 MB | GDDR3

AMD Radeon HD 4870 X2

2008 | 2 x R700 | 2 x 1024 MB | GDDR5

AMD Radeon HD 5970

2009 | 2 x Hemlock | 2 x 1024 MB | GDDR5

AMD Radeon HD 6990

2011 | 2 x Antilles | 2 x 2 GB | GDDR5

AMD Radeon HD 7990

2013 | 2 x Malta | 2 x 3 GB | GDDR5

AMD Radeon R9 295X2

2014 | 2 x Vesuvius | 2 x 4 GB | GDDR5

AMD Radeon PRO Duo

2016 | 2 x Capsaicin | 2 x 4 GB | HBM

AMD Radeon PRO Duo Polaris

2017 | 2 x Ellesmere | 2 x 16 GB | GDDR5

Asus's Republic of Gamers division took the concept even further with limited-edition cards that NVIDIA and AMD never made themselves. They often used full-spec chips where the official dual-GPU card was cut down, along with more memory and higher clocks.

Asus ROG Mars GTX 285

MARS/2DI/4GD3

Asus ROG Mars GTX 580

MARS-II/2DIS/3GD5

Asus ROG Mars GTX 760

MARS760-4GD5

Asus ROG Ares

ARES/2DIS/4GD5

Asus ROG Ares II

ARES2-6GD5

Asus ROG Ares III

ARESIII-8GD5

These were produced in small numbers, often around 1,000 units or fewer, and are now collector's items. Several trends ended the dual-GPU card: - Single GPUs got big enough. As flagship chips grew, one GPU could deliver what used to take two, without the scaling problems. - Game engines moved away from AFR. Modern rendering techniques reuse data from previous frames (temporal anti-aliasing, for example), which conflicts with having two GPUs each render alternate frames. - DirectX 12 and Vulkan shifted responsibility. Under the new APIs, multi-GPU support became something each game developer had to build in themselves, rather than something the driver handled. Very few did. - The vendors stepped back. NVIDIA stopped adding new SLI profiles from 2021 and limited SLI to its top cards before dropping it entirely. AMD quietly de-emphasized CrossFire around the same time. The last gaming dual-GPU cards were the GTX Titan Z and R9 295X2 in 2014. In the professional market, Apple's Mac Pro kept the format alive a little longer with the Radeon Pro Vega II Duo and W6800X Duo.


The basic idea of combining several pieces of silicon to go beyond what one chip can do has come back in the datacenter. The difference is that modern designs connect the dies so closely and with so much bandwidth that software sees one GPU. NVIDIA's Blackwell (B200) and Rubin accelerators put two large dies in one package. AMD's Instinct MI300 and MI400 series go further, with many compute chiplets in each package. In a sense, the dual-GPU card was a first attempt at what chiplets do properly today.