Little by little and indiscriminately, both AMD and NVIDIA are watching time pass without their multi-GPU technologies adapting to the new times. Why are you letting NVIDIA SLI die? Is NVIDIA really to blame for this? or instead, are there more culprits? Let’s find out by reviewing a little about the history of SLI and what’s to come in the future.
The gaming world does not stop and currently as in the past there is a great dependence on graphics power to reach the beloved 4K @ 144 Hz, 1080p @ 360 Hz or the new 8K @ 120 Hz. But for a while now, specifically from Pascal and the GTX 1080 Ti so far with the RTX 3090, NVIDIA has been backing down with its SLI technology and AMD has already declared that CrossFire is not a priority and indirectly excluded it from their plans. It seems that for now, not only SLI is dead, but any multi-gpu gaming technology.
SLI was never about performance per dollar, quite the opposite. The technology was always used by enthusiast gamers as a way of pushing a single rig to its maximum possible performance, even if that means diminishing returns in terms of value. Multi-gpu configurations powering desktops go as far in the past as 1998, and it seems it will end in 2020.
Believe it or not, in the 90s there were 3 major competitors in the GPU market. NVIDIA, ATI (Purchased by AMD) and 3dfx. The latter became popular for implementing hardware acceleration of 3D graphics. One of the company's successes was the Voodoo2, which was capable of running in multi-gpu configurations, as shown here:
Later 3dfx suffered because of Microsoft releasing Direct3D, which replaced Glide (3dfx's proprietary API) and therefore the value of the company's graphic cards. Nvidia acquired 3dfx on December 15, 2000.
With the introduction of the GeForce 7 series the gaming community was amazed by quad-gpu configurations. The next generations until 2020 continued to see SLI support, although sometimes it didn't make sense.
The GTX 590 is a card that by itself can still render modern games or at the very least popular games such as League of Legends at high frame rates. An impressive feat for a 9 year old graphics card.
Having a TDP of 250W this card was more than enough to make any power supply cry. Imagine those quad-Titan Xp configurations that soon started to appear after the card was released.
Officially the last bastion of SLI. The Ray tracing capable gpu was already more than enough for anything you throw at it. SLI is making less sense as the performance gains are marginal when compared to the money spent, and GPU's were powerful enough so that people wouldn't be tempted to spend more.
NVLink interface that allows to build graphics cards with multi-GPUs, but NVIDIA doesn't integrate this technology into its gaming graphics cards. In the specific case of the latest generation, the RTX 30 Series, only the RTX 3090 has a Link interface that allows NVIDIA if it wanted to make a Multi-GPU graphics card.
Future for multi-gpu configuration is still unknown and the market is too small for NVIDIA to worry about. Instead efforts are into AI acceleration and the purchase of arm.
In an industry there is rarely a single culprit, much less if the two main and only GPU manufacturers agree to forget about a technology where they have invested huge amounts of money and which boosts their sales of graphics cards in a small part. Why give up selling less when the outlay has already been made for years and now the investment is minimal compared to creating the infrastructure?
The fault lies with Microsoft and the developers, but let's go by parts, because neither NVIDIA nor AMD are spared from the technical bankruptcy of their systems.
Modern APIs have and do not bring us impressive technologies and performance improvements, but both DX12 and Vulkan were confident in putting optimization of something as extremely delicate as the multi-GPU in the hands of the studios. This confidence was intended to improve optimization to such an extent that scalability was very similar to that offered in the CPU cores, something very transparent for the hardware and above all very worked.
The reality is very different. The studios had to take care of a low-level optimization that until now the DX11 API and NVIDIA were working to optimize their games to their Multi GPU technologies, so when they realized the huge expense of money that they supposed to do what these two, the studies went backwards and little by little they have been optimizing for a single GPU.
Neither at the hardware level nor at the software level. Lisa Su's team continued with their concept of connection through PCIe with the consequent problem of internal bandwidth and the bottleneck with VRAM. On the other hand, NVIDIA was not far-sighted and to each generation after Pascal it has launched a new interconnection bridge that is not exactly cheap, causing the user to have to pay extra money to buy it, not being able to use the previous one.
As if this were not enough, the studies reproached the null expense of both companies in their technologies under the new APIs and assured that they should have both financial incentives and human resources from AMD and NVIDIA if they wanted SLI and CrossFire to be afloat.
The negotiation did not even have a beginning and is currently in a stalemate, on standby, just as a person may be in a coma awaiting either his death or a slow and late recovery.
AMD no longer offers any support and NVIDIA only does it for titles under DX11 or lower API, but nothing under DX12 or Vulkan logically. Chronicle of an announced death, a technology that had its peak, its nights of stars and now sees how it sits waiting as the grim reaper comes for them ...