Raja Koduri is one of the best known figures among video card enthusiasts: currently executive vice president and general manager ofAccelerated Computing Systems and Graphics (AXG) Group di Intel, Koduri has worked for ATI and AMD for many years, interspersing a period with Apple between the two companies. In Intel he presides over all activities related to the development of next generation GPUs and in general of all acceleration systems for parallel computing.
At Vision 2022, on the occasion of a roundtable with the European press, I was able to discuss with Raja on the topic of photorealism in graphics with virtual reality viewers starting from an anecdote from the past. In 2016, at an event organized by AMD’s Radeon Technology Group, he predicted that to achieve photorealism with VR headsets it would be necessary achieve a resolution of 16K per eye, with a vertical refresh rate of at least 240Hz. Back then I had calculated the reference computing power needed on paper in something equal to 100 NVIDIA GeForce GTX 980 video cards in parallel.
Raja Koduri holds up an Intel Server GPU Data Center card
Remembering this prediction, as well as the estimated time related to the required processing power, I asked Raja how much work has been done in these 6 years to reach that goal and how much distance there is, in the latest GPU architectures, to achieve true photorealism with VR headsets.
According to Raja, the last 6 years have passed we’re much closer to photorealism with VR than we might have thought when that prediction was made. We must first think that the complexity of VR content has changedhigher than we could have predicted 6 years ago, but at the same time the goal of photorealism in VR can now make the most of the techniques of “neural net based reconstruction“come quelle XeSS (Xe Super Sampling) e DLSS (Deep Learning Super Sampling), also taking advantage of part of the work developed for the neural radiance filter.
Therefore, the picture related to photorealism changes with VR viewers, an objective that can now be more easily achieved thanks to the contribution of these techniques that in 2016 could not yet be taken into consideration. But there is still a lot of work to be done in collaboration with the software developer communitywhich must be put in a position to be able to exploit these techniques in their apps, in particular the neural net based pipeline.
In addition to this it is necessary, for producers, making GPU architectures increasingly efficient in managing neural networks within the graphics pipeline, so as to arrive at a virtual reality that is, in the eyes of the user, indistinguishable from the real world. It is for this reason that according to Raja Koduri we are today much closer to that goal of photorealism in VR than what he thought in 2016, because in that year it was not yet clear what the series of technological innovations could be able to make accessible in the near future that result. Now we know: neural networks and all the techniques, within the graphics pipelines of the GPUs, which allow them to be exploited to the fullest.