Nvidia’s Tom Peterson answers Geforce RTX/Turing questions
Nvidia’s Tom Peterson answers Geforce RTX/Turing questions
During this podcast, Peterson explained several core aspects of the Turing architecture, discussing the increased parallelisation within Turing’s CUDA cores as well as all-new architectural features such Tensor cores and RT (Ray Tracing) cores, both of which are unique to Turing. Yes, Tensor cores are already available on Titan V-class graphics cards, but Turing is the first architecture to deliver the feature to consumer-grade Geforce cards.Â
When compared to their Pascal-based counterparts, Peterson stated that Turing graphics cards would offer a 35-45% performance boost over Pascal, ensuring a notable performance boost in games without support for Nvidia’s RTX Ray Tracing hardware or Tensor cores. I repeat, existing games are expected to see a 35-45% performance boost with Turing when comparing models like the GTX 1080 and its successor the RTX 2080.Â
These performance improvements come from the increased CUDA core counts on Turing graphics cards as well as Nvidia’s redesigned CUDA cores, which can now operate integer and floating point pipelines simultaneously. All of these improvements combined are what enables Turing’s 35-45% performance boost over Pascal, with Tom Peterson stating that he has overclocked Turing Founders Edition graphics cards to 2.13GHz.  Â
 Â
When discussing Nvidia’s DLSS (Deep Learning Super Sampling) technology, Peterson confirmed that the technique requires direct collaboration between Nvidia and the game developer in question.Â
To put things simply, Nvidia uses a pre-release version of a game and pumps data through their SATURN V supercomputer, data which is then used to create a bespoke DLSS algorithm that can run directly on Nvidia’s Tensor cores. One game’s implementation of DLSS doesn’t work on another, making DLSS a feature for AAA developers, or studios that can otherwise get this level of game support from Nvidia. Â
Developers are also able to utilise Nvidia’s Tensor cores to develop their own image upscaling features, though this would likely take up too much development time to be worthwhile.Â
You can join the discussion on Nvidia’s Tom Peterson and his Q&A about Turing-powered RTX graphics cards on the OC3D Forums.Â