Friday, January 17th 2025
NVIDIA Reveals Secret Weapon Behind DLSS Evolution: Dedicated Supercomputer Running for Six Years
At the RTX "Blackwell" Editor's Day during CES 2025, NVIDIA pulled back the curtain on one of its most powerful tools: a dedicated supercomputer that has been continuously improving DLSS (Deep Learning Super Sampling) for the past six years. Brian Catanzaro, NVIDIA's VP of applied deep learning research, disclosed that thousands of the company's latest GPUs have been working round-the-clock, analyzing and perfecting the technology that has revolutionized gaming graphics. "We have a big supercomputer at NVIDIA that is running 24/7, 365 days a year improving DLSS," Catanzaro explained during his presentation on DLSS 4. The supercomputer's primary task involves analyzing failures in DLSS performance, such as ghosting, flickering, or blurriness across hundreds of games. When issues are identified, the system augments its training data sets with new examples of optimal graphics and challenging scenarios that DLSS needs to address.
DLSS 4 is the first move from convolutional neural networks to a transformer model that runs locally on client PCs. The continuous learning process has been crucial in refining the technology, with the dedicated supercomputer serving as the backbone of this evolution. The scale of resources allocated to DLSS development is massive, as the entire pipeline for a self-improving DLSS model must consist of not only thousands but tens of thousands of GPUs. Of course, a company making 100,000 GPU data centers (xAI's Colossus) must save some for itself and is proactively using it to improve its software stack. NVIDIA's CEO Jensen Huang famously said that DLSS can predict the future. Of course, these statements are to be tested when the Blackwell series launches. However, the approach of using massive data centers to improve DLSS is quite interesting, and with each new GPU generation NVIDIA release, the process is getting significantly sped up.
Source:
via PC Gamer
DLSS 4 is the first move from convolutional neural networks to a transformer model that runs locally on client PCs. The continuous learning process has been crucial in refining the technology, with the dedicated supercomputer serving as the backbone of this evolution. The scale of resources allocated to DLSS development is massive, as the entire pipeline for a self-improving DLSS model must consist of not only thousands but tens of thousands of GPUs. Of course, a company making 100,000 GPU data centers (xAI's Colossus) must save some for itself and is proactively using it to improve its software stack. NVIDIA's CEO Jensen Huang famously said that DLSS can predict the future. Of course, these statements are to be tested when the Blackwell series launches. However, the approach of using massive data centers to improve DLSS is quite interesting, and with each new GPU generation NVIDIA release, the process is getting significantly sped up.


62 Comments on NVIDIA Reveals Secret Weapon Behind DLSS Evolution: Dedicated Supercomputer Running for Six Years
SSAA was simply increasing the internal rendering resolution (NV DSR and AMD VSR were a way to universally apply this without support in the game), MSAA has always been the clearest in my opinion but it's very resource-inefficient and has many compatibility issues with modern rendering techniques (iirc it doesn't play well with deferred rendering).
With nuclear fusion, energy could become much cheaper in the near future and allow exponential growth of civilization (technology and number of people)
chipsandcheese.com/p/amd-rdna-3-5s-llvm-changes
With the extra hardware bump, it should be able to run an upscaling CNN without much issues and no need for extra hardware (apart from what's in the GPU itself).Sony's PSSR did do something similar to what Nvidia has done with DLSS, by training a model with tons of compute during a long period of time, which can then be used by the actual consoles, they just did not announce it like Nvidia did now. And if Nvidia never gave up this detail away, you wouldn't be making this complaint.
FSR4, if truly based on ML, will also require lots of compute time beforehand in order to create a model that can perform this task in your local GPU.
Let me try to give you a better example: do you know that feature in phone's gallery that are able to recognize people or places?
That's a machine learning model that's running in your phone and tagging those images behind the scenes.
That "model" (think of a "model" as a binary or dll that contains the "runtime" of the AI stuff) has been trained by google/samsung/apple in their servers for long hours with tons of examples saying "this picture is a dog", "this is a car", "this is a beach", "this person X is different from person Y", etc etc. This part is the "training" part, which is really compute intensive and takes really long time. As an example, the GPT model behind ChatGPT took around 5~6 months to train.
The outcome of this model is then shipped into your phone, where it's able to use what it has learnt and apply it to your cat pictures, and say that it is a cat. This part is called the "inference" part, and is often really fast. Think how DLSS, even in its first version, was able to upscale a frame from a smaller res into a higher one with really fast FPS (so for each frame, it upscaled it in less than 10ms!). In a similar manner, think how your phone is able to tag a pic as a "dog" really quick, or how ChatGPT is able to give you answers reasonably fast, even though the training part for all of those tasks took weeks, months, or even years.
Aren't you thinking about AGI?
I think of AI as a general term. There are terms after AI that can distinguish it into multiple levels, from sub-human, on par, up to beyond.
All some form of AI though.
To me the idea of intelligence is not equal to something necessarily high.
Is not correct to say... this has low/very low intelligence? I think it is.
And when it is made it is artificial.