The point is addressing the cost of training the DLSS5 model.

The ‘models trained in giant data centers that devour energy and water’ are LLMs. DLSS5 isn’t an LLM, nor does it have anywhere near the training costs. DLSS5’s training costs are a rounding error when compared to LLMs.

However, DLSS 5 is not just upscaling anymore, it’s fully analyzing and generatively replacing parts of the image (lighting, material properties, …)

I know what Neural Rendering is. Calling it upscaling or generative upscaling doesn’t change my point about training.

And image based gen AI definitely cannot be trained on a phone, even less so than LLMs.

I know some CS professors that would be surprised to learn this.

I’m looking at the result of a lab where I completed training on 4x4 MLP denoiser on a TI-Nspire CX II… a graphing calculator with 64MB RAM and a 369MHZ processor. About 300k FLOPS/step, takes about an hour.

Yes, training a larger model on a phone would take longer but as long as a reasonable sized chunk of the model could fit into RAM it can be trained, eventually.

They’re not trained on phones as a rule because you can throw a lot more compute at them with dedicated hardware. You can train an LLM on an abacus if you’re patient enough.