Nvidia GTC live: Jen-Hsun lays out ‘the future of real-time rendering’ in today’s keynote and potentially ‘it’s gonna blow your mind’
Refresh
2026-03-16T17:34:09.482Z
Given how painfully expensive memory is for everything these days, however, I kinda think Nick ought to get on and patent RTX Ultra Memory.
“I’ve got it! Huang will hop onto the stage with a new GeForce RTX graphics card. It’ll have 10,000 CUDA cores and 4 GB of VRAM, but it will be the first to use RTX Ultra Memory, an AI-powered system that neurally compresses everything automatically and does it so quickly and so well that it effectively quadruples the VRAM and its bandwidth on your graphics card. Yours for just $1,299.”
2026-03-16T17:32:58.578Z
Our Nick had some good ideas about this earlier:
“How about DLSS Ray Full Construction, an AI system that doesn’t just denoise a ray-traced scene but actually uses machine learning to generate thousands of additional ‘fake’ rays?”
It would certainly make ray/path tracing a little lighter on GPUs if they only have to actually trace a couple real rays in the scene, rather than a couple per pixel.
2026-03-16T16:22:47.028Z
(Image credit: Nvidia/CD Projekt Red)
So, what could it be about? An update and some demo-ing of neural rendering in action in an actual game would be my first guess.
Introduced at CES 2025, alongside the RTX Blackwell series of GPUs, neural rendering—and neural shaders—promise big things. From enhanced AI compression techniques to deliver “next-generation asset generation” as a potental VRAM crutch (useful at a time when memory be expensive) to RTX MEGA GEOMETRY which is going to give Witcher 4 lots of pretty trees, sticking AI into the graphics pipeline has the potential to hugely up the fidelity of PC games.
