3 ms·
I agree. While GPUs have been indispensable for training and deploying large language models, their role in the field of computer vision is becoming increasingl
by Vox_Leone 2y ago
I agree. While GPUs have been indispensable for training and deploying large language models, their role in the field of computer vision is becoming increasingly critical. The expanding applications, increasing complexity of models, real-time processing needs, growing dataset sizes, and other factors, seem to indicate that the demand for GPUs is likely to rise in the mid term.
- CharlieDigital 2y agoI think if you take a device like the Apple Vision Pro or Meta Quest: imagine that you can procedurally generate any experience like a holodeck from just a few instructions and fine tuning the generated world. We're not even close. Imagine YouTube except you can fully generate shorts. Realistic, 2D animated, 3D, whatever your imagination desires. Imagine how that changes storytelling and content creation. More GPUs, please.
- dghlsakjg 2y agoI'm not super up on the latest gaming, but I imagine that procedurally generated in-game characters that you can have a voice conversation with are just around the corner. The ability to have a conversation with a character with a back story is definitely going to be an interesting addition in the bery near future.
- CharlieDigital 2y agoGoogle DeepMind has a fully procedurally generated Doom: https://gamengen.github.io/ https://gamengen.github.io/
- jononor 2y agoDriven by companies like Facebook and Meta, optimizing purely for engagement and ad viewing time and spending on content+products. The Matrix is ever closer. Neuralink already prepping gen 2.
- ActorNightly 2y agoThe interesting thing is that at some point, its very possible that LLM (or rather just transformer based large models) are going to be used to create ASIC or at least used to program much cheaper FPGAs for the specific applications, so in fact, GPU use will still decline.
- CharlieDigital 2y agoI feel like this is unlikely because it's not like a crypto algorithm where there's really only one way to compute the output. There are various optimizations, algorithms, and approaches and having more general purpose hardware yields flexibility.
- ActorNightly 2y ago>crypto algorithm where there's really only one way to compute the output. With ML, its basically just matrix math. You can easily build ASIC if you know your Tensor sizes for inference, because the memory locations become static, which means your circuitry gets way simpler.
- yencabulator 2y agoSome time after an LLM programs its own successor, without bugs. I won't be holding my breath.