To be fair, ram really is cyclical, and a commodity.
The Nvidia ecosystem… not so much. Most all of these pieces fail to recognise their already dominant position in the edge inference space - and most all of these pieces fall into the trap of AI=LLMs=ChatBots, and really, who wants to chat with a computer any more.
I don’t mind. This has been a great week to stock up on the equity, and there was enough blood in the water before the salacious spy sex story that I bought a bunch of very, very cheap $150 puts, as hysterical overreaction is the plat du jour, which I sold on Thursday.
CUDA is 20 years old. They weren’t caught by surprise. The groundwork was laid long ago.
Much in the same way as they have already laid the groundwork for inference and low cost/power edge inference, with jetson nano and the associated ecosystem, as they quite clearly understand the demand for GPUs will not be infinite, but the applications of the outputs near as damnit will be, in everything from robotics to cameras.
First you capture the developers. Then you build the technology out. Then you own the market.
You people are so naive, you probably think Jensen Huang was just practicing his Chinese with Mrs Huang...
Bryan Catanzaro, NVIDIA vice president of applied deep learning research, said when AlexNet emerged: "There really wasn't anybody at Nvidia working on AI."
This was in 2012, six years after the introduction of CUDA! Please...
> said when AlexNet emerged: "There really wasn't anybody at Nvidia working on AI."
Isn't that taking the wrong message away though? AlexNet was the deep learning setup that showed it could be done. The fact Nvidia had the tools to make that possible, even though they weren't already AI-focused, is a good thing is it not? They were ahead of the game without even really knowing what it was good for, and someone found the use case for it
I keep saying they had the infra and the good software and got lucky, to have the tools, as you are saying...and keep getting downvoted for it.
Its like they want the story of NVIDIA to be what never was. Just luck. Yes the merit to have the Technology, and the execution...but they had to make it to Vegas.
But that's too far in the other direction. They didn't have a use case for it, but they still built all the tech and it was still usable for when someone applied it to AI. They deserve credit for being forward-thinking enough on GPGPU stuff when most of their market then was video games
CUDA itself is just for GPU compute in general. The CUDA-based cuDNN library is where the neural net support is - the forward and backward pass kernels for convolution, activation functions, etc, etc (nowadays including LLM support).
When Alex Krizhevsky wrote AlexNet (the 2012 ImageNet entry that kicked off the deep learning revolution), he used CUDA (since cuDNN did not exist) - writing all his kernels by hand.
reply