3 things to know about Nvidia ahead of Q3 earnings

• by: @Mxyzptlk30
• 3 months ago

The reason why gaming GPUs are so popular for AI is because computer graphics requires a lot of matrix math (matrices). Matrices are super easy: mostly add, subtract and multiply. And that's needed to calculate the scene (3D positioning, texture mapping, shaders, etc). And the GPUs were designed to be able to do matrices super-fast and in parallel. Nvidia wrote the first CUDA libraries to help developers with graphics on their GPUs.

Machine learning/AI also requires a ton of matrix math. During an AI class at Stanford, one of the students discovered they could use this gaming library (CUDA) to do their AI-related school work really fast. Jensen got word that students were buying up Nvidia gaming cards and using CUDA for AI and not graphics, so he wanted to help by having the CUDA team add AI-related functionality into CUDA and well… here we are. AMD had no user-related software libraries for their cards, so AI students couldn't use AMD gaming cards. AMD completely relied on Microsoft and DirectX, which didn't allow AI developers the granularity of doing the matrix math they needed. AMD created ROCm late in the game to counter CUDA and to get on the AI bandwagon, but it's far behind where CUDA is today.

In any case, this is a simplified version of events. There's a lot more nuance, but this post is getting too long.

• by: @matthew.stevick
• 3 months ago

nothing to think or worry about for lifetime longs