As a former Groq engineer, the LPU branding is not really technically accurate. These are general purpose compute chips with excellent software for transformer models. In the sense that the GroqChip is general-purpose, it is basically a highly parallel, high performance, deterministic execution engine. It 'rivals' GPUs in the sense that obviously both are used for AI execution. The whitepapers cited are basically a perfect model for how the chips operate. As you can see, they are general purpose tensor units.
That being said, yes the software is impressive and the engineering team at Groq is top-notch.
Unfortunately, the chips are mainly aimed at inference, and it seems like a lot of the large investments at the moment are being driven by training.
EDIT: In the spirit of full disclosure, I suppose I should point out that I own lots of Groq shares, so have every interest in their success.
That being said, yes the software is impressive and the engineering team at Groq is top-notch.
Unfortunately, the chips are mainly aimed at inference, and it seems like a lot of the large investments at the moment are being driven by training.
EDIT: In the spirit of full disclosure, I suppose I should point out that I own lots of Groq shares, so have every interest in their success.