You're right but my understanding is that Groq's LPU architecture makes it inference-only in practice. Like Groq's chips only have 230MB of SRAM per chip vs 80

You're right but my understanding is that Groq's LPU architecture makes it inference-only in practice. Like Groq's chips only have 230MB of SRAM per chip vs 80GB on an H100, training is memory hungry as you need to hold model weights + gradients + optimizer states + intermediate activations.