Hacker News Reader

3080 ($700) apparently has 238 teraflops of "tensor compute." We're frighteningly close to a petaflop of compute for less than a thousand usd.

That is FP16 (rounding up to FP32). It is not sufficient for most HPC compute. Good for ML/AI, at least.

For all but the AI/DL crowd the move away from high precision to high-power low precision is a bit sad.