Core concept
4090 FP16-per-dollar: 100 TFLOPS/$1000. A100: 25 TFLOPS/$1000. 4x more per dollar.
Advertisement
How it works
A100 has ECC, NVLink, higher memory. Multi-GPU + long training benefits.
Advertisement
Trade-offs + gotchas
Small workloads + inference: 4090. Serious ML: A100+.