Non-ECC is fine for your RTX 4090 rig.

Don’t bother paying extra for ECC DDR5. For a single 4090 or even a few, consumer non-ECC RAM works perfectly well for deep learning.

I get the appeal — ECC was designed to catch bit flips that corrupt long training runs. But on a consumer GPU rig, the chance of a silent memory error crashing your 8‑hour training is basically zero. The GPU does its own error correction on VRAM, and the CPU side isn’t running 24/7 high‑stress memory operations like a server.

ECC DDR5 costs more, runs slightly slower, and gives you no real benefit for most ML workloads. The only case I’d consider it is if you’re building a multi‑socket, multi‑GPU beast that runs training 24/7 on critical models. That’s datacenter territory, not a desk rig.

Save the money for more VRAM or a faster SSD.