4090 - ECC ON vs ECC OFF

Maxim Saplin - Jun 25 - - Dev Community

Fine Tuning LLM via Huggin Face TRL/Torch:

  • ECC On: 2,22 epochs/day
  • ECC Off: 2,33 epochs/day [+5%]

Weights & Biases Dashboard

3D Mark TimeSpy Graphics Score:

  • ECC On: 36400
  • ECC Off: 37000 [+1,6%]

I have noticed the "Change ECC State" at Nvidia Control Panel and decided to check how enabling and disabling memory error corrections affected the performance of GPU.

Nvidia Control Panel

RTX 4090 is overclocked (VRAM +1440MHZ, Core +130Mhz at the bottom and +200Mhz at the top end of the voltage curve) and capped at 84% power (380W max).

. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .