3 ms·
Where did you get this number? A really, really big (like the biggest FAANG had last year) ML cluster would use at most four or five DGX-N’s to train a model,
by hedora 7y ago
Where did you get this number?
A really, really big (like the biggest FAANG had last year) ML cluster would use at most four or five DGX-N’s to train a model, because that is the current scaling limit for CNNs.
Each draw 3500W:
https://docs.nvidia.com/dgx/dgx1-user-guide/introduction-to-dgx1.html#topic_hwspec-power https://docs.nvidia.com/dgx/dgx1-user-guide/introduction-to-...
3500 * 5 is 17,500W. A US space heater maxes out at 1600 watts.
Training for a week (5 days) is 2100 kWh, which is less than 60 gallons of gasoline, or 2-6 tanks of gas, depending on your vehicle:
https://www.convertunits.com/from/kWh/to/gallon+%5BU.S.%5D+of+automotive+gasoline https://www.convertunits.com/from/kWh/to/gallon+%5BU.S.%5D+o...