Quote from: OGeneral on January 08, 2022, 16:24:21
450W is insane and a clear indication they want to make sure these cards are out of the compute/ml training market because you can not put more than one or two in one chassis power and thermals are simply too bad.
The 450W power limit is already typical for many AIB manufacturer 3090 cards. MSI, Asus, and EVGA all have RTX 3090 cards with power limits between 450 and 520 watts. I'm sure we will see 600 watt power limits on the 3090 ti cards from the various AIB partners.
You can simply use software in Windows or Linux to limit these cards and the 3090 ti to a lower power limit. The lower power limit won't have a huge impact in AI/ML performance or most 3D rendering performance. In reality, the current 3090 cards with 450-520 watt power limits rarely hit 400 watts power draw during 3D rendering, and AI/ML applications.
Also, you can safely put three 450 watt cards in one system and have them be cooled adequately. If you are using them in an industrial/commercial environment or at a home with a 20A breaker, you can even use 4 cards if you adjust the power limit. Increasing the power limit on RTX 200 and RTX 3000 cards shows diminishing returns in 3D Rendering/AI/ML applications.
If you are that concerned about cooling 3 or 4 of these 3090 ti cards in one system, there are many water cooling available parts available including professional/commercial water cooling solutions. The new 3090 ti compatible components should be available within a month or two of release of these cards. I currently have three RTX 3090s with a BIOS that has a 1,000 watt power limit per card(yes I said 1,000 watts per card) in one water cooled system that uses two power supplies. The cards don't draw 1,000 watts each all the time, but can sustain 600-700 watts on all cards while rarely going above 55C GPU temperatures due to water cooling of both sides of each card.
In reality, Nvidia raised the power limit for gaming use, not to prevent people from using them for 3D Rendering/AI/ML applications in place of quadro cards. There is no benefit to a higher power limit in AI/ML applications with RTX 3000 cards.