kopite7kimi, a veteran leaker who described Ampere weeks before anyone else, has updated his predictions of Ada Lovelace''s power limits. It''s important to note that power limits aren''t the same as TDPs or TBPs, but they''re an upper bound that comes into play when overclocking is involved. However, they''re still a good guideline for power consumption, just an overestimate.
The flagship AD102 has a limit of 800W. Going down a level, the desktop AD103 has a limit of 450W, the AD104 of 400W, and the AD106 of 260W. In the past, the xx106 has been used by entry-level GPUs, and the xx104 has been used by mid-range GPUs, but Nvidia might change that up.
The laptop power budgets in the UK are slim. Kopite reports that the AD103 and AD104 phones have a limit of 175W, and the AD106 has a limit of 140W. However, since they''re laptop components, manufacturers will have a lot of time to tweak them anyway.
Via Fritzchen Fritz, a disassembled Nvidia P100.
The numbers are correct, but you''re likely to be losing 800W. It''s possible with a pair of the new PCIe 5.0 power connectors, which can handle 600W each (if the card used 8-pin power connectors, it would require six!) But there aren''t quite many power supplies with two of these ports and that much power on the market.
Cooling will be a great challenge. For a moment, is there a 800W ceiling that would be unreasonable for a GPU with 18,432 CUDA cores (according to the information from February''s hack)?
The AD102 will consume 0.0434W per CUDA core in its fully unlocked state, which is only marginally greater than the 0.0419W per core consumed by the RTX 3090 Ti. This GPU is a power-hungry monster, but theoretically the AD102 is only equally as horribly.
Remember, 800W is not the TDP, but the (probable) hard limit. A previous study claimed that Nvidia was putting a 600W TDP for the 4090, which would improve the GPU''s performance and leave 200W as excess headroom.
