The flagship graphics playing cards of latest generations have been trending upwards in most energy consumption. The RTX 4090 can pull greater than 450W, the RTX 3090 Ti might do the identical, whereas AMD’s RX 7900 XTX is rated for 355W. Manufacturing unit overclocked fashions are sometimes a bit larger than that. However these are rookie numbers. It has been reported that Nvidia’s subsequent technology GB200 AI accelerator can pull as much as 1,000W.
The supply of the 1,000W declare is an announcement made by Dell COO Jeffrey Clarke throughout Dell’s This fall 2024 earnings name (by way of The Register). He is as prone to be within the know as anybody, so this 1,000W declare isn’t any mere rumor.
Clarke’s precise quote is “Our characterization within the thermal aspect, you actually do not want direct liquid cooling to get to the power density of 1,000 watts per GPU. That occurs subsequent 12 months with the B200.”
Be aware that he refers back to the B200, however there’s truly no B200 on Nvidia’s roadmap. Its subsequent technology flagship enterprise GPU is known as the B100. It is primarily based on the Blackwell structure. Nonetheless, there’s a GB200, which mixes the B100 GPU with the Grace CPU. It is probably Clarke merely misspoke and is referring to the latter.
Datacenter GPUs are very totally different beasts to shopper gaming GPUs, the place issues like fan speeds, bodily measurement and energy consumption are secondary concerns in comparison with outright efficiency. The GB200 is all about efficiency. The totally different variants of Blackwell primarily based accelerators will ultimately type the spine of cloud AI methods in 2025 and past.
By now, everyone knows how AI is large enterprise. Nvidia is now the world’s third largest firm by market cap, lately overtaking Saudi primarily based petro big Aramco, leaving solely Apple and Microsoft forward of it. And it is catching them rapidly, indicating demand for prime efficiency AI accelerators stays insatiable.
I am unable to think about we’ll ever see a 1,000W desktop graphics card. That is actually a step too far for a typical ATX PC, however with 450W changing into a brand new baseline for prime finish gaming GPUs, I’d count on issues to take a step up once more. The 500W barrier has just about come and gone already, and if Nvidia chooses to launch an RTX 4090 Ti, I might think about 600W or extra changing into a actuality. Simply take a look at the leaked photos of a gargantuan quad slot Founders Version cooler.
Personally, I detest the thought of playing cards utilizing that a lot energy. I take advantage of a 450W RTX 4090 recurrently, and it actually does warmth up my workplace room throughout lengthy periods. I hate to assume what a 600W+ card would do in the course of the hotter months of the 12 months.
It doesn’t matter what remaining type the long run RTX 5090 takes, it’s going to be a sizzling and energy thirsty card. If excessive finish Blackwell primarily based GPUs have TDPs above 500W, we will count on quad slot or AIO cooled playing cards to turn into the norm, significantly overclocked fashions.
Hopefully playing cards just like the RTX 5060 and RTX 5070 preserve the superb energy effectivity and efficiency per watt traits of playing cards just like the RTX 4060 and lately launched RTX 4070 Tremendous.