Nvidia Overclocked an RTX 4090 Beyond 3.0 GHz

Nvidia GeForce RTX 4090 Founders Edition
(Image credit: Nvidia)

One of the things that caught our attention during Nvidia's GeForce RTX 40-series announcement was the company's claim that it has already run its AD102 GPU at 3.0 GHz in its labs. This suggests the new GPU could be an overclocking champ with plenty of headroom that AIB partners could tap on factory-overclocked boards, as well as manual overclocking. The RTX 4090 is set to launch on October 12 and will likely join the best graphics cards — at least for those with deep pockets.

Nvidia's AD102 chip packs 76 billion transistors and is made using TSMC's 4N fabrication technology. Running a GPU this complex at 3.0 GHz is a big deal, as typically chips that run at high clocks are considerably less complex. Meanwhile, Nvidia's GeForce RTX 4090 has a boost clock of 2520 MHz, so overclocking an AD102 GPU to ~3.0 GHz shouldn't be too hard for a good air or closed loop liquid cooler. Naturally, liquid nitrogen should enable the GPU to be pushed even further.

The TSMC 4N fabrication technology that Nvidia Ada Lovelace GPUs use has been customized for the company, possibly to help with yields, process window, power, and clock speeds. Nvidia does not disclose the peculiarities of TSMC's 4N or how it customized the company's 'default' N4/4P node to achieve its goals in terms of complexity, performance, and power. Considering the clock speed potential of Nvidia's AD102, Nvidia and TSMC likely optimized the production node both for complexity and frequency. However, we do not know whether TSMC implemented libraries set to be used for upcoming N4P or N4X fabrication process technologies.

With its GeForce RTX 30-series graphics cards, Nvidia made thick three-slot cooling systems relatively common on custom graphics card, but with its GeForce RTX 40-series boards it looks like the company will make such ultra-large coolers truly widespread. Based on the custom GeForce RTX 40-series boards we have seen, they will all feature sophisticated cooling systems that will probably enhance their overclocking potential. How many of these boards will be able to hit 3 GHz remains to be seen, but it's certainly a possibility for at least some of the models.

One of the things we don't know with AD102-based graphics boards is how much power a 3 GHz GPU frequency will require. Over the past few months, we've heard rumors of flagship products with 600W or even higher power envelopes, but the reference GeForce RTX 4090 is only rated for up to 450W thermal graphics power. Some rumors suggest that Nvidia's partners might enable an up to 600W power envelope for their premium offerings. If this is true, we may soon see some heavily overclocked GeForce RTX 4090 boards.

Hitting a 3+ GHz frequency for a GPU is hardly a breakthrough, as AMD's Navi 21 XTXH was overclocked to 3.20 GHz in mid-2021 and the world's first GPU to break the 3 GHz barrier was an Nvidia Maxwell GPU, circa 2015. But AMD's Navi 21 is a chip that packs 26.8 billion transistors, made using TSMC's N7 technology (which was presumably optimized for AMD's needs), whereas Nvidia's AD102 is almost three times more complex. And Maxwell required liquid nitrogen to hit that speed, while it seems like Ada may reach that on air cooling.

We're definitely interested in seeing just how far custom RTX 4090 cards push clock speeds, and how much power they require to do so. We should have full reviews up on or slightly before the October 12 launch date.

Anton Shilov
Contributing Writer

Anton Shilov is a contributing writer at Tom’s Hardware. Over the past couple of decades, he has covered everything from CPUs and GPUs to supercomputers and from modern process technologies and latest fab tools to high-tech industry trends.

  • -Fran-
    Given how this thing is getting priced like Hardware from the 80's, maybe it did need them whole 1.21GW!

  • The Historical Fidelity
    Inside sources from AMD are saying they are seeing around 4Ghz on their RDNA3 engineering samples. I guess one of the benefits of using simpler chiplets to create high complexity architectures
  • thisisaname
    I wonder how Golden the chip was.
  • The Historical Fidelity
    thisisaname said:
    I wonder how Golden the chip was.
    Given that nvidia cherry picks only the best benchmarks and uses obscure and misleading settings to maximize gen on gen performance uplift. I would say it was an unobtanium chip