Dell Technologies is rumored to release the NVIDIA B200 next year, with a power consumption of up to 1000W. The ultimate benefit for AI is: energy?

Author: Li Xiaoyin

Source: Hard AI

After the "chip shortage," will the artificial intelligence market also face an "electricity shortage"?

Benefiting from the surge in demand for AI servers, Dell Tech, one of the world's largest server manufacturers, achieved revenue that exceeded expectations in its latest financial quarter, with its stock price more than doubling in the past 12 months.

Following the release of the financial report, Jeff Clarke, the Chief Operating Officer of Dell Tech, revealed in a press release that NVIDIA is set to launch the B200 product with the "Blackwell" architecture in 2025, with power consumption potentially reaching 1000W.

Clarke also mentioned that Dell Tech's flagship product, the PowerEdge XE9680 rack server, powered by NVIDIA, is the company's "fastest ever" solution.

B200 Power Consumption Increases by Over 40% Compared to H100

Currently, NVIDIA has not disclosed detailed information about the Blackwell architecture. From the perspective of chip manufacturing and basic heat dissipation principles (the maximum heat dissipation per mm² of chip area is 1W), considering:

The power consumption of NVIDIA's H100 (built on custom 4nm process technology) is about 700W (including HBM memory power), and considering the chip's die size of 814mm², the actual power consumption per square millimeter is less than 1W. This means that the power consumption of B200 will increase by over 40% compared to H100.

Some media outlets have analyzed that the H200 is likely to be built on another performance-enhancing process technology, such as using 3nm process technology.

Considering the power consumption of the chip and the required heat dissipation, B100 may become the company's first GPU produced with a dual-chip design, allowing for a larger surface area for heat dissipation. It is reported that AMD and Intel have adopted GPU architectures with multi-chip designs, which may become an industry trend.

In addition to the energy consumption requirements for chip design, when discussing AI and high-performance computing (HPC) applications, it is also necessary to consider how to balance the high power consumption required for these FLOPS and the heat energy released simultaneously.

FLOPS (floating-point operations per second) refer to the number of floating-point operations per second and are generally used to measure hardware performance.

For software developers, the key is how to efficiently utilize these FLOPS; for hardware developers, the focus is on how to cool the processors that generate these FLOPS effectively. And this is exactly where the advantage of Dell Tech Blackwell processors lies.

Clarke stated:

"(NVIDIA's next-generation AI and HPC GPU) will be implemented on B200 next year."

"We will have the opportunity to showcase our engineering technology and our speed of action, as well as the work we have done as industry leaders, using our professional skills to achieve liquid-cooled scalable performance, whether it's work in fluid chemistry and performance, or our interconnection work, telemetry work, and power management work. This indeed prepares us to push it to the market on a large scale to leverage the incredible computing power, intensity, or capability that will exist in the market."

B200 did not appear in NVIDIA's technical roadmap released last October. Currently, NVIDIA has not yet disclosed detailed information about B100, but may release relevant details at the developer conference to be held later this month.

The ultimate benefit of AI - Energy?

With the development of artificial intelligence technology, the current market demand for chips has surged, but this will be followed by a surge in electricity demand.

From an industry perspective, the booming artificial intelligence field has almost reshaped the already hot data center market. According to related data, ten years ago, the global data center market consumed 10 billion watts of electricity, and now the level of 100 billion watts is very common.

Although artificial intelligence currently accounts for only a small part of the global data center market. But according to the forecast of the U.S. Uptime Institute, by 2025, the proportion of global data center electricity consumption accounted for by artificial intelligence business will surge from 2% to 10%.

Some strategists believe that the development of AI technology is favorable for energy stocks:

"More and more people are beginning to realize that large artificial intelligence server clusters will require a large amount of energy, which is increasing the interest of some investors, expanding their investment scope to include energy-related fields such as electricity, oil, and gas, and nuclear energy is also receiving attention."

Musk has previously expressed concerns about the future of energy supply. At the end of last year, he stated in a podcast that the United States is currently facing a shortage of chips, which will lead to a shortage of transformers in a year, and a shortage of electricity in about two years.

According to reports, the current demand for transformers in the United States is mainly supplemented by imports. As the grid continues to expand towards a cleaner power system, the demand for transformers will surge. If further action is not taken, by 2030, the United States will face an insurmountable domestic supply gap.