NVIDIA CEO Jensen Huang, known as the "Godfather of AI," delivered a keynote speech at 7:00 PM on June 2nd at the National Taiwan University Gymnasium, warming up for the 2024 Taipei International Computer Show (COMPUTEX). He declared, "A new era of computing is beginning," and showcased various AI applications, including Nvidia Earth-2, Digital Human, and the Robot Factory. In his approximately two-hour speech, Huang highlighted what trends in accelerated computing are driving this trend, and how can it address the issue of increased power consumption driven by AI?
Trend 1: Accelerated computing can solve the problem of "computing power inflation"
At the outset, Huang Jen-Hsun mentioned that the computer industry has a 60-year history, starting with the launch of the IBM 360 central processing unit (CPU) in 1964. However, as computing demand has grown exponentially, the expansion rate of CPUs has slowed significantly, leading to the problem of "computation inflation." This means that data centers are using more electricity and computing costs are rising accordingly.
To address the issue of computing power inflation, NVIDIA has developed graphics processing units (GPUs) and provided CUDA programming tools, allowing developers to use GPUs to accelerate various computing tasks, including data processing and deep learning applications. Jensen Huang emphasized that GPU-accelerated computing can save significant costs and energy for data centers, achieving the goal of "buy more, save more."
Jensen Huang mentioned that "accelerated computing is sustainable computing." By combining a GPU with a CPU, it is possible to achieve up to 100 times acceleration, but power consumption only increases by 3 times, and the energy efficiency per kilowatt is 25 times that of using only the CPU.
Trend 2: As AI chips become more powerful, energy and time costs decrease.
Jensen Huang mentioned the goal of launching a new AI chip every year. The latest Blackwell is hailed as "the most powerful AI chip on the planet," and it has increased AI computing power 1,000 times in 8 years. "As GPU computing power increases, its cost also decreases." Huang explained that using Blackwell to train the large language model GPT-4 system requires approximately 1/350 of the original power consumption due to the increased computing power.
In the past, training an AI large language model (LLM) might have required 1,000GWh of electricity, but currently there isn't a single GW-level data center in the world, and the computational time would start at about a month. With the advent of the Blackwell acceleration chip, the power consumption has been reduced from 1,000GWh to 3GWh. "Only 10,000 GPUs are needed, and the computation can be completed in about 10 days," said Jensen Huang.
Trend 3: AI applications are booming, including climate technology, digital humans, and robots
Nvidia has also put the concept of accelerated computing into practical application. Its most ambitious project, Nvidia Earth-2, directly simulates a clone of the Earth, predicts climate change through AI, and helps countries and businesses cope with the impact of extreme climate change.
It is reported that the image resolution generated by Nvidia Earth-2 is 12.5 times higher than that of current numerical models, the speed is increased by 1,000 times, and the energy efficiency is increased by 3,000 times. It can better analyze, plan and simulate the impact of weather. Currently, Taiwan's Central Weather Bureau of the Ministry of Transportation and Communications has used it to predict the location of typhoon landing with more precision.
Furthermore, digital humans will transform all walks of life in the future, enabling digital nurses, customer service representatives, teachers, and more through generative AI. In other words, "computers in the future will be able to interact as smoothly as humans." For example, AI interior designers can provide decorating suggestions and procure materials and furniture.
Finally, Huang Jen-Hsun pointed out: "The next wave of artificial intelligence will be physical AI, artificial intelligence that understands the laws of physics and can work alongside humans."
Currently, many major Taiwanese electronics manufacturers, such as Foxconn, Delta Electronics, Pegatron, and Wistron, have integrated Nvidia's autonomous robotics into their factories. Using digital twin technology to plan factories and train robots, they aim to improve production efficiency and reduce costs. "In the future, all factories will be robot factories, and these robots will manufacture robot products," predicts Jensen Huang.
This article is reprinted with permission from RECCESSARY. The original title is "NVIDIA's Jensen Huang Unveils a New Era of Computing! What is Hashrate Inflation? How Do GPUs Solve the Power Consumption Problem?" by Chen Yingxuan. This article is not licensed under the Commonwealth Creative Commons License.
Sources:
Environmental Information Center https://e-info.org.tw/node/239276
