From MIPS to EXAFLOPS in decades: Computing power is explosive and will transform AI.

At the recent NVIDIA GTC conference, the company unveiled the first single rack system that can use one Exaflop (1 billion or Quintillion, Float Point Operations). This innovation is based on the latest GB200 NVL72 system that integrates NVIDIA’s latest Blackwell graphics processing device (GPU). The height of the standard computer rack is about 6 feet, 3 feet depth, and less than 2 feet wide.

Exaflop contraction: From Frontier to Blackwell

Some of the presentations struck me. First, the world’s first Exaflopable computer was installed just a few years ago in 2022 Oak Ridge National Laboratory. The “Frontier” supercomputer, which is produced by HPE and run by AMD GPU and CPU for comparison, consists of 74 racks. The new NVIDIA system has achieved about 73 times the performance density in just three years and is three times the performance every year. This development reflects the amazing progress of computing density, energy efficiency and architectural design.

Second, both systems have reached exacale milestones, but one should be said to be built for other challenges optimized for speed, and the other is built for precision. NVIDIA’s Exaflop specifications are based on low precision mathematics, especially 4 -bit and 8 -bit ploting points. It is optimal for AI workloads, including tasks such as education and LLMS (Lange Language Model) training. Such calculations give priority over precision. In contrast, Frontier’s Exaflop rating was achieved using the 64 -bit double presence mathematics, which is a gold standard for scientific simulation, where accuracy is important.

We have been a long way (very fast)

This level of progress seems to be almost incredible, especially when you remember the cutting edge when you start your career in the computing industry. My first professional task was the programmer of the DEC KL 1090, which provided 1.8 million (MIPS) per second, which is part of the DEC’s PDP-10-time main frame. In addition to CPU performance, the machine connected to the Catode Ray Tube (CRT) is displayed through a hard cable. There was no graphic function and only light text on a dark background. Of course there is no internet. The remote user was connected to the telephone line using a modem running at a speed of up to 1,200 bits per second.

50 billion times more computing

Comparing MIP with Flops, it provides a common progressive sense, but it is important to remember that these metrics measure other computing workloads. MIPS reflects integer processing speed, especially for general purpose computing in business applications. FLOPS measures many numbers behind modern AIs, such as Matrix Math and Linear Algebra, used to measure the important floating point performance for scientific workloads and train and execute the MA (Machine Learning) model.

It’s not a direct comparison, but the strict size of the difference between MIPS and Flops now provides a powerful picture of rapid growth of computing performance. To measure the work that did this, the new NVIDIA system is about 500 billion times more powerful than the DEC machine. This kind of leap shows the exponential growth of computing power on a single professional career and questions, what can the next five people will bring if this progress is possible within 40 years?

NVIDIA provided some clues for that. In the GTC, the company will provide 14 times the performance of the next-generation full lock system based on “Vera Rubin” Ultra Architecture, which will provide 14 times the performance of the Blackwell Ultra Rack SHIPPING this year and reach 14 to 15 exaflops in AI-AptialICED work for next year or 2 years. I shared a roadmap that predicts it.

What is noteworthy is efficiency. Achieving this level of performance in a single rack means that the absolute power demand of these systems is enormous, with less physical space per work unit, less materials, and less energy use per operation.

Does AI actually require computing power?

These performance benefits are actually impressive, but the AI ​​industry is now full of fundamental questions. Competition for building a large AI data center is led by more competent AI models and more competent AI models.

The most ambitious effort is the $ 500 billion project Stargate, and we plan 20 data centers across 200,000 square feet across the United States. Since the company and the state are sending to guarantee the infrastructure to support the AI ​​workload of tomorrow, the waves of other super sniper projects are underway at the world’s planning stage.

Some analysts are now worried that we will build excessive AI data center capacity. The problem has deepened since the launch of China’s deep sea reasoning model R1, which requires much less calculations than many colleagues. Microsoft later canceled the lease as a multiple data center provider and speculated that it was re -correcting expectations for the demand for AI infrastructure in the future.

but, register proposal This fullback can be more related to the planned AI data center, which is not enough to support the power and cooling demands of the next generation AI system. Already, the AI ​​model is currently promoting the limitations of what infrastructure can support. MIT technology review Reported This may be the reason why many data centers in China have difficulty and fail, and may have been built as specifications that are not suitable for the next few years.

AI reasoning requires more flops

The reasoning model performs most of the runtime through the process known as reasoning. This model allows some of today’s most progressive and resource -intensive applications, including the new wave of in -depth research assistants and agent AI systems.

DeepSeek-R1 initially made the industry to think that the future could be required by the future. Less Computing Power, NVIDIA CEO Jensen Huang was strongly pushed. speaking To CNBC he responded to this perception. He added that AI reasoning AI consumes 100 times more computing than non -rational AI.

As the AI ​​continues to develop autonomous agents in the reasoning model, the demand for computing will increase once again. The following breakthroughs can occur not only in language or vision, but also in AI agent adjustments, fusion simulations, or large digital twins.

Openai announced $ 40 billion in new funds, the largest private technology funding that seems to fit the queue. The company is A Blog The funds say, “We can further develop the border of AI research, expand the computing infrastructure and provide more powerful tools for 500 million people who use CHATGPT every week.”

Why are many capital flowing to AI? The reason ranges from competition to national security. As mcKinsey illustrates, one specific element is noticeable title: “AI can increase corporate profits of $ 4.4 trillion annually.”

What is coming next? Anyone guesses

At the core, the information system relates to an emergency vehicle routing system written on a modern AI system that accelerates a student’s accomplishment in COBOL through an emergency vehicle routing system, and accelerates tools or drug discovery. The goal was always the same. To better understand the world.

Now, as a powerful AI begins to appear, we go beyond thresholds. For the first time, we can have computing performance and intelligence to solve the problem that is once beyond human scope.

New York Times columnist Kevin Rouge I recently captured this moment.: “Every week, I meet with engineers and entrepreneurs working in AI and change, the world shaking, and the types of changes we have never seen before.” And it does not calculate the breakthrough that arrives every week.

I’ve seen the GPT-4O of Openai for the last few days. Almost perfect image In the text, it can be the most advanced reasoning model in the Google Release, but we have released a video model with a shot-shot character and a scene consistent on the Gemini 2.5 PRO and runway.

The next thing to come is truly guessing. We do not know whether powerful AI will help to solve fusion energy or to get new biological risks. But if more flops come online over the next five years, there is one certain thing. Innovation will come quickly and hard. It is also clear that it is necessary to talk about responsibility, regulation and restraint as a flops scale.

Gary Grossman is an EVP of technical practice Edelmann Global lead of Edelman AI Center of Excellence.