Interview with Google's Quantum AI Team: Why is Willow a milestone breakthrough?

 Author|Zhou Ya

When it comes to quantum computing, the American theoretical physicist Richard Feynman once said something memorable: "Nature is not classical, and if you want to simulate nature, you'd better turn it into quantum mechanics." ”

Over the past 30 years, quantum computing has had a fundamental challenge: as the number of qubits increases, the error rate rises dramatically. However, this seemingly insurmountable gap has now finally dawned on a breakthrough.

On November 23, 2019, Google published a quantum computing breakthrough in the journal Nature, Google's superconducting quantum chip Sycamore completed a calculation in only 200 seconds, and the world's fastest supercomputer takes 10,000 years. Sundar Pichai, Google's CEO at the time, likened it to "the Wright brothers' first 12-second flight."

On December 10, 2024, Google published the research results of the latest quantum chip Willow in Nature, which once again made a milestone breakthrough, mainly in two aspects:

· First, Willow achieved the goal of exponentially reducing the error rate as the number of qubits increased, cutting the error rate in half each time by gradually scaling the size of the qubit array from 3x3 to 5x5 to 7x7. This has been a formidable challenge for the field since 1995, when quantum error correction was introduced by Peter Shor.

Secondly, it is more concerned about the breakthrough of its computing power. In the Random Circuit Sampling (RCS) benchmark, Willow completed a calculation in less than 5 minutes, while Frontier, the fastest supercomputer available today, would have taken 10^25 years to complete, or 10,000,000,000,000,000,000,000,000,000 years. In order to give everyone an intuitive understanding of the concept of this number, Google says that "this time is older than the age of the universe".

"When we founded the Google Quantum AI team in 2012, our vision was to build a useful large-scale quantum computer that uses quantum mechanics (nature's 'operating system') as we know it today to drive scientific discovery, develop beneficial applications, and address some of society's key challenges. ”

Google's quantum AI team is led by Hartmut Neven, and Google has built a dedicated quantum chip manufacturing facility in Santa Barbara to invest deeply in the field.

Interestingly, Julian Kelly, director of quantum hardware at Google, explained at the briefing that Google's quantum chip, Sycamore, was previously built in a shared cleanroom at the University of California, Santa Barbara, which was announced in 2013 to give Google researchers more tools and more powerful features. This time, Willow is produced at Google's own dedicated superconducting chip manufacturing facility, which allows for better control over manufacturing process parameters and improved yield and consistency.

"You can think of Willow as basically inheriting all the benefits of Sycamore, but achieving a bigger milestone breakthrough." Julian Kelly said.

Exponential quantum error correction: below the threshold

Qubits are the units of computing of quantum computers, but they are very "unstable" and tend to lose information due to the surrounding environment, and usually, the more qubits used, the more errors there will be - so "errors" are one of the biggest challenges in quantum computing.

But Google did the opposite this time: when more qubits were used in Google's quantum chip Willow, errors were drastically reduced, and Google tested larger and larger arrays of physical qubits, from 3x3 encoded qubit grids, to 5x5 grids, to 7x7 grids — and the error rate was halved each time it was enlarged. In other words, Google has achieved an exponential reduction in error rates.

Here's a little explanation. Quantum error correction involves putting many physical qubits together and making them work together, that is, correcting errors by creating a "logical qubit" (logical qubit), a combination of 3×3, 5×5, 7×7 called "logical qubits".

One central physical qubit stores the actual quantum information (data bits), and the surrounding 8 physical qubits are auxiliary bits (also called synchronous bits or anchor bits), so that a 3×3 arrangement can actually only store 1 bit of information, but it can protect this information from being destroyed by environmental interference.

It's like transporting a fragile item (quantum information), the center is the fragile item itself (data bits), and the surrounding 8 positions are packaging foam (auxiliary bits), although it seems that 9 spatial positions are used, but the actual transportation of effective items is only the center one, but these "packaging bubbles" make transportation more safe and reliable.

This explains why quantum computers need so many "physical qubits", and the number of physical qubits seems to be a lot, but the number of "logical qubits" that can actually be used for calculations is much smaller: for example, to store 10 bits of information, 90 physical qubits (10×9) are needed with the 3×3 scheme, and 250 physical qubits (10×25) are needed with the 5××5 scheme 7 Scenarios require 490 physical qubits (10×49). This "redundancy" is necessary because it guarantees the reliability of quantum computing.

"We hope that as these sets get bigger and bigger, so that the qubits will become more and more accurate. The problem is that as these things get bigger, so so do the opportunities for error, so we need good enough equipment so that as we make these things bigger and bigger, the error correction capabilities can overcome these additional errors that we introduce into the system. Google Labs research scientist Michael Newman said at the briefing.

Google says this is a goal that has not been achieved for 30 years, until Willow has achieved a breakthrough - achieving an exponential decrease in error rate with each increase in the size of logic qubits, from 3×3 to 5×5 to 7×7.

This is like building blocks, the higher the blocks, the easier it is to fall, but now Google's research not only allows the blocks to be stacked higher, but also the higher the block, the more stable. This is a strong indication that the practical super-large quantum computers of the future can indeed be built.

This breakthrough is known in the industry as "below the threshold" – the ability to increase the number of qubits while reducing errors. In the paper in the journal Nature, the researchers wrote: "While many platforms have demonstrated different properties of quantum error correction, none of the quantum processors have clearly demonstrated performance below the threshold so far."

"If it doesn't fall below the threshold, then quantum error correction is really pointless, which is really a key factor in the future implementation of this technology." Julian Kelly added, "The quality of the qubits themselves must be good enough for error correction, and our error correction demonstration shows that at the integrated system level, everything works at the same time, and it's not just a matter of the number of qubits, T1 or double qubit error rates." This is one of the reasons why this challenge has been difficult to solve for a long time. ”

"Willow brings us closer to running practical, business-relevant algorithms that cannot be replicated on traditional computers." Hartmut Naven said.

5 minutes to complete the calculation, while Frontier takes 10^25 years

To measure Willow's performance, Google used the Random Circuit Sampling (RCS) benchmark. "Pioneered by Google's quantum AI team, RCS is now widely used as a standard in the field and is the most difficult classic benchmark in quantum computing today." Hatmut Naiven said.

Specifically, RCS is used to demonstrate the rapidly growing gap between quantum and classical computers, and highlights how quantum processors are stripped at double exponential rates and will outperform classical computers as qubits expand. It involves generating and measuring the output of a random quantum circuit (a random quantum circuit is a sequence of qugates applied to qubits in a seemingly arbitrary way).

As mentioned in the opening chapter, Willow's performance in the RCS test was amazing: it completed a calculation in less than five minutes, while today's fastest supercomputer, Frontier, took 10^25 years. "It confirms the idea that quantum computing occurs in many parallel universes, which is consistent with David Deutsch's first proposed idea of 'we live in a multiverse.'" Hartmut Naven said.

Figure: Compute costs are greatly affected by available memory. Therefore, Google's estimates take into account a range of scenarios, from the ideal case of unlimited memory (▲) to more practical, parallel-executable implementations on the GPU (?).

At the briefing, he was asked, "How far are we from seeing quantum computers in practical applications?" Hartmut Naven said that quantum computers have a place in drug discovery, nuclear fusion reactors, fertilizer production, quantum machine learning, electric vehicle batteries, etc.

In terms of drug discovery. "About 75% of small molecule drugs will be metabolized by the P450 enzyme, which is basically a hurdle that small molecule drugs must avoid, which is not yet fully understood, and quantum computers are expected to be able to model it better, and Google is working on this application to try to understand the enzyme complex P450 with quantum computers. ”

In terms of machine learning. "AI is everywhere now, but it's important to recognize that there are many foundational and computational problems, such as solving difficult optimization problems or factoring large numbers (Integer Factorization), that cannot be solved by learning alone because you need huge training data. This is where quantum computers can help. "

Charina Chou, director and chief operating officer of Google Quantum AI, added, "Today's AI mainly refers to machine learning, which requires a large number of training samples. For example, ChatGPT's amazing success is due to the large amount of training data available. In this regard, quantum computing can also help. Google has actually done some work on this front, which will give us algorithms that can get more value from magnetic resonance imaging (MRI) and nuclear magnetic resonance (NMR). These new quantum algorithms can act as an atomic ruler to give very precise distances between nuclei in a molecule. Therefore, quantum computing can help collect training datasets that would otherwise be unavailable, which is another important connection with AI. "

In addition, Charina Chou also pointed out that "the biggest opportunity to simulate nature may be in quantum mechanical systems", and Google is cooperating with many large companies, academic institutions and startups in the fields of physics, chemistry, and materials science to explore the application scenarios of quantum computing in various fields.

Systems engineering is key

In Hartmut Navin's view, systems engineering is key to designing and manufacturing quantum chips: all components of the chip, such as single and double qubit gates, qubit reset, and readout, must be carefully designed and integrated at the same time. If any component lags or the two components don't work well together, it can slow down system performance.

"As a result, maximizing system performance runs through every aspect of our process, from chip architecture and manufacturing to gate development and calibration. The result for Willow is to evaluate quantum computing systems as a whole, rather than just one factor at a time. ”

In addition to Willow's best-in-class performance in both of these system benchmarks (quantum error correction and random circuit sampling), Willow's T1 time (the length of time a measured qubit can retain excitation—a key quantum computing resource) is close to 100 μs (microseconds), which is five times better than the 20 microseconds of the Sycamore chip.

If you want to evaluate quantum hardware and compare it across platforms, here is the key spec sheet:

Chart: Willow's performance across multiple metrics

When asked, "From Sycamore with 53 qubits in 2019 to Willow's new achievements at 105 qubits now, Google's technical roadmap on quantum computing seems to focus more on quality than quantity, does this mean that the industry's general pursuit of 'more qubits' needs to be adjusted?" In response to this question, Hartmut Nevin told Techwalker:

Quantum computers need to have both "quantity" and "quality" at the same time. Simply increasing the number of qubits is not enough, because if the error rate is too high, these qubits cannot be used efficiently. It's like a computer, if it crashes frequently, it won't work even if it's high.

If a quantum computer's gate operation error rate is 1 in 1,000, then after a thousand operations, the system is likely to fail. In practice, each qubit needs to perform at least ten gate operations. So for a system with 100 qubits, the error rate needs to be controlled to 1 in 100,000 to be qualified.

In contrast, some other designs claim to have thousands of qubits, but have error rates as high as 1/50 or 1/200. In this case, it is simply not possible to use all the qubits at the same time until the entire system collapses. That's why Google chose to focus on improving the "quality" of qubits first, because it only makes sense to increase quantity if the quality problem is solved first. ”

Google's research team says they are developing new technologies to scale up the system. The current focus is on reducing the error rate and making it meet the requirements of quantum error correction. As the technology matures, the number of qubits will gradually increase.

Google's Quantum Computing Journey

So far, Google has conducted two different types of experiments with quantum computing.

On the one hand, run the RCS benchmark, which measures performance against traditional computers, but has no known real-world applications.

On the other hand, scientific simulations of quantum systems have also led to some new scientific discoveries, but these discoveries are still within the scope of conventional computers.

Figure: Random Circuit Sampling (RCS), while challenging for traditional computers, has yet to demonstrate real-world commercial applications.

During the video briefing, Google's Quantum Computing AI team unveiled Google's Quantum Computing Roadmap, which Google says is focused on unlocking the full potential of quantum computing by developing large-scale computers capable of complex, error-correcting calculations, and these milestones will lead us toward high-quality quantum computing hardware and software for meaningful applications. As you can see on the diagram, the roadmap contains six milestones, and Google has completed two milestones so far.

Google Quantum Computing Roadmap

Talking about devoting himself to this quantum computing journey, Hartmut Naven wrote on Google's official website:

"My colleagues sometimes ask me why I left the burgeoning field of artificial intelligence to focus on quantum computing instead. My answer is that both technologies will prove to be the most transformative of our time, but advanced AI will greatly benefit from quantum computing. That's why I named our lab Quantum AI. ”

"Quantum algorithms have basic scaling laws, and as we saw in RCS, many fundamental computing tasks that are critical to AI have similar scaling advantages. Therefore, quantum computing will be essential for collecting training data that is inaccessible to traditional machines, training and optimizing certain learning architectures, and modeling systems that are important for quantum effects. This includes helping us discover new drugs, designing more efficient batteries for electric vehicles, and accelerating progress in nuclear fusion and new energy alternatives. Many future game-changing applications that won't work on a traditional computer; They are waiting for quantum computing to unlock. ”