Source: Zeping Macro

On December 10, Google announced the latest generation of quantum chip - Willow, which caused a sensation in the global technology community. Musk even exclaimed "Wow"!

What is so powerful about Willow chips? How far is it from mass production?

1. Google’s latest generation quantum chip Willow was launched. The biggest breakthrough lies in its super computing power and error correction capabilities.

For a benchmark task called “Random Circuit Sampling”, the current fastest super It would take a computer 10^25 years to solve the problem, which is far longer than the age of the universe (26.7 billion years); Willow completed the task in less than 5 minutes.

Quantum computing has the potential to significantly increase calculation speed and surpass classical computers on specific tasks. This is called "quantum superiority". As early as 2019, Google had verified this fact and published it in Nature, showing that it used a 54-qubit quantum computer Sycamore to achieve tasks that traditional architecture computers cannot complete: the world's first supercomputer needs to calculate In a certain experiment on 10,000 years, Sycamore only took 3 minutes and 20 seconds. At that time, Google CEO Sundar Pichai said that this was the "Hello World" that researchers had been waiting for for a long time, and it was the most meaningful milestone in the practical application of quantum computing up to that time.

The release of Willow is undoubtedly another landmark event in the field of quantum computing.

However, "fast" is not Willow's most noteworthy breakthrough.

The biggest highlight of Willow is its super error correction ability.

In the past, during the data processing process of quantum chips, due to the fragility of quantum states, they were easily affected by environmental interference and decoherence occurred, resulting in errors in the state of qubits. Therefore, despite their "quantum superiority", quantum computers are susceptible to environmental influences and are very error-prone. Typically, the more qubits there are, the more errors can occur.

Therefore, "quantum error correction" has become a key technology. Quantum chips require special quantum error correction technology. This is also an important challenge in this field and once seriously restricted the practical application of quantum computing. and development.

The Willow chip has successfully solved the quantum error correction problem that has plagued researchers for nearly 30 years, achieving an exponential reduction in error rates. Google's research shows that the more qubits used in Willow, the lower the system's error rate.

When the number of qubits increases, expanding from a 3×3 array to a 5×5 to a 7×7 array, each expansion in Google’s Willow chip experiment can reduce the coding error rate by 2.14 times, the error rate drops faster and faster.

2. What is quantum computing? Why is it so powerful?

In 1935, the Austrian physicist Schrödinger proposed a great thought experiment: Put a cat in a box with radioactive material. There is a 50% probability that the radioactive material will decay and release poisonous gas to kill the cat. There is also a 50% chance that the radioactive material will not decay and the cat will survive. Before opening the box, no one knows whether the cat is alive or dead. It can only be described as "in a superposition state of life and death."

The quantum world, just like "Schrödinger's cat", is in an unresolved superposition state; the corresponding new computing theory is "quantum "Computing", the hardware layer is represented by quantum chips and quantum computers.

Quantum computing shows two advantages:

First, powerful data storage capability. Classical computing uses bits as the basic unit, while quantum computing uses qubits as the basic unit.

In classical computing, the state of a bit is determined, either 0 or 1; while a qubit is in a superposition state of 0 and 1. In other words, it can store 0 and 1 at the same time. .

A traditional chip with n bits can store n pieces of data at the same time; while a chip with n qubits can store 2^n pieces of data at the same time.

Second, it demonstrates powerful parallel computing capabilities for specific problems.

Traditional electronic computers are serial calculations. Each operation can only convert a single value into another value, which means that it must be calculated in sequence. A quantum computer can simultaneously convert 2^n data into new 2^n data in one operation.

3. Can future quantum chips replace GPUs and promote the development of AI?

Artificial intelligence technology and various applications have developed rapidly in recent years, and the demand for computing power has also increased exponentially.

Theoretically, the parallel processing capabilities of quantum computing give it a natural advantage in processing complex artificial intelligence algorithms, and can greatly improve the training speed and accuracy of the model. The emergence of Willow chips may provide powerful computing power for the further development of artificial intelligence.

In fact, GPUs, which are now widely used in AI, were originally designed to accelerate graphics processing. For example, 3D scene rendering in games, modeling and special effects processing in animation production, video visual effects in film and television production, etc. However, due to its powerful computing power, GPU was later widely used in the fields of scientific computing and artificial intelligence, especially in the neural network training and inference stages of deep learning. It performs well in processing large-scale data sets and high-parallel computing tasks. outstanding.

From this perspective, quantum chips will gradually make breakthroughs in the future, breaking through computing limitations and accelerating the training process of various AI machine learning algorithms. Quantum chips are currently mainly used in some specific fields that require extremely high computational complexity, such as encryption algorithm cracking in cryptography (such as the construction of traditional encryption methods based on the RSA algorithm).potential threats), quantum system simulation (simulating the physical and chemical properties of molecules, materials, etc. at the quantum level), solving complex optimization problems (such as logistics planning, resource allocation and other complex combinatorial optimization problems), etc. In these areas, the advantages of quantum computing can be fully exerted, and it is possible to solve tasks that traditional computers cannot complete within an acceptable time.

The growth of quantum chip computing power is mainly related to the increase in the number and quality of qubits. In the future, as the number of qubits increases, the computing power of quantum computers will increase exponentially. Each additional qubit doubles the number of possible state combinations. For example, 2 qubits have 4 state combinations, 3 qubits have 8 state combinations, and so on. At the same time, the quality of qubits (such as coherence time, fidelity, etc.) also has an important impact on computing power. High-quality qubits can maintain quantum states more effectively, thereby achieving more accurate and complex calculations.

However, in the short term, it is difficult for quantum chips to shake the status of GPUs. Quantum chips have stronger computing power than GPUs and can theoretically be replaced. But the moat of GPU is only one aspect: computing power. What is more important is: programmable architecture and developer ecological advantages, manufacturing technology and industry maturity.

GPU’s programmable architecture and developer ecosystem are core barriers. The “AI computing power revolution” launched by NVIDIA with GPU has been paving the way for more than ten years.

CUDA (Compute Unified Device Architecture) is the first GPU programming architecture platform developed by NVIDIA in 2006. Its value lies in building a GPU developer ecosystem, and algorithm engineers can customize the capabilities of the GPU according to their own needs. Discover, this also expands the application field of GPU from graphics rendering to general field.

If you develop new software based on new hardware (such as quantum chips), you need to achieve forward compatibility. However, existing major AI software basically relies on CUDA platform development, so breaking away from the CUDA architecture requires high costs. cost. Coupled with the moat effect of the development community, many high-performance computing developers have accumulated development experience in the CUDA ecosystem. CUDA has up to five million downloads every year. It will take ten years to push the developer community to other programming models. planned project.

GPU chip manufacturing technology and industrial chain are mature, with a broad consumer market and a positive industry cycle.

It has been 25 years since the birth of GPU, and downstream commercial application scenarios such as personal PCs, customized development, and AI data centers have been formed for 10 to 30 years. Currently, GPU takes one year from chip project establishment to tape-out, and one year from tape-out to mass production. With GPU development as the main tone, a corresponding linkage cycle has been formed such as lithography equipment development and wafer foundry process iteration. It is difficult for such a solid industrial chain to survive under the positive cycle of more than ten years.Be broken.

It is difficult for quantum chip manufacturing and GPU industry chains to overlap. The design and manufacturing process of quantum chips are extremely complex and require a highly pure experimental environment, precise quantum control technology and stable qubits. Therefore, for a long time, a few top technology companies have been "working alone" and have not yet matured. industrial supply chain. Therefore, it is a big problem to achieve mass production and commercial application of quantum chips in the short term.

4. The areas where quantum chips have the greatest impact: cryptocurrency and “HPC+AI”

4.1 Quantum chips may be the “nemesis” of cryptocurrency

Take Bitcoin as an example. Its security Built on two key mechanisms. The first is the "mining" mechanism. Bitcoin output is based on proof of work (Proof of Work) that relies on hash functions. The higher the hash rate, the greater the possibility of successful mining. The second is transaction signature, which is based on Elliptic Curve Digital Signature Algorithm (ECDSA) and is equivalent to the user’s “identity wallet”. The design of these two mechanisms makes Bitcoin almost impossible to crack in traditional computing, and quantum chips will pose a direct threat to Bitcoin.

The first is quantum computing’s violent cracking of the “mining” mechanism. Quantum computing algorithms can accelerate the calculation of hash functions, that is, the mining speed is accelerated, and the extent is greater than that of all previous traditional equipment. As a result, the mining success rate increases, the supply of cryptocurrency increases sharply, causing significant fluctuations in its market price. On December 10, Bitcoin fell from US$100,000 to US$94,000. Coinglass data shows that a total of 237,000 people liquidated their positions from December 10 to 12.

The second is the direct threat of quantum computing to transaction signatures. There are two types of credentials for cryptocurrency transactions: "public key" and "private key". The former is equivalent to a bank card number, and the latter is equivalent to a wallet password. Normally, the disclosure of public key addresses does not affect the security of users' funds, but quantum computing can use public keys to crack signatures and forge transactions. For example, Shor's algorithm in quantum computing is specially used to crack the prime factorization and discrete logarithm problems of large integers, which will pose a serious threat to transaction signatures.

Although Willow poses little threat to Bitcoin at the moment, it is very likely that cryptocurrencies will be broken through by quantum computing in the future. Theoretically, to launch an attack on Bitcoin's signature and mining mechanism, approximately several million physical qubits are needed. Compared with the 105 physical qubits Willow currently possesses, the gap is still very large. But if Willow iterates like a general-purpose GPU and achieves mass production and computing power leaps, then it is not impossible for Bitcoin to be "compromised" in the next ten years.

4.2 Quantum chips will promote "HPC+AI" and promote the development of high-end artificial intelligence

According to OpenAI's classification of AI, from L1 (Chatbot) to L5 (AGI), currently The development of large AI models is only in the transition stage from L1 to L2. AGI at L5 level is defined as “having organizational level capabilities”.Dynamic and complex real-world environments enable judgment, reasoning, prediction, and planning of actions. The industry believes that "HPC+AI" will be a key step in realizing AGI.

High performance computing (HPC) refers to the use of powerful computer capabilities to solve science, engineering and technology implementation problems, and today's AI big The models are homologous to a certain extent, but have different directions and focuses.

HPC focuses on "complex problem solving". For example, the application of supercomputers in meteorology, physics, astronomy and other fields has brought major scientific research breakthroughs.

The AI ​​model focuses on "inference and generation". Although it is not good at solving complex models, it has good versatility.

The implementation of quantum chips is a revolutionary breakthrough in the field of HPC. The solution of complex problems no longer requires the long-term "brute force calculation" of traditional HPC, but can be developed in a new direction - combined with AI For more complex general training.

First, traditional AI training cannot process qubit data, while quantum computing can optimize specific learning models that cannot be processed by traditional computing and build system models that are sensitive to quantum phenomena. That is to say, future AI models will have the ability to reason and predict complex worlds, reducing or even eliminating the phenomenon of "AI hallucination" compared to current large models.

The second is the advantage of quantum error correction technology. Willow chip has overcome the key challenges of quantum error correction and achieved a significant reduction in error rates. In high-level AI training, the application of quantum error correction technology can ensure the accuracy and reliability of models when training and processing large amounts of complex data, reduce calculation errors caused by the fragility of qubits, thereby improving the effectiveness and reliability of AI training. reliability.

Although current AI training does not yet have the conditions to apply quantum chips, it is very likely that quantum chips will be needed as the core support of computing power in the future. Because qubits are extremely sensitive and easily affected by external environmental factors, including temperature and electromagnetic fields, these factors may cause decoherence of quantum states, thereby affecting the accuracy of calculation results. Although Willow has made certain progress in quantum error correction technology, in actual artificial intelligence training applications, in order to achieve long-term stable operation, the stability and anti-interference performance of the quantum system still need to be further improved.

Google released Willow, a new generation of quantum computing chip, which caused a huge sensation in the global technology community. This is not only a major breakthrough in the field of quantum computing, but also the next cutting-edge of global technology.

There are still thorns in the road to the development of quantum computing technology in the future, and there are still many problems to be solved before large-scale application in AI training.

Technological progress has never been a smooth road, just as GPU has gone from obscurity to brilliance.