Google’s AI division DeepMind once became famous for defeating the chess king with AlphaGo. This week, DeepMind announced the latest AI model research and development result AlphaDev, which can speed up the operation speed of the data center and has energy-saving effects.
AlphaDev is an improved model of AlphaZero (which once defeated AlphaGo) and MuZero, who are proficient in chess and video games. The two later changed to optimize the data center and image compression as the main axis. AlphaDev is a specialized version of AlphaZero, which discovers new data sorting and hashing algorithms that can speed up the execution of software code.
AlphaDev’s search for new algorithms begins with low-level programming instructions that computers read, rather than humans writing high-level languages such as C++. They believe that it is easier to find room for improvement in low-level combined instructions than in high-level programming languages. Computer storage and computing are more flexible at this level, which means that it is easier to have breakthrough technologies to increase speed or reduce energy consumption. They hope that AlphaDev can find new data sorting and hashing algorithms, because these are the two most basic processes for people’s data sorting, storage and retrieval today. Ranking algorithms can affect how digital devices process and display information, rank posts from searches or social networking sites, or recommend users.
As a variant of AlphaZero, AlphaDev is also a reinforcement learning model. To train AlphaDev to find new algorithms, the researchers turned sorting into a “combination game” for one player. In each round, AlphaDev observes the algorithm it generates and the information obtained by the CPU, and then selects an instruction to add to the algorithm before starting the next round. AlphaDev must find a sorting algorithm from a very large number of instruction combinations, and find a faster and better algorithm every round. The number of command combinations is roughly equivalent to the number of molecules in the universe, or chess moves (10 to the 120th power) and Go (10 to the 700th power). And if you take any wrong step, the entire algorithm may be useless. Researchers will reward AlphaDev’s sorting from two aspects, one is correctness, and the other is efficiency and speed.
Finally, AlphaDev discovered a new algorithm that can improve the speed of the low-level virtual machine (LLVM) libc++ sorting library, which can increase by 70% in short (3 to 5 element) sequences, and in the long sequence of more than 250,000 elements. Sequences increased by 1.7%. Algorithms for short sequences are the focus of DeepMind’s research because they are more commonly used. DeepMind pointed out that on simple user search tasks, the AlphaDev algorithm can improve the sorting speed, but once it is applied to a larger-scale environment, such as a data center, it will save a lot of energy and cost.
Finding hash algorithms is also one of AlphaDev’s tasks. Hash is often used in data storage and retrieval, such as in databases. The algorithm found by AlphaDev can increase the efficiency by 30% when it is applied to the hash function of 9 to 16 bytes in the data center.
Millions of downloads since publishing the sorting algorithm for the LLVM standard C++ library to replace sub-routines that had been used for over a decade, and the hash algorithm for the abseil library , these algorithms have been used in various industries, including cloud, online shopping and supply chain management.
DeepMind expects that, just as its algorithm has crossed from chess playing to data center computing, more general-purpose AI models will be further used in modern life in the future.