Enchanted by computers since his childhood
The word “architecture” refers to a kind of “structure.” Computer architecture, then, is the study of the structure of a computer with all its dynamic machinery, just like a factory or a castle, concerned with both the practical and theoretical aspects: the foundation, the layout, the positions of machinery, the flow of people, and the use of energy on the one hand, design philosophy and engineering technology on the other. The aim is a computer that is ever faster, more powerful, and more energy efficient, a future device that we cannot even imagine yet. In the realm of computer architecture, Associate Professor Takamaeda says he wants to be a “warrior wizard.”
“In the game “Dragon Quest,” if you master both the warrior and the wizard, you can advance to the level of a “warrior wizard.” If a warrior is an expert in hardware such as semiconductor circuits, and a wizard is an expert in theory such as AI, then there must be ideas that come to you not because you are the best warrior or the best wizard, but because you are both, a warrior and a wizard. I think that is what researching computer architecture requires. That said, in the world of “Dragon Quest,” warrior wizards are very weak (laughs), which makes me want to become an unparalleled warrior wizard.”
Takameada has been interested in computers since he was a small child. To amuse himself in his early elementary school years, he would draw pictures of keyboards on a binder notebook, creating almost a mock-up of a laptop computer. He got his first real laptop in his first year in technical college, bought with money saved from his part-time job, right around the time Windows XP appeared. He wanted to use the Internet, which was starting to become widespread at the time.
“I took a computer science class in technical college and fell in love with programming. I felt like I had everything I needed in my hands. It was exciting that on a computer, I could do whatever I set out to do. I especially enjoyed programming, and my interest in computer engineering grew, so I joined Tokyo Tech.”
He planned to get a job after he had finished his master's degree doing research on prototyping systems, technologies that make possible the evaluation of future CPUs without simulation. He changed his mind when he saw some of his seniors interact at a conference.
“It was when we were having drinks at an after-party that I saw assistant professors and post-doctoral fellows of different ages having heated discussions as equals. They seemed to be having a lot of fun which made me want to become a researcher too. Until then, I had been thinking of becoming a management consultant for the earning potential, but those aspirations disappeared in an instant," he says with a nostalgic smile.
As a part of the doctoral program, Takamaeda spent some time at Carnegie Mellon University, where his research became his major driving force.
“I was there for less than two months, but those were truly precious two months. Carnegie Mellon University’s computer science division is world-class. I did research on compiler software for various hardware designs under Professor James C. Hoe, famous for computer architecture and field-programmable gate arrays (FPGAs). It was summer, and the days were long. The sun often stayed up until 10 p.m. So, I spent my days writing code.”
Soft hardware
FPGAs are what Takamaeda calls “soft hardware”: semiconductor devices that allow users (programmers) to change the configuration of digital circuits at will.
The heart of a computer consists of an engine called a CPU combined with various elements such as arithmetic and control circuits, register files, and cache memory. Once manufactured in a factory, these hardware components cannot be changed by tinkering with the circuits. In FPGAs, however, you can manipulate the circuits electrically as you wish, using software. The property that they can be freely assembled like toy blocks is why FBGAs are called "soft" (electrically, of course, not materially). A hardware description language is, so to speak, a blueprint for assembling the blocks.
In fact, FPGA research is what Takamaeda has been working on since he was an undergraduate student and is still his primary interest.
“With FPGAs, you can create any kind of CPU you envision, for example, one dedicated to AI. It's a device that can shapeshift like Ditto in Pokemon (laughs). My doctoral research investigated arrangements that used FPGAs to mimic future CPUs of 100 or 200 cores in a row.”
Veriloggen and NNgen, developed by Takamaeda but made available as open-source software, are compiler software that enable the free design of various circuits in FPGAs using the general programming language Python.
“If exploiting the power of software made hardware design easier, then more engineers would be able to create all the complicated and intricate circuits they can think of. If that resulted in computers that are 1% more efficient, that would lead to computers worldwide becoming 1% more efficient, which would be quite an achievement.”
The hardware design tools developed by Takamaeda give birth not only to broadly applicable new technologies but also to more efficient, specialized computers developed for specific purposes using FPGAs. But how does this connect to computer architecture?
“In the past, most research on computer architecture focused on what CPUs should be like, with all the detailed mechanisms inside them. However, I take a more extensive view. I think that precisely because computational mechanisms are conjointly implemented by software and hardware, both are important research targets of computer architecture. In any case, I like taking a wider approach and trying different things. That is my research style.”
Future Computers and Octopus Tentacles
Takamaeda strongly feels that hardware design technology is his life's work.
“I am fascinated by technologies that interact with the structure of circuits in CPUs and AI chips. They can be thought of as the hardware version of programming languages. I believe that leveraging them will be crucial in the future. I am confident that I can technically manipulate the lower layers of a computer. Building on this as a foundation, I want to improve the efficiency of significant computational methods, such as AI. We are also enthusiastic about working on the research and development of AI chips suitable for neural network calculations.”
The "lower layers" that Takamaeda refers to are the layers of circuitry and other hardware. In "Dragon Quest," this would be the world of the warrior. The layers above are the highly abstract layers of theories and programs, the world of the wizard. The “hardware version of programming languages” could otherwise be described as the magic of the wizard implemented in the hands of the warrior. The goal is still higher speed and higher energy efficiency because that is what society and the Earth need.
“There are many new emerging applications based on neural networks and deep learning that no one could have imagined,” Takamaeda says. “ChatGPT is one such example: by investing enormous amounts of computing power and data, we are creating an intelligence powerful enough to be a potential threat to humans. However, if we are serious about creating human-like intelligence, the computing power of today's computers is absolutely inadequate. It is often said that humans can use their brains to do advanced calculations all day long with the energy of a sandwich, but computers are far from that level. There is no computer in the world yet that is as energy efficient as a human being which means we need computers to be faster and smaller. That is a social necessity as well, and that is why I am working on AI hardware.”
Of course, steps toward higher speed and efficiency have already been taken in various areas.
“Developing digital neural networks that work efficiently is crucial and our lab is doing research in this area. However, we also think that it might not be necessary to stick to digital 1s and 0s. For example, an idea that is worth investigating further is “approximate computing,” a kind of computation that allows for ambiguous properties. Instead of aiming for perfect accuracy, people can decide how much calculational error they can tolerate and use much less operational energy to get “ballpark numbers.” Also, using analog circuits similar to natural phenomena to perform calculations was proposed decades ago, but then it was abandoned because of the dramatic progress in semiconductor integration due to Moore's Law. Recently, however, Moore's law seems to have reached its plateau, and the idea of computing with analog circuits, which may fluctuate but require less power, is emerging once again.”
Research worldwide is being conducted on reservoir computing, a method that uses liquids, or even octopus tentacles, for computation. There are approaches to future computers other than quantum computing.
“Quantum computation requires error correction done by a classical computer. Therefore, when thinking about future computers, it is not enough to work solely on quantum computers. It is also necessary to be familiar with classical computers and to know about the more unusual methods like reservoir computing as an alternative to quantum and digital circuits. There are truly many “ingredients,” so I think computer architecture is about how to combine and “cook” them together.”
Having a great impact on the world
Takamaeda describes himself as a carpenter who is more interested in the work that he does and the tools he uses, rather than the house he builds. Or, in terms of the two founders of Apple, he is Wozniak, not Jobs. Even so, he also wants to pursue the “truth” of what a computer could be. What drives his research is that he doesn't know what that “truth” is.
“I am personally excited about computation in memory (CIM), or processing in memory (PIM), where calculations are run in the memory of a computer. Normally, the circuits for computation and memory are separated and data are transferred in and out of the memory when computations are performed. CIM, however, is the concept of performing calculations in the memory, without the very energy-intensive process of transferring data. This makes CIM very low energy. There are many papers on this subject, but it is challenging to create CIM circuits that work properly. That motivates me to try to build them.”
Once technology cannot keep up with Moore's law anymore and transistor integration stabilizes in the future, trying new ideas to increase energy efficiency will be the role of computer architecture, Takamaeda believes, because every aspect of society needs faster computers and higher energy efficiency, be it AI, drug discovery, disaster prevention, or weather forecasting. So, Takamaeda has a message to young people, inviting them to join him in researching computer architecture.
“If you understand not only the computers and AI we are using right now, but also the underlying mechanics and computer science, then you can create the advanced structures of the future. You cannot create novelties if you do not understand how things work currently. This is true not just for computer architecture and CPUs, but also for AI, machine learning, and natural language processing. Come and join us, let’s study thoroughly, and create the next big thing together. I am sure we can create something that will have a great impact on the world.”
In his private life, Takamaeda is a car-loving father of two. When asked what he enjoys most about his daily research, Takamaeda replies like this.
“The discussions I have with my students every Thursday when I'm in the lab. “Could it be better this way? This doesn’t seem right… Oh, this is what it’s about!” Having such conversations with my students forces me to think deeply as well. I have to constantly use my head, so those are the times when I feel the smartest (laughs). I come up with many ideas that I couldn't come up with on my own, so it is definitely the most enjoyable part of the job,” says the wizard warrior with an unparalleled smile.
laboratory: https://sites.google.com/view/casys/
※Year of interview: 2023
Interview/Text: Minoru Ota
Photography: Junichi Kaizuka