Program Operation Performance
The 2nd company of 2020 Fall Startup-ting is FuriosaAI! FuriosaAI joined us for LunchTalk, followed by Startup-ting. Due to COVID-19, the session was held online through Zoom.
FuriosaAI is a startup that designs and develops customized AI inference coprocessors, and received more than 10 billion won in investment from Naver D2SF, etc. Like NVIDIA and Google, FuriosaAI is growing as a competitive company.
FuriosaAI’s Startup-ting was divided into parts – company introduction, semiconductors, hardware, software – followed by a Q&A session.
In this Startup-ting, the CEO gave a brief company introduction, answered questions asked in advance, and gave a lecture on FuriosaAI and semiconductor chips.
Competition in the Global Battlegrounds: a Breakthrough to Conceptual Design – CEO June Paik
The first part of Startup-ting started off with a lecture focused on FuriosaAI by CEO June Paik. FuriosaAI is a developer and manufacturer of AI inference coprocessors, founded in 2017. AI coprocessors are similar to those of NVIDIA and Google. FuriosaAI develops AI as well as products.
Semiconductor chips are difficult to make, but it is even more difficult for startups to compete with semiconductors because NVIDIA is used widely in the market. However, if it is developed, it will become competitive not only in the domestic market but also in the global market. NVIDIA, Qualcomm, and Intel also started off as startups, not as big companies. Currently, AI is emerging in the market, so many related startups are bound to appear. FuriosaAI said that at this point, there is a possibility for startups to succeed in the market.
CEO Paik briefly introduced FuriosAI, then proceeded with a Q&A session. Since AI is an area of interest for KAIST students, there were a lot of questions asked not only in advance but also in real-time chatting. A memorable question was “What is the most important thing to study as an undergraduate to research and develop AI chips?”, to which CEO Paik answered that developing AI coprocessors involves the development that targets microprocesses, so it’s important to study the basic concepts at school and practice making semiconductor chips. FuriosaAI makes and designs these chips itself.
The next question was “Some people say that the environment of doing a fab-less business in Korea is relatively poor compared to the overseas. Did you ever feel inconvenienced while leading a startup?”, and CEO Paik replied that doing a fab-less business is difficult whether it is in the U.S. or Korea. Because it is a global product without borders, it must be done as a global business. He said that if the semiconductor chip is good enough, it will have a competitive edge in the global market.
FuriosaAI is a semiconductor company, but there are 60% software and 40% hardware engineers working.
There was also a question like “AI chips have a high entry barrier, but how did FuriosaAI decide on such business goals?”, and Paik said that he didn’t know he would enter the semiconductor business, but when he took a sick leave in 2014, he thought of AI as a comprehensive study and studied architecture, signal processing at school. After returning to the company, he thought AI semiconductors was an opportunity and started a new company with people around him. A new road will open if one continues to do what he or she has been doing. In the case of early team building, CEO Paik worked with CTO Hanjoon Kim previously, and other team members were introduced by acquaintances. He mentioned that building a team with those who previously worked together or are introduced by acquaintances is probably the most proven way. Currently, FuriosaAI has more than 30 employees, with talented people from Samsung, KAIST, Seoul National University, Postech, and MIT. An important thing while doing a business is to have ideas, set high but realistic goals, which gives the power to decide for a company.
It seems like there were a lot of people who fell for FuriosaAI during the Q&A session. Paik added that anyone with the fundamental technology, passion, desire to create a GPU and experience the engineering process is welcome in FuriosaAI.
Pursuing the best computational structure – CTO Hanjoon Kim
The second part of the Startup-ting was a lecture about hardware architecture by CTO Hanjoon Kim. CTO Kim earned a Bachelor’s and Integrated PhD degree in the department of Computer Science at KAIST. The lecture was as follows.
What problems are we trying to solve?
We need a programmable architecture for deep learning
- New deep learning algorithms can be optimized on it
- Provide a good abstraction to tools & algorithms
- Should be based on a fundamental understanding of algorithm, architecture, compiler, and system
New methodology, team, and infrastruture are necessary
CTO Hanjoon Kim was worried about which problem to solve while studying for 10 years. His desire to solve problems in the world was so large that he decided to study computer architecture, which is a field about designing chips like Intel’s CPU and GPUs. After trying implementing chips in research papers that he studied, he saw that the performance lacked compared to that of Intel, and realized that the architecture inside Intel’s CPUs were much more sophisticated.
After 10 years of study, he entered a company where he encountered practical problems, and decided to create a system to solve the problems. CTO Kim did prior studies and worked hard to solve problems that would be helpful to the society. Kim found out that Architecture, Compiler & Tools software stack, Deep learning Algorithm, Network Architecture Search, Board & System, Chips, and etc. must all come together to make an AI coprocessor. Just like the amp lab shocked the world by making a spark, CTO Kim wanted to solve an impactful problem in the world and recommended that KAIST students should also do such thing.
The first half of CTO Kim’s lecture contained his philosophy, while the second half was about Challenges in Architecture. Below is a brief summary of the lecture.
- Pursuing the best computational structure
- how to build chip for deep learning
As various algorithms are discovered and developing, there may be cases when there are tenfold differences in performance between algorithms. It is important to find a fast algorithm. However, the building speed cannot keep up with the speed of algorithm development, and hardware isn’t set up delicately. To solve this problem, it is important to maximize the use of transistors, which must be configurable enough for certain accelerator architecture and for all algorithms. Since algorithms are constantly being introduced, architecture is not exactly specialized. It is important to understand deep learning algorithms correctly to design architecture and compilers. There are a lot of tasks that require a lot of engineering to design compilers. Instead of just focusing on accelerating the process, it is important to think about how to design the soft steps and the hardware (see below).
Software Architecturing/Implementation-Physical Design/Manufacturing-Verification-RTL Implementation-Performance Modeling/Architecturing
As algorithms change, the framework, architecture, and compiler are changing as well. Different from problems arising while making IPUs in previous fab-less productions, making AI coprocessors should be done in a comprehensive way. While making such chips, Methodology, Team, and Infrastructure are important.
Pushing the impossible limit-Software Jaeseung Ha
The third session was about the software team at FuriosaAI. Jaeseung Ha of the software team previously worked at Nexon, NC SOFT, Neople, and has 10 years of game development experience. He gave a lecture in a self-Q&A section structure, and introduced himself. He left a game company to work at FuriosaAI, because he thought that FuriosaAI’s vision of deep learning will be more widely used in the future and that deep learning accelerators were essential technologies. The tasks of performance, testing environment, and demonstrations were similar, which is why he decided to move. To the question about what kind of work a software developer does at a hardware development company, he answered that he develops a deep-learning related software and a deep-learning accelerating compiler.
FuriosaAI’s compiler also provides a TensorFlow-lite compatible API, which enables running deep learning production models. In addition, Neural Architecture Search is important, so a research on finding a smaller and faster model that does the same task is being conducted. In the current deep learning development environment, if you do not have a software, then even if you have a good chip you need a software stack. Developer Ha is in charge of making Python bindings, and various software tasks. FuriosaAI’s deep learning accelerators does not pale in comparison to NVIDIA’s RTX. In addition, he mentioned that developing the architecture as a software is also important.
Developer Jaeseung Ha says that FuriosaAI is a startup with a free environment and a lot of opportunities and smart people. At FuriosaAI, you can solve valuable problems, contribute to the development of AI, as well as compete with NVIDIA and Google.
Q&A and Discussion
After Developer Jaeseung Ha’s lecture, there was a followed up live Q&A session. One of the questions was about what kind of strategy Furiosa has in competing with NVIDIA, and CEO Paik answered that there will be a deep-learning accelerator that will eventually have better performance than Google, Tesla, and NVIDIA’s GPU. However, not all companies can make these chips, and Furiosa can meet the needs of companies like Naver and Kakao who demand customized chips. It’s difficult for startups to survive in the semiconductor market, because developing chips require a huge budget, but Furiosa believes that they can attract investment if they own a competitive edge.
During the session, there were also a lot of participants who were interested in the FuriosaAI internship. Information about the internship program will be sent by e-mail to interested participants in the future.
Although FuriosaAI’s Startup-ting was online, participants stayed until the end and communicated in real time. Students seemed to be interested in the AI field, and thanks to CEO Paik and others explaining terms used in the field, participants were able to understand fully and become more interested. FuriosaAI became well known in the AI semiconductor chip market within 2 years of its founding, and became a company that competes with NVIDIA and Google. We look forward to the future of FuriosaAI, and hope that it becomes the leading company in the industry just like Furiosa of Madmax.
Although it was online, thank you for participating in FuriosaAI’s startup-ting! Please look forward to the next startup-ting ????