Build the strongest shogi AI machine for professional shogi players!Everything started from a consultation (ITmedia NEWS) --Yahoo! News

"Ryzen Threadripper 3990X" with an amazing number of cores of 64 cores and 128 threads

"I would like to use deep learning shogi software ..." ──In mid-November, a consultation with the author came in. [Image] Image of Akihito Hirose 8th Dan, who is sometimes referred to as "a shogi player whose soul was pulled out by Yoshiharu Habu" The main speaker is Akihito Hirose 8th Dan, a professional shogi player. He has also won the titles of the throne and the Dragon King, and he is one of the top professionals (in the A class, which is the highest class in the ranking battle). Even if he is not familiar with shogi, if you are familiar with net memes, you may have seen the image of "a shogi player whose soul was pulled out by Yoshiharu Habu" (he said, "just look up and think". That).

An era in which professional shogi players study on computers

In the shogi world, "Ponanza" beat Shinichi Sato 4th Dan (at that time) in 2013, and in the Go world, "AlphaGo" beat Lee Sedol 9th ​​Dan (at that time) in 2016. The research used is in full swing. For shogi fans, it is well known that Souta Fujii, who is thrilling now, is building his own PC with AMD's high-end CPU "Ryzen Threadripper 3990X" for analysis. Since I've been walking around ITmedia PC USER, Mobile, etc. and now I'm in NEWS, I'm in a position to know a little more than the average in the world about hardware specifications and requirements required by AI. Since I had a personal relationship with him, I was asked to consult with Hirose Hachidan (?).

Is it not enough to just strengthen the GPU?

I see, I want to study shogi with deep learning. I heard that I have a machine with a 9th generation Core i9 9900K, but I don't have a graphics card (a card with a GPU). Then, the first impression was that it would be enough to insert one or two additional high-end GPUs "GeForce RTX 3090" from NVIDIA. However, there is also the idea that if you win the title, you will get a prize of millions to tens of millions of yen, so you should make as many PC specs as you can. However, I don't want to propose useless things such as large-capacity SSDs and HDDs that are not used. The idea was that only an additional GPU would be sufficient to achieve the required effect at the lowest cost. However, there was something I was worried about. The point is that Ryuo Fujii uses Threadripper, which is a high-end CPU. Generally, GPU is more suitable than CPU for deep learning calculations. The CPU can perform general-purpose calculations from the operation of the OS to the execution of Web browsers and game software, but it is not "generally" very strong in parallel computing. The GPU is originally a device that specializes in calculating the video output to the display. For example, if you want to display an image on a full HD 60fps display, you need to draw 1920 x 1080 pixels = about 2 million pixels 60 times per second. If one drawing of one pixel is one calculation, it will be calculated 2 million pixels x 60 times = 120 million times per second. If it is a more familiar form (this expression is not accurate), it can be expressed as 120MHz, so it does not seem to be a big calculation, but what I want you to keep in mind here is that "drawing of video is parallel calculation". The GPU does not have a "processor that can do anything" like the CPU, but instead has many "processors (cores) that can do simple calculations". By calculating at the same time with many cores, the pixel group of the appropriate color is output for each frame. In this way, the GPU can be regarded as a device specialized in parallel computing, so it has come to be used for parallel computing other than video processing. And deep learning is compatible with parallel calculation because of its calculation algorithm (repetition of matrix calculation).

1/2 page