본문으로 건너뛰기
0%
약 3분 (713 단어) ko

Introduction to Preferred Networks

Categories DeepThinking AI
Tags #DeepThinking #ToyProjects #TensorFlow #AI #DL #ML #Preferred Networks #GithubPage #Retrospect #AdaptiveAUTOSAR #AUTOSAR #ClassicAUTOSAR #ECU #CPU #GPU #OTA

About this post

This article is adapted from a post by Jihyo Lee, formerly of Bain & Company.

Preferred Networks

Among the countless AI startups I’ve seen recently, this is the largest and most ambitious company. It’s a Japanese company, so it’s somewhat hidden, but it’s incredibly impressive to an almost absurd level.

Initially, they created their own AI framework like TensorFlow, then expanded downward to libraries like CuDNN and upward to tools for optimizing AI model hyperparameters. Eventually, they built their own dedicated GPU cluster, and ultimately manufactured their own dedicated AI chips to construct an AI supercomputer cluster. The chip is also remarkable - it’s a 2018 chip with FP32 performance of 131 TFLOPS.

For reference, NVIDIA’s latest A100 has FP32 performance of 19 TFLOPS. The implementation using chiplets to attach four dies is also admirable… The cluster built with this became the officially recognized #1 in the global supercomputer Green500 rankings as of 2020. (Thinking about how those who boast about making AI chips in Korea talk big but haven’t produced anything properly working, and exactly zero have made proper training systems…)

After completely internalizing all the HW and SW stack of this infrastructure, they develop AI applications directly on their own dedicated data center.

All major Japanese corporations are attached as partners/customers.

For automobiles, there’s Toyota; for robots, FANUC; for consumer goods, KAO; for telecommunications, NTT. @_@ The direction and vision of the business they’re pursuing might be understandable, but seeing them actually accomplish all this made me flinch.

But why isn’t such a company well-known?

This startup is competing with Google at a comparable level (semiconductors - AI data center - AI framework - AI applications…).

Japan shouldn’t be underestimated.

With just the domestic market, they can achieve this scale, so they don’t need to be known externally.

All Korean AI startups combined don’t even come close to this scale.

Wow… wow… the world is vast and we must be humble… I can’t even sleep now..

Share this article

Found this helpful? Share it with your network

Join the Discussion

Share your thoughts and connect with other readers

댓글

GitHub 계정으로 로그인하여 댓글을 남겨보세요. 건설적인 의견과 질문을 환영합니다!

댓글을 불러오는 중...