Kakao Brain to introduce Google¡¯s TPU to upgrade its hyperscale AI model

2021.11.26 11:42:25 | 2021.11.26 11:42:54

À̹ÌÁö È®´ë
Kakao Corp. will install Google¡¯s super computing system to improve its Korean language artificial intelligence (AI) model KoGPT developed by Kakao Brain, its AI research and development subsidiary, so that it can catch up its rival Naver that has already unveiled a faster hyperscale AI platform.

Kakao Brain announced Thursday that it will increase KoGPT¡¯s capacity by more than 100 times by using Tensor Processing Unit (TPU), Google¡¯s custom-built AI chip which accelerates machine learning efficiency.

¡°It is the latest tool developed by Google and has a capacity of 1 exaflop, meaning that it is capable of carrying out 10 quadrillion computations per second,¡± according to Kakao. ¡°It is the most advanced deep learning super computing infrastructure introduced by Korean companies.¡±

Kakao¡¯s move is expected to intensify AI competition by Korean tech companies. In October last year, its rival Naver adopted NVIDIA¡¯s AI supercomputer Superpod for the first time in Korea for its hyperscale AI platform Hyper Clova. Superpod has a 700 petaflops capacity, capable of processing 1,000 trillion computations per second.

Google claims that its TPU is much better in performance and energy efficiency than graphic processing unit (GPU) that is used to enhance graphics and AI workloads. Kakao has chosen Google¡¯s tool to take over Naver in the hyperscale AI competition.

Naver unveiled in May the country¡¯s first hyperscale Korean language AI platform Hyper Clova. The system has a capacity of 204 billion machine learning parameters, the largest in the country so far, compared with Kakao Brain¡¯s KoGPT with 6 billion parameters. Kakao aims to increase KoGPT¡¯s capacity by 100 times with Google¡¯s TPU.

LG and KT have also joined the fierce race to develop hyperscale AI platforms. LG¡¯s AI research institute plans to unveil its hyperscale AI with 600 billion parameters in the latter half, while KT has been developing its hyperscale AI model with NVIDA¡¯s supercomputing tool with an aim to commercialize it in the first half of next year.

By Lim Young-shin and Choi Mira

[¨Ï Pulse by Maeil Business Newspaper & mk.co.kr, All rights reserved]