Rebellions' AI chip (Courtesy of Rebellions) Rebellions Inc., a South Korean artificial intelligence chip design startup, beat global chip giants Qualcomm Technologies Inc. and Nvidia Corp. in the latest AI chip performance tests by an internationally renowned machine learning engineering organization, raising hopes for the future of the country’s memory chip-centered semiconductor industry.
According to the latest results of the industry-standard MLPerf Inference v3.0 benchmark suites announced on Wednesday, Rebellions’ ATOM chip performed the best, showing off the fastest speed in processing large language and vision models.
The Korean AI chip’s processing speed was 1.8 times and up to 2 times faster than the same class chips by Qualcomm and Nvidia, respectively, in a single stream latency test in the large language model area by MLPerf Inference v3.0 test. This means that ChatGPT run on an ATOM-fitted computer delivers the quickest answers to its users.
The Rebellions’ AI system-on-chip on its original ION core architecture also showed 1.4 times and 3 times faster processing performance in the vision area than the chips of Qualcomm and Nvidia, respectively, topping the test.
Rebellions is the first Korean AI chip developer to score top in both language processing tests by MLPerf, said its Chief Product Officer Kim Hyo-eun.
Its win follows the victory of two other Korean fabless startups -- FuriosaAI Inc. and SAPEON Inc. -- in the previous MLPerf Inference tests in 2021 and 2022, respectively, raising hopes that Korea, home to two global memory chip giants Samsung Electronics Co. and SK Hynix Inc., will be able to diversify its semiconductor industry beyond memory chips.
GLOBAL NODS TO KOREAN AI CHIPS
MLPerf is a benchmark suite launched in 2018 as a comprehensive set of system tests to measure machine learning performance to define and run AI performance metrics. It releases a new round of benchmarks for AI Inference processing once every six months.
It is run by MLCommons, a leading open AI engineering consortium founded in the same year to accelerate machine learning innovation by some 50-plus members of companies, academics and non-profit organizations from around the world, including Harvard University, Google and Qualcomm, Samsung Electronics and more.
Rebellions debuted its first performance results in the latest test round, according to MLCommons.
Korea needs to foster novel chip technologies beyond memory chips and nurture experts like Rebellions to stay competitive in the global semiconductor market, industry observers said.
Rebellions was co-founded in September 2020 by Chief Executive Park Sung-hyun, a former quant developer at global investment bank Morgan Stanley and Elon Musk’s aerospace company SpaceX, and Chief Technology Officer Oh Jin-wook who previously designed AI chips at IBM. CPO Kim was a former deep learning technology developer at AI medical device maker Lunit.
SAPEON started as an in-house venture project at SK Telecom and was spun off as a separate entity in 2021, headed by CEO Ryu Soojung, a former professor at Seoul National University leading research on NPU and PIM and ex-vice president of Samsung’s R&D center overseeing GPU projects.
SAPEON’s neural processing unit (NPU) (Courtesy of SAPEON)
After getting the nod for their AI chips’ performance from global top-tier peers, these novice Korean fabless startups should prove their technological prowess with the seamless mass production of their chips, a move that will pave the way for the development of the Korean non-memory chip industry.
Rebellions plans to kick off the mass production of 5-nanometer ATOM chips with Samsung Electronics’ foundry early next year, while Taiwan Semiconductor Manufacturing Company Ltd. already began commercial production of FuriosaAI’s chips last year.
Write to Joo-Wan Kim at kjwan@hankyung.com Sookyung Seo edited this article.
We use cookies to provide the best user experience. By continuing to browse this website, you will be considered to accept cookies. Please review our Privacy Policy to learn our cookie policy.