생성형 AI는 데이터 학습량이 많을수록 우수한 결과를 내기 때문에, 대량의 데이터를 다루면서 저전력·고속으로 작동하는 메모리가 필수적이다.
▲Prototype of the accelerator card 'AiMX' that further improves performance by connecting multiple GDDR6-AiM (Accelerator-in-Memory) (Photo: SK Hynix)
AiMX prototype first unveiled and demonstrated at the US AI Summit
10x faster response time and lower power consumption compared to GPU systems Since generative AI produces better results the more data it learns, memory that operates at low power and high speed while handling large amounts of data is essential.
SK Hynix announced on the 18th that it had unveiled and demonstrated for the first time a prototype of the AiMX, a generative AI accelerator card based on GDDR6-AiM, at the 'AI Hardware & Edge AI Summit (hereinafter referred to as AI Summit) 2023' held at the Marriott Santa Clara Hotel in California, USA for three days from the 12th (US time).
AiMX is an accelerator card product of SK hynix specialized for large-scale language models (LLM) such as ChatGPT using GDDR6-AiM chips. At this summit, SK hynix presented the slogan, 'Boost Your AI: Discover the Power of PIM with SK hynix's AiM', and introduced GDDR6-AiM and several of these products. First prototype of AiMX accelerator card with improved performance through connection
▲SK Hynix on-site booth at the 'AI Hardware & Edge AI Summit 2023' (Photo: SK Hynix)
At this event, SK Hynix demonstrated Meta’s generative AI, the ‘OPT (Open Pre-trained Transformer) 13B’ model, on a server system equipped with an AiMX prototype. The AiMX system equipped with a GDDR6-AiM chip is 10 times faster in response speed than a system equipped with a GPU, but consumes only 1/5 of the power.
This demonstration showed that it can achieve high performance compared to the latest accelerators, attracting the attention of global companies that provide artificial intelligence services.
Along with this, the company also held a session presentation on-site to specifically explain the advantages of AiMX. SK hynix Vice President Lim Eui-cheol (in charge of memory solutions) compared the performance of GPUs and AiMX under the topic of ‘Cost-Effective Generative AI Inference Acceleration using AiM’ and explained the future of next-generation intelligent memory semiconductors.
Vice President Lim said, “SK Hynix’s AiMX is a solution that offers high performance, low power, and cost savings compared to existing GPUs,” and added, “We will continue to develop memory technology that will lead the artificial intelligence era.”
.jpg)
