Research

Research Highlights

EE Prof. Myoungsoo Jung and Panmnesia joint research team showcases Innovation Award-winning CXL-Enabled AI Accelerator at CES 2024

[EE Prof. Myoungsoo Jung and Panmnesia joint research team showcases Innovation Award-winning CXL-Enabled AI Accelerator at CES 2024]

 

Inline image 2024 01 17 10.54.55.413

<(from left) CXL-Enabled AI Accelerator Image & System>

 

Professor Myoungsoo Jung and Panmnesia’s joint research team showcased the Innovation Award-winning ‘CXL-Enabled AI Accelerator’ in world’s premier IT industry event CES 2024.
 
The award-winning accelerator features virtually limitless memory capacity that brings significant performance boost for large-scale AI-driven services. This innovative feature empowers AI-driven services to handle increasing amount of data, leading to substantial enhancements in accuracy and quality of the services.
 
In the evaluation using a representative large-scale AI service, AI-powered image search, the system with their accelerator exhibited 101 times faster performance than conventional SSD-based AI acceleration system.
 

In order to provide better service to users, AI services from global tech giants such as Google and Microsoft are rapidly increasing the amount of data their AI services are handling. However, conventional accelerators have limited memory capacity which makes it hard to load the dataset into their memory. Thus, to manage the large dataset, the conventional systems falls back to employ SSD, whose long latency significantly degrades the overall performance.

 

To this end, the joint development team showcased the CXL-enabled AI Accelerator that addresses the above problem. The accelerator adopts Compute eXpress Link (CXL), a cache-coherent high-speed interconnect, to connect itself with the system, Such interconnect allows the system access the accelerator’s internal memory.

 

This feature allows the system to expand its memory space by connecting multiple accelerators to the CPU through CXL switch. In addition, the joint development team further improved the overall system performance by employing specialized hardware acceleration module for parallel processing of AI-based image search inside the accelerator.

 

The joint development team showcased their outcomes at CES 2024, held in early Jan ‘24 at Las Vegas, US. The accelerator bestowed the CES 2024 innovation award for their technical superiority. The news of the CXL-Enabled AI Accelerator winning the innovation award and exhibiting at CES was reported by international media, including IEEE Spectrum, TechRadar, Storage Newsletter, and Design and Reuse.
 
The news was also widely covered by multiple domestic media, including Chosun-Ilbo, The Korea Economic Daily, Seoul Economic Daily, ChosunBiz, and ZDNet Korea. 
 
“We had an exclusive meeting with ARM, the world’s largest IP company, and confirmed their interest in our technology,” said Panmnesia’s exhibitor. “HPE (Hewlett Packard Enterprise), a global cloud and data center company, is also very interested in our CXL technology, and we’ve been invited to HPE headquarters for an exclusive meeting later this month,” he added. 
 
Panmnesia is a fabless startup with technological leadership in CXL Intellectual Property (IP). Panmnesia was born out of the collaborative efforts of a talented group of Ph.D. graduates from KAIST EE. They gained exposure from their recent $12.5 million seed funding.
More information on Panmnesia can be found in its website(http://panmnesia.com). More information on Innovation Award-winning accelerator can be found in Panmnesia’s Youtube video (https://youtu.be/ZujXVEr3nO0)
 
 
Inline image 2024 01 17 10.55.36.861
<Exhibition Hall of KAIST Research Team>
 
 
*Link to press coverage on this news: https://www.etoday.co.kr/news/view/2321263