Samsung Electronics is expanding its lineup of memory semiconductor products equipped with AI engines. Samsung Electronics plans to lead the convergence of memory and system semiconductors and rapidly expand the next-generation memory semiconductor ecosystem through collaboration with various global companies.

At the Hot Chips conference held online on the 24th, Samsung Electronics introduced not only the HBM-PIM, which it developed for the first time in the industry in February, but also various product groups and application cases that apply PIM technology.

At the conference held that day, Samsung Electronics introduced 'AXDIMM (Acceleration DIMM)', which has an AI engine installed in a DRAM module, 'LPDDR5-PIM' technology, which combines mobile DRAM and PIM, and actual system application cases of HBM-PIM.

AXDIMM is a product that expands PIM technology from chip unit to module unit and equips an artificial intelligence engine on a DRAM module.

By installing an AI engine in each Rank, which is the operating unit of the DRAM module, parallel processing is maximized to increase performance, and the AI engine enables calculations within the DRAM module, reducing data movement between the CPU and DRAM module, thereby increasing the energy efficiency of the AI accelerator system.

Additionally, it can be applied without changing the existing system by installing the AI engine in the buffer chip mounted on the existing DRAM module.

Currently, AXDIMM is undergoing performance evaluations in server environments of global customers, and performance has been confirmed to have increased by approximately two times and system energy has been reduced by more than 40%.

Samsung Electronics unveiled its 'LPDDR5-PIM' technology, which expands the application of PIM not only to the ultra-high-speed data analysis field but also to the mobile field.

When PIM technology is combined with mobile DRAM, it is expected to effectively solve the performance and energy issues of On-Device AI, which can perform AI functions independently on mobile phones without connecting to a data center.As a result of the experiment, performance was improved by more than 2x and energy was reduced by more than 60% in speech recognition, translation, and chatbots.

Additionally, Samsung Electronics announced the verification results of installing HBM-PIM, which was released in February, in an actual system at the Hot Chips conference.

Xilinx, an American FPGA developer, drew great attention from attendees by announcing that when HBM-PIM is installed in an AI accelerator system that is already being commercialized, performance is increased by about 2.5 times compared to a system using the existing HBM2, and system energy is reduced by more than 60%.

Samsung Electronics emphasized that by incorporating the innovative technology called PIM into the DRAM process, actual performance and energy efficiency have been significantly improved in various applications, and that PIM technology will become a new memory paradigm in the big data era.

“HBM-PIM is the industry’s first AI-specific memory solution, and has already been evaluated by customers in their AI accelerators, demonstrating potential for commercial success,” said Kim Nam-seung, senior vice president of DRAM Development at Samsung Electronics’ Memory Business Division. “In the future, we plan to expand it to HBM3 for next-generation supercomputers and AI, mobile memory for on-device AI, and DRAM modules for data centers through the standardization process.”

“Xilinx has been working with Samsung to support high-performance solutions for data center, networking, and real-time signal processing applications, starting with our Virtex UltraScale+ HBM family, and recently introduced new and exciting Versal HBM series products,” said Arun Varadarajan Rajagopal, senior director of product planning at Xilinx. “We look forward to continuing our collaboration with Samsung on system evaluations to enhance the performance and energy efficiency of HBM-PIM in AI applications.”.

Oliver Rebholz, head of HANA core research & innovation at SAP, an enterprise software company, said, “We expect performance improvements and high energy efficiency in performance prediction evaluations of systems utilizing AXDIMM,” and “SAP will continue to work with Samsung Electronics to improve the performance of its in-memory database platform, ‘SAP-HANA.’”

Samsung Electronics plans to expand its AI memory market by completing the standardization of its PIM technology platform and ecosystem for AI accelerators for various customers within the first half of next year.