Samsung’s In-Memory processing, A Wider Range Of Applications

Samsung has shown off its latest advances with processing-in-memory (PIM) technology at the largest semiconductor conference. The company announced that it had successfully integrated its PIM-enabled High Bandwidth Memory (HBM-PIM) into a commercially available accelerator system, as well expanding PIM applications to include DRAM modules and mobile memory in order to accelerate the intersection of memory and logic.

First AI accelerator to integrate HBM-PIM
Earlier this year, Samsung launched an industry-first HBM-PIM (Aquabolt-XL) that integrates AI processing into HBM2 Aquabolt to enhance data processing in supercomputers and AI applications. With the HBM-PIM, performance of the Virtex Ultrascale+ (Alveo) AI accelerator was nearly 2.5X higher as well as energy consumption was more than 60% lower.

Samsung Electronics’ senior vice president of DRAM Product & Technology, Nam Sung Kim, said, “The standardization of the technology will open up a variety of applications, including HBM3 for next-generation supercomputers and AI applications, and even mobile memory for data center memory modules.”

The Virtex UltraScale+ HBM family of Xilinx products is designed for use in data centres, networking and real time signal processing applications. Xilinx is collaborating with Samsung on evaluating HBM-PIM systems for the possible benefits they could bring to AI applications,” said Arun Varadarajan Rajagopal, senior director, Product Planning at Xilinx.”

Powering DRAM modules with PIM
AI accelerator systems can be made more energy efficient using Acceleration DIMMs (AXDIMMs), which move processing to the DRAM module itself, thereby minimizing large data movements between the DRAM and the CPU.

Rather than accessing just one DRAM chip at a time, the AXDIMM can process multiple ranks (sets of memory chips) at the same time thanks to an AI engine built within the buffer chip. This greatly improves system performance and efficiency. With an AXDIMM, the traditional DIMM form factor can be maintained, so the module can be replaced with no system modifications needed.

“In the past few years, SAP has been working with Samsung on developing memory technologies so SAP HANA can run more efficiently and accelerate the database”, said Oliver Rebholz, SAP’s head of HANA core research and innovation.

Mobile memory technology that brings AI to the device
Samsung’s LPDDR5-PIM mobile memory technology can provide independent AI capabilities without data center connectivity. Simulation tests have shown that the LPDDR5-PIM can more than double performance while reducing energy usage by over 60% when used in applications such as voice recognition, translation and chatbot.

Enhancing the ecosystem
As part of its AI memory portfolio expansion, Samsung will work with other industry leaders to standardize its platform for the PIM by the first half of 2022.