Seoul, Korea, August 24, 2021 / AP / — recently, Samsung demonstrated its latest progress in in memory processing (PIM) technology at the hot chips 33 conference. As an important conference in the semiconductor industry, the hot chips 33 conference will have high-profile microprocessor and IC innovative products unveiled every year.
Samsung Semiconductor hbm-pim information released by Samsung includes that Samsung has successfully integrated PIM based high bandwidth memory (hbm-pim) into commercial accelerator system for the first time, and expanded the application scope of PIM to DRAM module and mobile memory, so as to accelerate the integration of memory and logic.
Samsung integrates hbm-pim into AI accelerator for the first time
In February this year, Samsung launched its first hbm-pim (aquabolt XL) to integrate AI processing capacity into Samsung hbm2 aquabolt to enhance the high-speed data processing capacity of supercomputers and AI applications. Subsequently, hbm-pim was tested in Xilinx virtex ultrascale + (alveo) AI accelerator, which improved the system performance by nearly 2.5 times and reduced the energy consumption by more than 60%.
“Hbm-pim is Samsung’s first AI customized memory solution tested in customer AI accelerator system, showing great commercial potential,” said Nam Sung Kim, senior vice president of DRAM products and technology of Samsung Electronics, “With the development of technology standardization, the application of this technology will be further expanded to hbm3 for next-generation supercomputers and AI applications, and even mobile memory for AI on devices and memory modules for data centers.”
“Xilinx has been cooperating with Samsung Electronics to provide high-performance solutions for data center, network and real-time signal processing applications starting from virtex ultrascale + HBM series. Recently, both sides have launched exciting versal HBM series products.” Arun varadarajan Rajagopal, senior director of Xilinx product planning department, said, “We are pleased to continue our cooperation with Samsung. We help evaluate the potential of hbm-pim system to achieve significant performance and energy efficiency improvement in AI applications.”
DRAM module driven by PIM
The accelerated DIMM (axdimm) can “process” in the DRAM module to reduce a large amount of data exchange between CPU and dram as much as possible, so as to improve the energy efficiency of AI accelerator system. Because the AI engine is embedded in the buffer chip, axdimm can process multiple memory permutations (DRAM chipsets) in parallel instead of accessing only one group at a time, which greatly improves the system performance and efficiency. Because the axdimm module can retain the overall dimensions of the traditional DIMM, it can be replaced directly without modifying the system. At present, axdimm is being tested on the customer’s server, which can provide about twice the performance in AI based recommended applications and reduce the energy consumption of the whole system by 40%.
Samsung Semiconductor axdimm “sap has been cooperating with Samsung in the application of in memory database (IMDb) for sap-hana.” Oliver rebholz, core research and innovation director of SAP Hana, said, “based on performance prediction and potential integration solutions, we expect in memory database management (imdbms) The performance of will be significantly improved, and more energy efficiency will be achieved through decomposition calculation on axdimm. SAP hopes to continue to cooperate with Samsung in this field. “
Mobile memory brings AI from the data center to the device
Samsung lpddr5-pim mobile memory can provide independent AI functions without connecting to the data center. Simulation tests show that lpddr5-pim can improve the performance by more than twice and reduce the energy consumption by more than 60% when used in speech recognition, translation and chat robots.
Stimulate ecosystem vitality
Samsung plans to expand its AI memory portfolio by working with other industry leaders to standardize the PIM platform in the first half of 2022. Samsung will also continue its efforts to cultivate a highly sound PIM ecosystem to ensure the wide applicability of the entire memory market.
*Actual performance may vary depending on the device and user environment
*The product pictures, model, data, function, performance, specification parameters, features and other product information (including but not limited to product advantages, components, performance, availability and capability) in this document are for reference only. Samsung may improve the above contents without notice or restriction. For specific information, please refer to the actual product Product manual. Unless otherwise specified, the data involved in this paper are Samsung’s internal test results, and the comparison involved in this paper is compared with Samsung products.