Samsung Launches Open Innovation Contest for AXDIMM Technology
Your imagination could win you $60,000. Submissions due November 15th.
If you have an idea on where the future of near-memory processing should be headed, Samsung Memory wants to hear from you.
On September 30th, the South Korean memory giant kicked off its first-ever Open Innovation Contest for AXDIMM Technology. Held to foster collaboration between industry and academia, the contest invites professors and students from universities in South Korea and the U.S. to propose innovative applications for one of Samsung Memory's key areas of research focus: AXDIMM.
Short for "Acceleration DIMM," AXDIMM is a powerful yet flexible near-memory acceleration platform. It allows researchers and developers to implement various functions in memory modules, and can be used to accelerate memory-intensive applications such as deep learning recommendation and genomic computation.
DDR4-compatible and FPGA-based, AXDIMM features a smart buffer chip that not only functions as a traditional buffer chip for a high-capacity DIMM, but also provides Arm CPU cores and FPGA fabrics which can help accelerate various functions. The platform supports simultaneous access to each rank within a DIMM, which means that it is capable of multiplying the process bandwidth and improving the performance of memory-intensive applications.
Because AXDIMM processes data locally in the DIMM, it also enables reductions in energy consumption. These qualities and more make AXDIMM a perfect platform for near-memory processing – one that can be easily integrated with a host through the standard memory channel interface.
What could this look like in practice? A good example of a use case would be the integration of a deep learning recommendation model (DLRM) with AXDIMM, in which the FPGA is programmed to accelerate the embedding operations, thereby improving performance and energy efficiency.
Samsung has established an ambitious research agenda for AXDIMM that emphasizes two areas in particular: artificial intelligence and databases. Through this competition, the company hopes to use the AXDIMM architecture to develop and recommend hardware and software models for in-DRAM databases, recommendation systems, natural language processing, image processing and more.
The goal of the contest is to identify ways that AXDIMM could be used to create future value. Students who are interested in joining the competition should form teams of up to five people, including a professor, and submit their proposals by November 15th. Five finalists will be selected on December 15th, and will be provided with a virtual AXDIMM system to perform R&D on their project until July 15th, 2022. Proposals will be evaluated in each stage based on criteria including their level of innovation, the creativity of their use of AXDIMM, and the feasibility of their hardware and software architectures.
Each of the five finalists will receive a one-time R&D incentive of $10,000. Winning teams, announced on August 15th, 2022, will receive $50,000 for first place, $30,000 for second, and $10,000 for third.
More information, including submission instructions, can be found on the contest's website.