Consortium

Founded in 2015, UPMEM is a pioneering flabless semiconductor company headquartered in Grenoble, France. UPMEM is recognised for its innovative Processing-in-Memory (PIM) technology, which addresses the limitations of traditional compute-centric architectures, particularly for data-intensive applications.

UPMEM´s PIM solution integrates powerful computing capabilities directly into DRAM memory dies, significantly reducing the energy and cost associated with moving data between CPUs and memory. This architecture mitigates the high latency and limited bandwidth bottlenecks inherent in traditional systems. The company's flagship product, PIM DRAM, embodies this advancement by embedding multiple Data Processing Units (DPUs) within each DRAM chip, enabling parallel processing and efficient data handling.

In a typical UPMEM PIM server configuration, standard Xeon x88 servers are equipped with both traditional DIMMs and UPMEM PIM DIMMs. Each DDR4-2400 compliant PIM DIMM houses 16 PIM chips, each containing 8 DPUs, for a total of up to 2560 DPUs in a standard configuration. This configuration enables significant performance improvements and energy savings, making it highly suitable for applications in AI, big data analytics, genomics and more. UPMEM´s technology provides fully programmable, scalable solution without requiring changes to existing hardware architectures. The company provides a comprehensive software stack that enables developers to write DPU programs in C and use APIs in popular languages for their programming. This approach makes it easy to customize and optimize applications to take advantage of PIM technology.

The company has garnered attention and support, including a significant €7 million funding round in 2023 to scale its revolutionary PIM technology globally. Through this innovation, UPMEM is poised to transform data centers by dramatically improving computing efficiency and reducing overall energy consumption.

Within the SustainML project, UPMEM collaborates closely with the University of Kaiserlautern-Landau (RPTU) to develop energy-efficient hardware libraries. UPMEM´s primary focus is on an innovative PIM chip that integrates Ai capabilities directli into the memory architecture. UPMEM´s PIM-DRAM technology minimises data movement by processing data closer to the memory, enabling significant energy savings and performance enhancements.

The PIM AI chip, central to UPMEM´s contributions, features a novel AI Compute Fabric that includes a processor and its specialized AI units optimized for AI workloads. This fabric sits on a logic die stacked on top of the DRAM die, ensuring high-bandwidth access to DRAM banks while optimizing compute performance and flexibility. UPMEM has also separated the orchestration and system integration logic between the DRAM and logic die, enabling efficient management of AI computations and data flows for improved performance and energy efficiency.

To support the architectural innovation of the PIM AI chip, UPMEM has developed a dedicates simulator. Designed as an x86 program, on top of the pytorch machine learning library, this simulator provides detailed insight into execution times, power consumption and execution profiles. It enables comprehensive benchmarking of the PIM AI chip against industry standards. Within the SustainML framework, the simulator allows users to evaluate and benchmark different hardware solutions' performance and energy consumption for a given AI application. Schedules for release as open-source software in S2 2024, the simulator aims to democratize access to this advanced technology, enabling the global technical community to use it.

By integrating its cutting-edge PIM technology into the SustainML framework, UPMEM is helping to significantly reduce the carbon footprint of AI operations. Their work fosters sustainable growth in AI research and applications and positions SustainML as a leading solution for greener AI.