Groundbreaking Announcement

In a groundbreaking announcement from Barcelona, Spain, dated December 9, 2024, Semidynamics has been selected by UPMEM to provide vital IP technology for their upcoming LPDDR5X Processing In Memory (PIM) device. This collaboration is set to transform how large language models (LLMs) and AI applications operate on mobile devices.

About UPMEM and Semidynamics

UPMEM, a trailblazer in the semiconductor industry, specializes in innovative PIM solutions. By integrating co-processors directly into main DRAM memory, UPMEM is significantly enhancing data-intensive operations such as those seen in generative AI applications. The newly selected Semidynamics IP features a standardized RISC-V architecture enhanced with an integrated Tensor Unit and a sophisticated latency optimization tool called Gazillion. This robust architecture enables the seamless and efficient integration of cutting-edge AI models, providing unmatched performance.

Impressive Technology Performance

What's particularly staggering about this technology is its impressive internal bandwidth, which reaches a phenomenal 102.4GB/s, combined with a remarkably low energy consumption of 1pJ/bit. The processing capabilities allow for up to 8 TFLOPs FP16/BF16 and an astounding 16 TOPs int8. This means that smartphones utilizing this technology will be able to perform model inference with unparalleled efficiency—setting a new standard in the industry.

Quotes from Leadership

Gilles Hamou, the CEO of UPMEM, expressed enthusiasm for the alliance, stating, "SemiDynamics GenAI compute IP combines the power and efficiency needed for our disruptive Processing In Memory DRAM chips for mobile. The integration of RISC-V architecture is crucial, as it positions us at the forefront of the evolving ecosystem. Our joint technologies are uniquely capable of delivering the energy efficiency and cost-effectiveness needed to process generative AI computations right from smartphones."

Roger Espasa, CEO of Semidynamics, echoed this optimism: "We are thrilled to support UPMEM with their pioneering Process In Memory approach, which is set to redefine the capabilities of deploying large language AI models. Our partnership promises a substantial leap forward in mobile technology, paving the way for exciting advancements in generative AI applications."

Efficiency and Performance Breakthroughs

With the ability to preprocess and analyze AI models locally, UPMEM’s next-generation PIM chips can run extensive models such as LLama and Mistral more efficiently than current offerings—achieving processing speeds capable of handling 15 times more queries per second while consuming a staggering 90% less energy compared to leading System on Chip (SoC) solutions. This innovation not only prioritizes performance but also addresses the essential need for battery longevity in modern smartphones.

Future Outlook

Located in Grenoble, France, a hotspot for microelectronics and technological advancements, UPMEM is primed to revolutionize how generative AI tasks are executed on mobile devices. This pivotal partnership with Semidynamics is expected to attract significant attention from the tech community as it signifies a leap into the future of smartphone capability.

Conclusion

Stay tuned as these two industry leaders embark on a journey that promises to reshape our digital experiences!