MASE: A Middleware Framework for Machine-Learning Accelerator System Exploration

Prof Yiren ZHAO’s talk on “MASE: A Middleware Framework for Machine-Learning Accelerator System Exploration” was a riveting exploration of cutting-edge developments in machine learning and hardware acceleration. With over 60 students in attendance, the talk showcased the immense interest in this field. Prof ZHAO presented the capabilities of MASE, highlighting its software stack’s efficiency in exploring the design space and achieving lossless sub-4-bit post-training quantization. The live demonstration of MASE’s proficiency in mapping language models onto many-FPGA systems, surpassing Nvidia GPUs in FPS/Watt performance, was particularly impressive. The talk left attendees inspired and eager to delve deeper into model efficiency and computational performance optimization.

趙一人教授的講座探索了機器學習和硬體加速的前沿發展, 內容引人入勝, 成功吸引超過60名學生參加，展現了學生對此領域的濃厚興趣。 趙教授介紹了MASE的能力，重點介紹了其軟體堆疊在探索設計空間和實現無損子4位後訓練量化方面的高效性。 展示了MASE在將語言模型映射到多個FPGA系統方面的熟練能力，超越了Nvidia GPU在FPS/Watt性能上的表現，給人留下了深刻印象。 會議期間不少學生提問及提出自己的見解，講座也成功推動及鼓勵更多學生深入研究模型效率和計算性能優化的領域。