אירועים והרצאות בפקולטה למדעי המחשב ע"ש הנרי ומרילין טאוב

The memory wall bottleneck is throttling the performance of data-intensive applications as the data transfer between the processing units (e.g., CPU, GPU cores) and the memory is significantly slower than the compute itself. Therefore, emerging digital processing-in-memory systems overcome this bottleneck by performing parallel bitwise logic within the memory arrays themselves. This enables a drastic reduction in data transfer for vectored operations since the same instruction may be repeated across the entire vector dimension, thereby leading to the effectiveness of PIM for applications that require high-throughput vectored arithmetic.

In this seminar, we explore the extension of these high-throughput bitwise operations to reliable large-scale applications. We begin by constructing high-throughput arithmetic for both fixed-point and floating-point numbers based on the supported parallel bitwise operations, incorporating a variety of techniques to convert the control flow of floating-point operations into data-flow. We continue by demonstrating the effectiveness of these parallel vectored arithmetic operations for the acceleration of large scale applications such as matrix multiplication and Fast Fourier Transform (FFT), and propose an end-to-end programming framework that automatically converts high-level python instructions to underlying bitwise operations. Lastly, we explore algorithmic techniques that aim to improve the reliability of the memory and logic, thereby compensating for soft errors.

Orian is a PhD student supervised by Prof. Shahar Kvatinsky.