Event Actions
Abstract:
The continuous increment in computing performance driven by Moore's law has not only shaped
our society, i.e., with sci-fi processing machines that fit in our pockets, a.k.a. mobile phones. It
has also been a significant factor in unprecedented scientific discovery, with high-performance
computing systems reaching the Exaflop mark. Yet, with the slowdown of Moore's economic
model and the power and memory walls, How do we go beyond Exascale while keeping power
consumption under control? Some promising techniques to answer these questions are (i)
extreme heterogeneity and specialization (accelerators), (ii) beyond silicon/CMOS technologies
such as superconductors, spintronics, photonics, or quantum computing, (iii) massive
parallelism, and (iv) alternative computing paradigms, such as stochastic or unary computing. In
this talk, I will share my work in some of these areas to meet future throughput and power goals
of computing.
First, I will introduce my work in superconducting digital computing, a promising technology for
high-performance computing that can reach at least 10X the speed and 600X better power
consumption than CMOS. More interestingly, superconducting circuits operate in cryogenic
environments (4K), opening up an avenue to bring quantum computing control/readout and
sensing/instrumentation inside the cryogenic refrigerator. Then, I will shift gears toward extreme
heterogeneity. I will introduce MoSAIC, a tool for cycle-accurate, fast exploration of
heterogeneous, message-driven computing architectures. Research on heterogeneous
architectures is usually based on models/simulators that (i) capture the dynamics of limited
components, (ii) approximate the cost of interaction between accelerators, memory, peripherals,
and host, and (iii) are orders of magnitude slower than running prototypes. To address this,
MoSAIC leverages the flexibility of FPGAs to enable fast/easy co-location of RISCV cores,
specialized accelerators, and distributed memory that communicate through a lightweight
network on a chip. A key feature of MoSAIC is the hardware message queues placed next to the
cores/accelerators to enable fast and inexpensive communication among tiles. During my talk, I
will show our ongoing work and examples of using MoSAIC as a research platform.
Bio:
Patricia Gonzalez-Guerrero is a Research Scientist in the Applied Mathematical and Computing
Research (AMCR) department at Lawrence Berkeley National Laboratory. Patricia received her
Ph.D. (2019) and M.Sc. (2015) from the University of Virginia, Charlottesville, VA, USA, and her
Bachelor's Degree from the Pontifical Xavierian University (2008), Bogota, Colombia. Her
research intersects non-conventional computing paradigms with emergent technologies from
circuits to architectures.
Host: Dr. Mircea Stan
Organizer: Cong Shen