Data Pump and Processing Architecture
Tuesday April 8, 2008
Hamerschlag Hall D-210
4:00 pm
Qian Yu
Carnegie Mellon University
Main memory accesses always experience long latency due to Memory/CPU speed
gap and demand-driven access. In many applications, certain specific kernels
exhibit deterministic data flow. This characteristic can be used to prefetch
data before they are processed and write results back to main memory when
necessary.
In this talk, I will present a parameterizable algorithm-specific
architecture -Data Pump and Processing Architecture (DPPA). This
architecture improves computing and data access performance of specific
data-intensive applications. It explicitly manages on-chip memory hierarchy
with dedicated software and minimum hardware mechanisms. In this
architecture, two dedicated processors separately control data transport
flow and computation flow. As a consequence, memory access and computation
will be parallelized ideally. I will also introduce the instruction set
architecture and programming model of DPPA, and briefly describe the
implementation.
Qian Yu is a post-doctor in the Computer Architecture Laboratory and SPIRAL
group at Carnegie Mellon University. She received her Ph.D. degree in
Electronic Engineering from Chinese Academy of Sciences in 2006 and spent
one year as a post-doctor in University of Illinois at Urbana-Champaign.
Qian's research interests include VLSI design for high performance computing
and multi-processor communication.
|