A New Microarchitecture Paradigm for High-Performance General
Purpose Processor Synthesis
Tuesday September 4, 2007
Hamerschlag Hall D-210
4:30 pm
Eriko Nurvitadhi
Carnegie Mellon University
The complexity and intricacy of modern high-performance microprocessors
are beyond the reach of state-of-the-art high-level design and
synthesis. The goal of practical microprocessor synthesis may be better
served by rethinking fundamentally the underlying microarchitecture. We
observe that a high-performance microprocessor spends a great majority
of the time in exception free execution of a small, basic subset of the
ISA. It is only this core set of behaviors that decides the overall
performance of the microprocessor.
We propose a new microarchitecture paradigm comprising high-performance
primary datapaths for only exception-free execution of the common-case
instructions. By limiting the scope to only a highly regular set of
instruction behaviors, we expect that even the complexity of superscalar
out of-order execution can be brought within reach of template-based
high-level synthesis. Unsupported instruction behaviors are relegated
to a secondary microcoded datapath of the complete ISA behavior, without
bearing on the overall processor performance.
This talk will introduce this new microarchitecture paradigm we call
Chimera (a creature from mythology with the heads of a lion and a goat)
and show instruction profiling results to provide insights on design
simplification potentials provided by Chimera. Then, the talk will
report on our initial progress on automatically synthesizing the
secondary microcoded portion of the Chimera microarchitecture from a
graph-based ISA specification.
Eriko Nurvitadhi is a graduate student in the Electrical and Computer
Engineering Department at Carnegie Mellon. He received his BSs, BA, MS,
and MBA degrees from Oregon State University. His advisor is Prof. James
Hoe.
|