[Acpc-l] Vortrag: Prof. Thomas Sterling, 20. 6. 2000

[Maria Cherry]

                 UNIVERSITAET WIEN INSTITUT FUER SOFTWAREWISSENSCHAFT
                                      gemeinsam mit
                  FWF-Projekt Spezialforschungsbereich F011 "AURORA"


               EINLADUNG ZU EINEM VORTRAG IM RAHMEN DES AURORA-KOLLOQUIUMS


               The MIND Chip Architecture: Towards the Next Generation 
                           of Processor-in-Memory Devices


                              Prof.  Thomas Sterling

                      Center for Advanced Computing Research, 
                   California Institute of Technology and High
                           Performance Computing Group, 
                          NASA Jet Propulsion Laboratory

	         
	         ZEIT:  Dienstag, 20. 6. 2000, 16.00 Uhr s.t.  
	           ORT:  Institut fuer Softwarewissenschaft
	              1090 Wien, Liechtensteinstrasse 22, 
	                      Seminarraum, Mezzanin


Abstract

The steady increase in integrated circuit device density, reflected if not 
driven by the
so-called Moore's Law, is making innovative classes of computer architecture 
practical,
which were infeasible only a short time ago.  Of equal importance, conventional 
processor
architecture techniques are incapable of optimally exploiting the availability 
of logic
gates per chip, also suggesting the need for alternate structures.  Examination 
of
effective computation per transistor over the recent microprocessor history 
demonstrates
declining efficiency even as overall performance increases.  Other trends as 
well
motivate new structures for exploiting the wealth of chip level resources.  
While device
density is increasing at a significant rate, microprocessor clock rate is 
increasing much
more slowly.  Even so, memory access rate is improving much more slowly than 
that.  The
implication is that the "memory wall", the gap between the data requirements of 
high
speed microprocessors, and the ability of memory chips to supply them, is 
worsening.
Ironically, DRAM memory density is increasing more rapidly than CMOS logic 
density,
requiring a diminishing number of memory modules per processor chip.  Finally, 
recent
advances in fabrication methodology has made practical the co-integration of 
both CMOS
logic and DRAM memory cells on the same semiconductor die.  Recognition of these 
trends
is motivating new classes of computer architecture that combine multiple system 
units on
a single die.  System on a Chip (SOC), SMP on a chip (SMPoC), and Processor in 
Memory
(PIM) including IRAM are all examples of this revolution in computer 
architecture.  But
of these, only PIM exploits the important property of DRAM that exposes enormous 
memory
bandwidth on-chip compared to that available off chip; on the order of 100 
Terabits per
second per chip for certain organizations.  In its most initial form, PIM places 
simple
ALUs across the memory stack row buffer making possible several Gops of 
performance per
PIM chip without the need for temporal locality and at order of magnitude 
reduced power
consumption than conventional systems.  Yet practical operational considerations 
limit
the usefulness of such basic PIM structures and their entry in the commercial 
market has
been slow.

The Gilgamesh (Billion Logic Gate Assemblies - Mesh) project is developing the 
MIND
(Memory, Intelligent, Network Devices) chip, an advanced scalable processor in 
memory
architecture for spaceborne and general robotic embedded applications.  It 
overcomes the
logical and efficiency limitations of earlier PIM devices by incorporating 
several
critical mechanisms not previously found in PIM.  Message driven computation 
employing
light weight Parcels permits direct memory chip to memory chip interaction.  A 
new
scalable address mapping scheme permits the direct manipulation of virtually 
addressed
objects in a distributed system.  A multithreaded task switching and management
capability provides overlapping use of parallel on-chip resources for high 
efficiency of
memory and I/O channels.  A distributed object-based computing model called
"macroservers" provides a logical framework for organization multiple MIND
processor/memory nodes to perform a single parallel task for high performance 
and large
data sets.  This seminar will present the technology opportunities and 
architecture
challenges defining the design tradeoff space of advanced PIM devices.  The 
architecture
approach and some detailed mechanisms will be described demonstrating the 
concept of the
MIND architecture.  The talk will conclude with a brief discussion of future 
work and
immediate plans.

                        ***************************
Im Anschluss an diesen Vortrag (ca.  17.00 ct) wird Prof.  Zima im
Forschungsprivatissimum einen Vortrag ueber "Macroservers" halten.