What Do We Do With 10^12 Transistors? The Case For Precision Timing

Google Tech Talks
February, 21 2008

ABSTRACT

It is will be too costly to design many of these chips at the polygon
or even gate level, so they must be highly programmable. Furthermore,
they should not just be FPGAs as we now know them because with that
many transistors, we should specialize more for power efficiency. I
envision FPGA-like chips where the computational elements combine CPUs
with more traditional FPGA-like fabrics.

For embedded real-time applications, which I argue will dominate, I
argue that the temporal behavior of these processors should be as easy
to analyze and control as their functional behavior.

I present a vision such a precision-timed (PRET) processor, which
incorporates a variety of techniques. At the ISA level, it provides
cycle-accurate timers, a predictable memory hierarchy based on
scratchpad memories, and an interleaved pipeline that provides
predictable, hardware-efficient concurrency. It will be programmed in
a C-like language that includes user-specified timing constraints and
concurrency, perhaps with synchronous semantics. Both compile- and
run-time checks will ensure the program meets timing constraints,
similar to array bounds checking.

Speaker: Stephen A. Edwards
Stephen A. Edwards received the B.S. degree in Electrical Engineering
from the California Institute of Technology in 1992, and the M.S. and
Ph.D degrees, also in Electrical Engineering, from the University of
California, Berkeley in 1994 and 1997 respectively. He is currently
an associate professor in the Computer Science Department of Columbia
University in New York, which he joined in 2001 after a three-year
stint with Synopsys, Inc., in Mountain View, California. His research
interests include embedded system design, domain-specific languages,
and compilers.

Category:

People & Blogs

Tags:

• google
• techtalks
• techtalk
• engedu
• talk
• talks
• googletechtalks
• education

[Films4You]

And I remember counting Clock Cycles on a Z80 CPU

[Films4You]

Von Neumann Bottle Neck is a term I haven't heeard for a long time...

The same as the 640K barrier..

Not taked about because no one wants to fix it.

[Dudex58]

can someone make a vid for me what capacitors do and actually show all the diferent ones and send me the best

[Dudex58]

wow how do u getto make a 51.30 vid?

[DaveLG526]

I was a little surprised that the thought this could help with parallel processing wasn't a bigger part of the discussion.

In addition the embedded world is very different than the area Intel plays in. Also FPGAs are pretty expensive for the embedded world.

I wonder what could be done with 100 simple CPUs?

[xb360]

I think it is silly to be talking about how to organize large number of processors, without talking about what the software is going to look like.

Even more interesting is what we would want to do with 10^12 transistors... Is excel or word going to be a good application for such hardware?

[rektide]

The general thesis is pretty flawed: You'd have to eliminate out of order processing, reduce superscalar execution to predictable conditions, & know where the data you are accessing is located ahead of time. If you want to enforce rigid constraints you just drag performance down to a slowest common denominator, its all missed opportunity for speed. Go to 42m, theres a 10s slide showing many of these dilemmas in bullet point form: the problem is caches & pipelines and packet switching.

[taylortails]

it's more interesting to search for 'tile', 'raw processors', 'exposed wire delay'.

[semiliteratedgod]

the guy has absolutely no idea what terms like "mathematically chaotic" mean.

Not speaking about the problem of algorithmic determination of whether a Turing machine will stop on empty input or not

and other basic stuff...

[semiliteratedgod]

a moron at google?