Macro Architectural Trends; IC Fabrication
Tom Kelliher, CS 220
Sept. 11, 2009
Read 2.1-2.6.
First homework will be posted later today.
Performance.
The power equation:
- Capacitance: What is it, what contributes to it? Trends.
- Voltage: Trends and limits.
- Clock Rate: Trends.
``Sea change'' from uniprocessing to multiprocessing:
- Easy to increase uniprocessor performance if you can constantly
overclock.
- Explicit parallelism -- the ``third rail'' of CA.
- Explicit parallelism -- we've been here before: supercomputers.
- What's the future of this trend?
How did we get here?
- The ``Clock'' wars.
- Implicit parallelism -- ILP:
- Deep pipelining, branch prediction, speculative execution.
- Trace caches.
- Multiple issue.
- Register renaming, out of order execution.
The cost of high clock rates and all this complexity: power.
- Heat dissipation in commodity microprocessors.
We've been here before: the CISC/RISC crossroads
- CA was more art than science -- few quantitative studies.
- Trends:
- Move software functionality into hardware -- more and more HLL
support.
More ``powerful'' instructions are not always faster -- the 11/780
INDEX instruction.
- Historically, memory bandwidth had to be conserved.
Achieved by:
- Dense, complex instruction sets.
- Microcode techniques.
- Feature creep.
- Consequences:
- Faster memory hierarchy.
- ``Irrational'' implementations.
- Increased design time.
- Increased design errors.
- Faster memory and VLSI technology broke the complexity trend.
Moral of the story: re-evaluate your assumptions on a regular basis.
(Additional background: See ``The Case for the Reduced Instruction Set
Computer'' on the web site.
A little more background...
- Die, package details.
Core 2 Duo ``Allendale'' die size:
, 65 W, 775 pin LGA
- Die photomicrograph.
- All features/layers are added in separate steps...
- It all begins with design, either standard cell, schematic, VHDL, or
hand-layout, resulting in a set of masks.
Thomas P. Kelliher
2009-09-11
Tom Kelliher