Course Review and Technology Predictions

The unifying principle of this course has been parallelism:

Circuits operate with gate-level parallelism. ASICs or FPGAs expose this directly.
Pipelining and superscalar provide instruction-level parallelism. These are programmer-transparent.
SIMD provides data parallelism. This is often exposed directly to programmers.
Multicore provides thread level parallelism. This is also exposed directly.
GPUs provide parallelism across both thread warps (SIMD), and thread blocks (multicore), but hide them both from programmers inside the "kernel" abstraction.
Networks allow distributed-memory parallelism. This is exposed directly to programmers.
Quantum computers use parallel universes. This is creepy, but can allow exponential speedups.
Future nanotechnology may allow molecular-level parallelism. It is not clear how to program these machines.

We live in interesting times--by the end of your career, it is clear that massive and radical changes will have occurred in both hardware and software. This is our chance to shape the future!

Technology Predictions

Silicon semiconductors have given us a very good run, but they're facing major challenges in the next few decades. The high switching voltage of silicon, about 0.6 volts, is a limiting factor in reducing energy consumption of modern chips. There are also serious obstacles to scaling semiconductor dopants below the few-nanometer (dozen-atom) regime, and maintaining reliability despite electromigration, especially at elevated temperatures.

We've radically switched computing technologies every few generations for the past 200 years, so it's quite possible for another radical transformation to occur.

Technology	Past	Present	Future	Issues
Electronic digital computer	Sequential computer (1950s)	Multicore (2000s)	Many-core?	Programming model is restrictive. Data movement problems such as energy and race conditions dominate performance.
Hydraulic computer	MONIAC, hydraulic automatic transmission (1940s)	Microfluidics	Nanofluidics?	Viscosity limits cycle rates to KHz range. Reactant recycling is difficult.
Mechanical computer	Babbage's Analytical Engine (1830s)	Keyless lock	Rod logic? Nanogears?	Viscosity determines operating rate. Wear limits operating lifetime. Routing mechanical power via line shafts is harder than routing electricity or fluids.
Electromechanical computer	Hollerith's tabulating machine (1890 census)	Piezoelectric transistors	Nano electro-mechanical systems?	Mechanical creep affects electrical properties.
Analog computer	Tube amplifiers, the gears of war	Probabilistic Processor	Quantum computers?	Noise and nonlinearity corrupt the computation results.
Networked computer	Mainframe & terminals (1970s)	Cloud server & phone apps	Skynet?	Latency, scalability, and reliability.

CS 441 Lecture Note, 2014, Dr. Orion Lawlor, UAF Computer Science Department.