David Kanter (dkanter.delete@this.realworldtech.com) on June 21, 2011 12:19 pm wrote:
> After a brief hiatus due to travel for conferences and other matters, we
> are returning to our regularly scheduled reading with a new article.
>
> Enthusiasts and engineers know cooling is vital; it raises frequency and dramatically lowers power by reducing
> CPU or GPU temperatures. The world’s fastest supercomputer shows that thermal management can increase
> CPU performance/watt by 20% and 3D integration will make cooling more critical in the future.
>
> http://www.realworldtech.com/page.cfm?ArticleID=RWT062011114950
>
> As always, comments, discussion and feedback are encouraged.
>
> David

One assumption here is that "logic" means "CMOS". An inherent flaw in CMOS is the once-per-cycle transient short from Vcc to ground (the N-channel FET turn on faster than the P-channel FET can turn off). The more clock cycles per second then the more shorts and the higher power dissipation.

One way to avoid this problem is to use a different technology such as I²L or tunnel diodes. I worked with tunnel diodes in 1968 at 0.5GHz. One problem was that none of the test equipment worked at those speeds. In 1984 General Electric was selling 70GHz tunnel diode logic. In 2021 researchers are working at 100GHz to 2THz. This looks very promising. Especially since the speed * power product is 100,000 times that of CMOS. That means 1000 times the performance at 1% of the power.

I saw the engineers struggle to go from TTL to CMOS. At TI most engineers could not make the transition from TTL to I²L. (A transistor with multiple emitters is NATURAL and a transistor with multiple collectors is UNNATURAL). (Don't tell them about ULMs.)

Once a major manufacturer actually tries tunnel diode logic, all the questions about the problems with CMOS will just go away. It may take a couple of decades for engineers to think in threshold logic (or another generation). Neural networks will help. The "perceptrons" are just threshold gates.

Robert Pearson