Speed now & then

Paul Koning paulkoning at comcast.net
Wed Apr 11 13:40:24 CDT 2018



> On Apr 11, 2018, at 2:31 PM, Jecel Assumpcao Jr. via cctalk <cctalk at classiccmp.org> wrote:
> 
> Chuck Guzis wrote on Wed, 11 Apr 2018 11:09:23 -0700
>> I thought that Moore's "law" dealt only with the number of transistors
>> on a die.   Did Gordon also say something about performance?
> 
> That is correct. The observation that transistors would be faster and
> use less power as they became smaller is called "Dennard scaling" from
> 1974:
> 
> https://en.wikipedia.org/wiki/Dennard_scaling
> 
> This led to the MHz wars of the 1990s. Sadly, as the isolation barriers
> (the "O" in "MOS") became thinner and thinner we could no longer ignore
> leakage currents. In addition, going with lower voltages no longer was
> possible as we got closer and closer to the transistor threashold
> voltages. So we got stuck at 3 GHz or less.

For silicon VLSI logic circuits, yes, give or take a GHz or so.  CMOS can go very much higher in RF circuits.  And if you switch to different semiconductors, you can go higher still.  People have built oscillators that get close to a THz, but those are not logic circuits, let alone computers.

Propagation latencies also become an issue, though you can design fast computers with substantial propagation delays if you have to.  Cray did this decades ago; the basic stage delay in the CDC 6600 is 1/20th of the clock period, and the delay on chassis interconnects is 1/4 of the clock period.  In that design, all sorts of things are carefully paced to match the known propagation delays.  But most other computers were not designed with such complex techniques.

	paul




More information about the cctalk mailing list