13 ms·
Hmm. In a CMOS structure with a weakly turned off device, you'd need the effective off resistance to be more than 1/20th of the effective on for the logic value
by microarchitect 15y ago
Hmm. In a CMOS structure with a weakly turned off device, you'd need the effective off resistance to be more than 1/20th of the effective on for the logic value to degrade by roughly 5%. Did we have such leaky transistors for 130nm/90nm when the shift happened?
A quick calculation I did with a 90nm model for I have some parameters with vt=0.3, vdd=1.2, subthreshold slope=90mV/decade and dibl coeff=0.1 seems to suggest that leakage would still be < 1/100th of the on current.
- sliverstorm 15y agoI would guess the concerns are less about corrupting results, and more about leakage power and switching speed. If you have a very leaky pulldown, switching to "1" is going to be slower.
- microarchitect 15y agoI'm not convinced for the same reason I posted above. The on-current is at least an order of magnitude higher than the off-current for all technology nodes that I know of. The only way I see this having a significant effect is through the vt vs. speed trade-off I mentioned above. Lower vt's exponentially increase leakage power, so we push the vt up to combat leakage, but this reduces transistor speed because frequency is roughly proportional to the on-current which is roughly proportional to gate overdrive vdd - vt.
- mbell 15y agoI believe we're talking about two separate issues, maybe this will help: http://en.wikipedia.org/wiki/Drain_Induced_Barrier_Lowering http://en.wikipedia.org/wiki/Drain_Induced_Barrier_Lowering
- microarchitect 15y agoI don't think so. The effect of DIBL is an increase in leakage current with the drain to source voltage. The first-order model I was taught in school was that leakage current not considering DIBL used to be proportional to exp(vgs-vt) but thanks to DIBL leakage is now proportional to exp(vgs - vt + eta*vds). The eta here is the DIBL coefficient, which AFAIK is 0.1 or thereabouts.
- mbell 15y agoThe first-order model falls apart as channel length decreases.