# Clock Speed Regressions

By: anon (spam.delete.delete.delete@this.this.this.spam.com), May 19, 2019 11:13 am
John H (john.delete@this.not.com) on May 19, 2019 7:30 am wrote:
> Curious - there was an AMD talk hinting that TSMC 7nm/5nm may actually show clock speed
> regression vs. prior nodes 'without substantial effort to overcome' that regression.
>
> What physics are at play that are limiting clock speed on these processes? Is it too small wires/transistors
> increasing resistances for the current needed for higher clocks/performance?
>

For simplification let's assume everything scales perfectly. Then if we scale length by s and area by s^2 the RC delay shouldn't change. R ~ length/area and C ~ area/thickness so ideally they'll cancel each other out. Of course everything is usually less than perfect so you can assume it's worse.

Either way though if transistor delay goes down but RC delay remains constant then at first scaling transistor delay by s will scale total delay by almost s but eventually the scaling will approach 1 as RC delay starts to dominate.

Now throw in the new and exciting problems like barrier sizes not scaling and taking up more of the cross-sectional area than the conductor itself and you've got yourself a regression.

Increasing V is out of the question because we can't afford even higher power consumption and higher I would affect reliability or necessitate larger wires which would reduce area scaling and eat up the last few percent of cost reduction the nodes offer. Decreasing R same problem. Decreasing C? Well you could try keeping the thickness the same since thinner barriers are a problem anyway but I'm not sure how well the FETs will work with even lower gate capacitance and reducing parasitic capacitances has never been easy.

tl;dr
wire delay is constant at best, at 7nm/5nm things are far from ideal -> regression if you don't try really hard

> Also, is it the same effect that 'killed Dennard scaling around 65nm' ?

No, that was due to the threshold voltage not scaling anymore.

>
> Thanks..
>

TopicPosted ByDate
Clock Speed RegressionsJohn H2019/05/19 07:30 AM
Clock Speed RegressionsMichael S2019/05/19 06:06 PM
Clock Speed Regressionsanon2019/05/19 11:13 AM
Clock Speed RegressionsDavid Hess2019/05/21 06:25 AM