Max Overdrive Supply Voltage in SoC 5nmFinFet

[Mister_hass]

Dear community,

Core devices in TSMC 5nm FinFet are rated for nominal supply voltage Vdd of 0.8V. For consumer applications lifetime/reliability matters, I believe they can be safely over-driven to 1V Vdd. I heard this morning that 5nmFF CPUs and GPUs can be over-driven to 1.25V Vdd for the highest performances. My knowledge of physics of deep-submicronic devices is that device is to blow up at this over-voltage, if not their lifetime is exponentially shortened to seconds/minutes.
I would appreciate if some one in the field can confirme the statement above (1.25V Vdd in 5nmFF). Otherwise what would be the max Vdd in 5nmFF for consumer applications (3 years lifetime)
Thank you very much.

 

[dick_freebird]

I would not take anyone's opinion besides the foundry's
reliability documentation (you want it in writing, yes, you
do). Especially since you seem to want a specific life
expectancy, only the foundry's guarantee is worth
anything. Other than maybe a very well conducted reliability
qualification exercise (sound like fun? It ain't, even when it
comes out as you wanted).

I find the numbers unconvincing. I have looked some at
GF 12/14LPP and their nominal is 0.8V with a max of 0.9V
or so in the "overdrive" case. Your numbers are a 50%
overvoltage. And this, at less than half the L I've seen.

Taking supply from 0.8 up to 1.25V is likely to at least
double the power dissipation, already a problem in many
CPU / GPU devices and stands to compound the gate ox
reliability problem with temperature acceleration.

Check for "weasel words" in any description of these
"overdrive" modes, regarding time-in-overdrive, reduced
case temperature limits, etc. Couching reliability info in
"use model" terms has been an increasingly popular play
in the consumer electronics segment but put one toe
over the line and you can be exposed to a lot of field-
return hurt (financial and reputational). Will you accept
the "overdrive" constraints, as a product use limit that
you have to force on your customers and theirs? Mine
aren't having it - they have flowed-down environment
requirements and wouldn't be interested in any opinion
of lesser authority.

 

[Mister_hass]

Dear community,

Core devices in TSMC 5nm FinFet are rated for nominal supply voltage Vdd of 0.8V. For consumer applications lifetime/reliability matters, I believe they can be safely over-driven to 1V Vdd. I heard this morning that 5nmFF CPUs and GPUs can be over-driven to 1.25V Vdd for the highest performances. My knowledge of physics of deep-submicronic devices is that device is to blow up at this over-voltage, if not their lifetime is exponentially shortened to seconds/minutes.
I would appreciate if some one in the field can confirme the statement above (1.25V Vdd in 5nmFF). Otherwise what would be the max Vdd in 5nmFF for consumer applications (3 years lifetime)
Thank you very much.

I found an article where in fact Overclockers can overdrive a 7nm AMD Ryzen 9 series CPU to 1.2V. This obtained in binning parts that can withstand that high voltage
https://wccftech.com/amd-ryzen-9-3950x-16-core-pre-binned-cpus-available-silicon-lottery/

www.siliconlottery.com is specialist of binning CPUs for overclocking and overdriving

 

[ThisIsNotSam]

I would think 1.2 is really pushing it. But one thing worth mentioning is that just because you supply 1.2V externally, it doesn't mean the transistors actually get 1.2V. Maybe the degradation is actually helping you to operate within bounds (IR drop).