OC will shorten the hardware life span?

[Idontcare]

I assume these effects play on each other though? Like the lattice structure of ions created from the voltage also increases joule heating which then increases diffusion and so on.

Would it be possible to create a processor with this lattice structure already in place?

Yep, there is a cumulative effect from the interplay of all the various degradation mechanisms.

Process development is 1/3 hitting your electrical parameters at your target design rules (Ion/Ioff, IDDQ, etc and gate pitch), 1/3 creating a process integration that is manufacturable (yield entitlement), and 1/3 mitigating reliability and lifetime degrading mechanisms such that the device can survive 10yrs on the field.

(this is why AMD/IBm could show results of 45nm HKMG but they couldn't ship/implement it into production until 32nm - the yield and reliability of those 45nm devices was simply unacceptable and it literally took years for the engineers to address it all)

A process node won't "qualify" for transfer form the R&D team to the production team until it meets all three requirements (meets parametrics, minimum yield entitlement, and meets or exceeds reliability specs) and typically that last year of process development is entirely focused on improving the intrinsic reliability while increasing yield.

The challenge with yield enhancement is that increasing yield is easy if you don't mind trading off your lifetime reliability. But of course you can't do that, so a lot of time and money goes into that last year of a node's development to get the yields to entitlement (without cratering reliability) while also getting reliability up to meet the internal spec.

Is that 10 continuous years (24/7 at load? 24/7 at idle? some combination?), or 10 years of "normal" usage, say 8 hours per day or less, most of it idle? Just curious

You are referring to what is called the "duty cycle" and the duty cycle rating will be for 24x7 operation.

This is partly why we enthusiasts find so much excess headroom in our CPU's. Why we can undervolt at stock clocks by so much (because we aren't running them at TJmax while simultaneously undervolting) and why we can overclock and overvolt by so much without killing our chips in a week.

Still, that said, I wouldn't OC a work computer or even a home PC that I used for anything I actually care about or depend on. I've been bit by data corruption on more than one occasion and potential upside in terms of time saved with an OC'ed rig versus the potential downside in terms of unrecoverable losses is an unsatisfactory ratio for me now.

 

[BonzaiDuck]

Stock voltage for my 2600K at stock clockspeeds is 1.34V.

At TJmax and stock voltage my CPU is expected to last 10yrs. (I base this expectation on my personal experience, not with CPU's but with process development experience in the industry)

Raising the voltage from 1.34V to 1.5V is expected to reduced the operating lifetime by a factor of of <2.

Just as device physics (Arrhenius equation) gives us a rule of thumb that the operating lifetime is cut in half for every 10°C increase in operating temperatures, operating lifetime goes as the inverse of the 4th power of voltage.

(1.34^4)/(1.5^4) = 3.22/5.06 = 0.64 (that means the expected lifetime at 1.5V is 64% that of the expected lifetime at 1.34V)

I can understand the reluctance to believe me in this school of thought, I would too if I didn't do this for a living in my professional life. But that is the way it works out.

There is a reason you don't see a massive out-cry of 2600k/2500k owners that their chips have died at 5GHz, it is also the reason why the number crunchers at Intel decided that offering a performance tuning program for a mere $25 was money in the bank despite the liklihood that people would slam their processor with crazy Vcc after buying it.

Device physics is a killer for lifetime, but once Intel enabled thermal and current throttling it becomes nearly impossible to kill your processor by adding enough volts at high enough clocks without tripping that TJmax or simply not getting any higher clocks (fully saturated Ion, no more GHz regardless the Vcc increase).

Well, you've revised my thinking. It's easier to be reasonably cautious about what seems more uncertain, and "uncertain" describes my understanding of the physics. I apparently had a better grasp of how they plan and organize the development around the cost- and risk accounting, only confirmed through your "insider" understanding of it.

 

[Ferzerp]

Still, that said, I wouldn't OC a work computer or even a home PC that I used for anything I actually care about or depend on. I've been bit by data corruption on more than one occasion and potential upside in terms of time saved with an OC'ed rig versus the potential downside in terms of unrecoverable losses is an unsatisfactory ratio for me now.

So true.

Sorry to quote without any meaningful input, but this can't be said enough.

 

[taltamir]

Idontcare, are there any thermal degredation processes that occur if:
1. The processor is powered off.
2. Temperature is reasonably low (say, 60c)
3. Temperature does not fluctuate up and down?

I said earlier that no but one of the things you mentioned sounds like it might fit the bill.

 

[Idontcare]

Idontcare, are there any thermal degredation processes that occur if:
1. The processor is powered off.
2. Temperature is reasonably low (say, 60c)
3. Temperature does not fluctuate up and down?

I said earlier that no but one of the things you mentioned sounds like it might fit the bill.

There's the academic answer and then there's the practical answer.

Academically speaking I am compelled to point out that unless the IC is at absolute zero kelvin then all thermally activated processes (including the diffusion of dopants) are still occurring, albeit at ridiculously slow rates.

At a practical level, yeah, nothing is happening to the CPU under those conditions. It would take a thousand years for a dopant atom to move just a few nanometers, technically it is degrading but you wouldn't be alive to notice it happening.

(the same can be said of diamonds, which are metastable at atmosheric conditions here on Earth and as such they do not last forever)

 

[taltamir]

thanks idontcare.

 

[CTho9305]

Stock voltage for my 2600K at stock clockspeeds is 1.34V.

At TJmax and stock voltage my CPU is expected to last 10yrs. (I base this expectation on my personal experience, not with CPU's but with process development experience in the industry)

Raising the voltage from 1.34V to 1.5V is expected to reduced the operating lifetime by a factor of of <2.

Just as device physics (Arrhenius equation) gives us a rule of thumb that the operating lifetime is cut in half for every 10°C increase in operating temperatures, operating lifetime goes as the inverse of the 4th power of voltage.

(1.34^4)/(1.5^4) = 3.22/5.06 = 0.64 (that means the expected lifetime at 1.5V is 64% that of the expected lifetime at 1.34V)

That doesn't fit with my experience. If I understood device lifetime correctly, it seemed like there was a pretty severe cliff at some point - I would expect a devices to show excessive failure rates in well under a year at that level of overvolting. Now, the issue may be that "excessive" for me is something like 1%, but the failure rate seemed to go up really fast above a certain point. Your equation shows a usable lifetime at 1.7V (~30% of original)...that seems implausible. Where's the disconnect here?

 

[apathy_next2]

i would like to stress loss of data as well. i recently went through this. system was overclocked, during gaming it just shut off and would nor power back again. turns out mobo was dead, i had a raid setup and had not backed up recently, so i was really sol. i had to get an same excat mobo to get all my data back. if u are going to oc as people said treat it as if it would be np at all if ur comp died tonorrow

 

[BonzaiDuck]

. . . ummmm . . . .

. . . time to clone my wkstation and backup my server . . .

This system has been flawless now for 15 months, 24/7. Easy to become complacent, so . . .

 

[Idontcare]

That doesn't fit with my experience. If I understood device lifetime correctly, it seemed like there was a pretty severe cliff at some point - I would expect a devices to show excessive failure rates in well under a year at that level of overvolting. Now, the issue may be that "excessive" for me is something like 1%, but the failure rate seemed to go up really fast above a certain point. Your equation shows a usable lifetime at 1.7V (~30% of original)...that seems implausible. Where's the disconnect here?

The fourth power is simply a rule of thumb meant to be a catch-all for the bailiwick of degradation mechanisms that are voltage dependent and as such it is only going to give meaningful indications of changes in lifetime reliability when operating close to the stock voltage. I would not be comfortable using this rule of thumb to predict lifetime impact from overvolting to 1.7V (or 1.9V for suicide runs).

The actual equations have an exponential dependence with the applied electric field as well as an exponential dependence on the local temperature within the leakage conduction pathway. (Semiconductor Device Qualification)

There is a severe cliff, naturally, which is when the e-field approaches the intrinsic dielectric breakdown strength combined with the self-heating effect that transpires simultaneously in the area in which e-field induced leakage is manifesting.

In practice we have to take devices up above 2V to force voltage-induced failures within a meaningful timeline (<60days) when performing accelerated TDDB lifetime reliability testing.

 

[aigomorla]

actually.... idc your statement doesnt apply to the earily 45nm cpu's.

maybe SB has that kind of durability, but im fairly sure a wolfdale is a crystal vase in comparision to it.

So would a yorkie's and gulftowns.

Ha! We're you been, Aigo?! I agree about IDC's Sandy. You still up there in LA . . . wherewuzit . . . . near Arcadia?

mah... busy with work, and got impatient at the slow down rate of cpu's.
I'll be back with fury when haswell comes out.