Intel Turbo Boost

[crazymonkeyzero]

Had a quick question concerning turbo boost on intel chips.
I am aware turbo boost allows the cpu to jump up in frequency (up to a specified limit) automatically when within thermal limits and there is heavy load, however, is it safe to disable this and have the chip overclocked to run at it's max turbo frequency all the time, given I have appropriate cooling of course. (ie overclock the, 3930k @ 3.2 to run at 3.8ghz all the time) When I say "safe" I mean in the perspective of a workstation user, which implies the need for stability and data integrity. I have heard that overclocking can lead to data corruption, but since the chip at stock is designed by the manufacturer to run at that speed anyway is it really considered "overclocking" at all?

 

[Sheep221]

Only overclocking over turbo frequency is considered harmful. Although you don't need to do this specifically because the CPU is underclocking when idle and at least from my experience any slight load will trigger turbo to max, so it does not really change anything if you leave all settings on auto or you set it manually.

 

[Idontcare]

When I say "safe" I mean in the perspective of a workstation user, which implies the need for stability and data integrity. I have heard that overclocking can lead to data corruption, but since the chip at stock is designed by the manufacturer to run at that speed anyway is it really considered "overclocking" at all?

Within this context I do think it is absolutely safe. Intel and AMD must validate the circuits in every core as being fully functional at the max turbo speedbin at the TJmax for the chip.

If you are operating the chip below TJmax then your stability is only improved even more so.

Taking that verified chip and boosting all cores to the turbo clockspeed isn't going to take you over TJmax and it isn't going to suddenly make an already verified core unstable.

Where you enter no-mans land is when you take any core, let alone all of them at once, above the top-most binned turbo multiplier. Now you are taking the circuits to a place that Intel or AMD never checked to confirm them to operate correctly (regardless of voltage or temperature).

On a work computer I would feel comfortable operating all the cores at the max turbo bin, provided the temperatures were well below TJmax. But I wouldn't take my work computer above that.

 

[crazymonkeyzero]

Thanks for the info guys. Currently, when ever I run computations on my computer (typical load is 40-70% on all cores),the frequencies always jump all over the place. like from 3.2-3.5-3.8 and back down to 3.3 every couple seconds even though my temps are pretty good (always less than 60c on load) It's actually kind of annoying so now I think I'll just overclock to lock in the max turbo value.

 

[frozentundra123456]

Thanks for the info guys. Currently, when ever I run computations on my computer (typical load is 40-70% on all cores),the frequencies always jump all over the place. like from 3.2-3.5-3.8 and back down to 3.3 every couple seconds even though my temps are pretty good (always less than 60c on load) It's actually kind of annoying so now I think I'll just overclock to lock in the max turbo value.

If you are only running 40-70% load on the cpu, do you really need to overclock?

 

[crazymonkeyzero]

If you are only running 40-70% load on the cpu, do you really need to overclock?

Well if the computer can run at 3.8, why shouldn't it? If it runs at 3.2 half the time that's about 16% slower than it can be running. When you run computations 24/7 as a part of your job, saving time is a good thing, no?

And when you say only 40-70%, are you implying that this is low (typically its 60% for me on average)? I don't think any software runs all cores at 100% (especially new processors such as ivy/sandy), except prime 95, which is solely for benchmarking purposes, correct me if I'm wrong.

 

[GPz1100]

^^Why run at the higher speed if your load doesn't demand it. The whole beauty of speed step/turbo boost is to run the cpu at low speeds when under light loads/idle, and ramp up to max speed when the load requires more processing power. If you run it at high speed all the time, you're wasting electricity.

If you monitor your cpu speeds, you'll notice as soon as any load is applied, it ramps up. Download this and take a look.

http://www.intel.com/content/www/us...ology/turbo-boost/turbo-boost-technology.html

http://downloadcenter.intel.com/con...om/19105/eng/Setup.exe&lang=eng&Dwnldid=19105

 

[Saylick]

Well, the whole point of "turbo" is to improve the efficiency of how the chip manages its thermal budget. Theoretically, you extract the most performance out of a chip across all workloads by having the chip run at its rated TDP 100% of the time. Got a multi-threaded app? No problem, increase the frequency of all cores until the max TDP is reached. Got a single-threaded app? No problem, idle all cores but one and increase that one core's frequency until the max TDP is reached. So what's the tradeoff? Power efficiency. Power efficiency diminishes with increased clocks since you need to raise voltages to hit higher clocks.

Side comment: I've always wondered too if you can increase the CPU utilization just by reducing the frequency. From what I understand, I assume that the CPU works only as fast as it is fed. If you are telling it to run a simple loop without any complicated dependencies on prior calculations, you can effectively get the CPU to be executing every clock cycle (or 100% load, as Task Manager would report). In real life, what I imagine is that the CPU utilization is a measure of the ratio between the CPU frequency and the memory frequency. If the CPU is running faster than memory can feed it, the usage drops below 100% since the chip has to wait for the memory. If you run the CPU too slow, then the rate at which the chip can retire instructions becomes the deciding factor and you get 100% CPU utilization. Of course, when you factor in calculation dependencies, I'd guess that there will be cases where the CPU ends up waiting on itself instead of waiting on the memory. Please correct me if I am mistaken.

 

[Borealis7]

why not enjoy both? overclock the 4 turbo levels AND keep turbo on! a 3.4GHz IVB CPU becomes 4GHz when 1 core is active.

 

[Idontcare]

Side comment: I've always wondered too if you can increase the CPU utilization just by reducing the frequency. From what I understand, I assume that the CPU works only as fast as it is fed. If you are telling it to run a simple loop without any complicated dependencies on prior calculations, you can effectively get the CPU to be executing every clock cycle (or 100% load, as Task Manager would report). In real life, what I imagine is that the CPU utilization is a measure of the ratio between the CPU frequency and the memory frequency. If the CPU is running faster than memory can feed it, the usage drops below 100% since the chip has to wait for the memory. If you run the CPU too slow, then the rate at which the chip can retire instructions becomes the deciding factor and you get 100% CPU utilization. Of course, when you factor in calculation dependencies, I'd guess that there will be cases where the CPU ends up waiting on itself instead of waiting on the memory. Please correct me if I am mistaken.

This is true for some situations, but not true for others. And that is the problem.

If the stalls are coming from hard-drive paging or ram paging then lowering the clockspeed of the CPU will increase the utilization rate while at the same time keeping the total computation time the same.

I.e. a 20min job takes 20min whether you have a 4GHz or 3GHz CPU, but the 3GHz CPU has a higher utilization rate for those 20min.

However, if the <100% utilization is coming from other factors such as L2$ or L1$ stalls, or from actual stalls in the pipeline including branch mispredicts and so forth, then scaling the CPU clockspeed will not change the utilization rate. In those cases the utilization rate will remain the same regardless whether the CPU is running at 3GHz or 4GHz.

The easiest way to find out which it is for you is to keep the job the same but vary CPU clockspeed and note the job completion time and average utilization. Then you will know which is the case for your specific type of workload.

 

[VirtualLarry]

Where you enter no-mans land is when you take any core, let alone all of them at once, above the top-most binned turbo multiplier. Now you are taking the circuits to a place that Intel or AMD never checked to confirm them to operate correctly (regardless of voltage or temperature).

I thought that you said in my thread that Intel does test EVERY multiplier on the "K" chips, and assigned a VID to them, based on shmoo plots.

So which is it, IDC. Either Intel tests ALL the multis, and assigns VIDs for them, or they only test up to their max turbo, in which case, is the VID for multis above them just assigned a max value, or a default (untested) value, or what?

Please, resolve this apparent contradiction in your statements.

Edit: My thread:
http://forums.anandtech.com/showthread.php?t=2283583

 

[Idontcare]

I thought that you said in my thread that Intel does test EVERY multiplier on the "K" chips, and assigned a VID to them, based on shmoo plots.

So which is it, IDC. Either Intel tests ALL the multis, and assigns VIDs for them, or they only test up to their max turbo, in which case, is the VID for multis above them just assigned a max value, or a default (untested) value, or what?

Please, resolve this apparent contradiction in your statements.

Edit: My thread:
http://forums.anandtech.com/showthread.php?t=2283583

What's with the attitude? Did I shat in your wheaties or something?

It is clear you are not understanding something here, there is a critical difference between stability versus verification.

Verification is exhaustive, covers a range of instructions mixes that spans the entire ISA, and is iterative to ensure random fluctuations are not problematic.

Stability testing is just a spot-check on a subset of instructions taken from the overall ISA.

We OC'ing enthusiasts just do stability tests, LinX and Prime95 and the like, we do not do verification.

The VID's for multipliers above the top turbo bin are set for stability purposes, but they are not set for passing verification because they are not verification tested above the top speed bin.

Its not really clear to me what it is that you want to argue about but I generally make it a rule to avoid confrontations with people who are showing bruised egos (as your post above is indicating).

If you have a general, everyday, problem with my presence in these forums then let me know and I will simply refrain from interacting with you and your threads. If you have a specific issue with me because of something I have posted here in these forums then either lose the attitude and engage me over it in a respectful manner conducive of friendly conversation or continue onwards with this approach of yours and I'll just steer clear of you from now on.

If my posts in your specific thread are the problem then I can delete them, or if this is more just about publicly shaming or disgracing me in some fashion as a matter of catching me up in an apparent contradiction then let me be the first to say I've fvcked up more things in my life, and been more wrong about big huge things that actually matter, than you probably ever will, and it would be of no surprise to me if I was/am/are/will be dead wrong about anything and everything else I post here in these forums.

If the shame train needs to start with your thread and VIDs then let's just cut to the chase and I'll say I was dead wrong. The VIDs are randomly generated numbers that an intern at Intel cooked up in Excel and ported over to be hardcoded into the CPUs. Does that answer suit you better?

 

[VirtualLarry]

Where you enter no-mans land is when you take any core, let alone all of them at once, above the top-most binned turbo multiplier. Now you are taking the circuits to a place that Intel or AMD never checked to confirm them to operate correctly (regardless of voltage or temperature).

This above is the statement that I take issue with, given this other statement in the other thread:

Are you suggesting that Intel tests each CPU, and programs all of the multis up to 57x or whatever they go to, with a VID that will allow the chip to OC using the stock cooler?

Yes they do. But the chip will thermal throttle from hitting TJmax long before it gets to the 57x multiplier.

When I wrote "tests" in the other thread, I was implying Intel-style testing, thorough validation. Meaning that they validation-test up to max multi on the 'K' chips, to assign a VID.

You said that they did.

But then you said in this thread, that they don't test (ok, "validate") them past stock turbo multis.

My mind is somewhat like a computer. I accept statements from you as factual, based on the evidence of your intelligence and experience (which I value highly). But when I recieve two seemingly (in my mind) contradictory statements, I don't know how to process them, how to file them away in my brain.

So the real question I have is, does Intel actually do validation testing, on multis above stock turbo speeds, on 'K' chips. And if so, then would not running the CPUs at those multis (and thus pre-programmed VIDs), not be a problem, as long as cooling was taken care of?

If so, then would not this statement be the factually incorrect one?

Where you enter no-mans land is when you take any core, let alone all of them at once, above the top-most binned turbo multiplier. Now you are taking the circuits to a place that Intel or AMD never checked to confirm them to operate correctly (regardless of voltage or temperature).

Since Intel does (if the answer to the above question is "yes") do validation testing on higher than stock turbo multis.

Edit: Ah, but this:

The VID's for multipliers above the top turbo bin are set for stability purposes, but they are not set for passing verification because they are not verification tested above the top speed bin.

So if they don't do validation testing on 'K' chips above max stock turbo multis, then that would mean that this statement was the incorrect one:

Are you suggesting that Intel tests each CPU, and programs all of the multis up to 57x or whatever they go to, with a VID that will allow the chip to OC using the stock cooler?

Yes they do. But the chip will thermal throttle from hitting TJmax long before it gets to the 57x multiplier.

Edit: Or perhaps it was my error, in simply using the word "tests", and making the assumption that everyone knew that Intel did validation testing when they tested a chip's frequency and voltage.

Or possibly, they do different types of testing, in order to both verify correctness, as well as set the VIDs for all of the multis?

Edit: Just for clarification, I have nothing against you, IDC. Only with the facts at issue. I think that you are a thoughtful, respectful, and intelligent poster.

 

[Idontcare]

Ah, understood, I completely see where you are coming from now.

Chances are high that I used the wrong words somewhere in my posts, either in trying to be brief while getting a general point across or I was truly sloppy in my word choice and not only confused you but I gave an outright wrong statement in the process.

I will have to go back and read what I wrote to figure out which is the case, but I'm leaning towards the latter (sloppy word use on my behalf) at this point.

My bad :$, sorry for the confusion I've caused in the meantime. Let me dig back into it and see what I can do to set the record straight

 

[Ed1]

I doubt very much Intel would validation much past to tier speed . They may run some tests (random picking of chips) to see there not on edge of what chip can do .

For sure there not running rigorous testing on all higher multipliers that are available .

If they did you probably see "tested" OC as a services added to cost of standard K .

The Vid is probably just a value and the higher the multiplier the higher the vid goes, one reason its to high if you use auto .

The above is just guessing of course .

 

[crazymonkeyzero]

So how high do you think I could overclock a 3770k past 3.9ghz on all cores, while maintaining data integrity. I am under the impression since different chips behave slightly differently, thus Intel has probably put some extra headroom for stability? I mean most people OC 25-30% above stock. So do you think 5-10% is safe? That would be running an i7 3770k at around 4.2ghz.

 

[Saylick]

So how high do you think I could overclock a 3770k past 3.9ghz on all cores, while maintaining data integrity. I am under the impression since different chips behave slightly differently, thus Intel has probably put some extra headroom for stability? I mean most people OC 25-30% above stock. So do you think 5-10% is safe? That would be running an i7 3770k at around 4.2ghz.

Hm... Like you said, I would believe that data integrity depends on the chip quality and how hard you decide to push it.

Off the top of my head, what I'd do is run Intel Burn Test for the same amount of time as your expected computation duration and check to see if there are any inconsistent results. If there are none, then your chip should be stable (at least from a hardware perspective... code stability, that's another issue). You can run as many passes of it as you'd like until you feel comfortable that the chip will produce consist results.

 

[crazymonkeyzero]

Off the top of my head, what I'd do is run Intel Burn Test for the same amount of time as your expected computation duration and check to see if there are any inconsistent results. If there are none, then your chip should be stable (at least from a hardware perspective... code stability, that's another issue). You can run as many passes of it as you'd like until you feel comfortable that the chip will produce consist results.

Thanks! I will try the intel burn test once I get to overclocking.

Would running prime95 also be a good indicator?

 

[Saylick]

Thanks! I will try the intel burn test once I get to overclocking.

Would running prime95 also be a good indicator?

From my own experience with overclocking, I've had times where P95 would pass for 24 hours while Intel Burn Test failed after half an hour. My temperatures were also typically higher with IBT than with P95. However, I have read of situations where the reverse was true.

For IBT, if I recall correctly, you can tweak the amount of memory dedicated to IBT before you run it, so if you know the problem size, you can enter that into IBT as well. Of course, you can always go balls-to-the-wall and max out your RAM usage. As a warning, I recall experiencing extreme amounts of unresponsiveness/lag when I told IBT to use >95% of my RAM during the test. It is nothing to worry about though as it didn't crash; the CPU was just far too busy handling the calculations to be doing what I wanted it to do. You just gotta give it some time and remain patient.

Another thing I should point out though is that the stability benchmark portion of P95 has the option to use large or small FTT, as well as a blend of both. Small FTT will be more of a test on the cache system within the chip as it will be executing primarily out of cache while large FTT will stress the memory system more as the data is too large to fit in the cache, so a lot of RAM paging will be needed.

As for you, I say go with whatever is comfortable. Some people swear by the 24-hour P95 benchmark, and some are comfortable with 30 minutes of IBT, followed by 8 hours of P95, and then some games. You understand your situation far better than I can, so feel free to mix and match whatever you see fit. The only real recommendation I can give is to run a variety of tests so as leave nothing to chance. Focusing on a specific benchmark or test is worthless in my opinion. Running a variety of tests is perhaps the only real way to make sure you covered all the bases. And for that reason, I suggest IBT, P95, and Memtest86+.

 

[crazymonkeyzero]

Thank you for the information Saylick. I think I will try both haha. 24 hours of p95 and 24 hours of intel burn test. That will certainly be a safe bet.

 

[Phynaz]

So do you think 5-10% is safe?

No, it is not safe.

You don't overclock a computer that earns you your living. It is far, far far safer to just buy a faster one.

 

[Saylick]

No, it is not safe.

You don't overclock a computer that earns you your living. It is far, far far safer to just buy a faster one.

Yeah, I probably should have mentioned this in my last point as well...

Crazymonkey, if these calculations are vital to your paycheck then I agree with everyone else and not take the risk. The benefit of an extra 10 - 20% in speed could be offset entirely by 1 vital calculation, and by that I mean reputation, etc. By the nature of your posts, it sounded like you were just asking about a hypothetical situation so I figured I give my honest opinion on your options assuming you are already well aware of the consequences.

 

[crazymonkeyzero]

Yea, then I'll just leave all core at max turbo. This way It will guarantee reliability. I was just trying to test the water, and see what other members on the forum might say. Guess the general consensus is obvious.

 

[Dave3000]

No, it is not safe.

You don't overclock a computer that earns you your living. It is far, far far safer to just buy a faster one.

I agree with this. I don't just game on my PC even though I specifically built my PC for gaming, but I use it for my job as well, which has lots of calculations involved. I still have the Intel turbo boost enabled because it's an Intel feature that is part of the design of my i7 3820 and I don't consider having Intel turbo boost enabled overclocking since I'm running the turbo boost at stock settings (per core on auto). I'm running my CPU within manufactured specifications even with Intel turbo boost enabled as I'm not overclocking the turbo boosts. If I want more CPU speed in the future, I'll buy a faster CPU instead of overclocking.

 

[bononos]

Had a quick question concerning turbo boost on intel chips.
I am aware turbo boost allows the cpu to jump up in frequency (up to a specified limit) automatically when within thermal limits and there is heavy load, however, is it safe to disable this and have the chip overclocked to run at it's max turbo frequency all the time, given I have appropriate cooling of course. (ie overclock the, 3930k @ 3.2 to run at 3.8ghz all the time) When I say "safe" I mean in the perspective of a workstation user, which implies the need for stability and data integrity. I have heard that overclocking can lead to data corruption, but since the chip at stock is designed by the manufacturer to run at that speed anyway is it really considered "overclocking" at all?

I would think that its unsafe since the turbo speed is only meant for a single (or fewer than the total) core running while the rest of the cores are sleeping.