"7 Watt" Ivy Bridge my arse!

[Ferzerp]

So you're saying that better cooling alone will decrease power consumption for the same workload.

But then you said you don't know what workload actually measures, that it could be anything. If what you're saying is true then it's whatever the same workload 10W cTDP is. Now I have no idea what that represents, but cTDP isn't really new so maybe there's some information out there. I expect Intel should still be able to run the CPU cores at its base clock at least, under "normal" programs that fully load it (ie not doing weird power virusy things). And with the GPU only doing minimal GUI stuff.

The problem is that nowhere do I see that Intel is clearly describing this. Maybe to you it's implied by the slide, but I think there's still some room for ambiguity. 10W to 7W is a huge reduction just from 25 degrees C at TJ, which I imagine is nowhere close to saying a 25C reduction in ambient; the temperature of the cooling elements on the tablet are going to be nowhere close to 25C over the air around you.

The savings should be less dramatic if already starting at lower voltage levels, especially given that Intel's 22nm voltage/clock curve rises more steeply than on its older processors. For your graph you have things fixed at a high 1.29V which I'm sure exaggerates the results.

Where are you arriving at the assumption that it is anything less than 100% theoretical load at 80C? I didn't see this other than information introduced by posters trying to criticize the idea.

 

[Exophase]

Where are you arriving at the assumption that it is anything less than 100% theoretical load at 80C? I didn't see this other than information introduced by posters trying to criticize the idea.

I'm NOT assuming that. It's others who are assuming that it necessarily can do 100% theoretical load (relative to 10W at Tjmax) at 80C Tj consuming 7W. I'm saying it isn't clear to me that this is being guaranteed by Intel.

 

[exar333]

Here's a blast from the past:

Different article from a different time/market.

Intel is aggressively pursuing the mobile and tablet arena. If they have a fantastically efficient 3ghz 2C4T with turbo, they can make a number of lower-power variants. Whether they cut HT, lower the clockspeed, or cut the TJMax to a specific level. It's only going to get more confusing because Intel is moving to a more flexible 'TDP' at lower levels with Haswell.

2C/4T @ 3ghz = 17w
2C/2T @ 2.4ghz = 11w
2C/2T @ 2ghz = 9w
2C/2T @ 2ghz w/ low TJMax = 7w

MB makers could possibly find a way to override the TJMax or allow the CPU to operate above the designed 'TDP'. In this case, you could exceed the marketed wattage ratings by doing so. OEMs will have a lot of responsibility here to (1) design with adequate cooling and (2) offer a compelling product.

It is all about giving OEMs the tools to create a device that offers enough performance while giving plenty of battery life. In the end, this is what matters.

If I am using a '11w TDP' laptop with similar computing power as a different '7w TDP' laptop, and the former gets more battery life, which device would you choose? I don't care about the TDP for marketing, I care about what device actually does.

TDP matters more to me on the Desktop and in the server arena.

 

[NTMBK]

Where are you arriving at the assumption that it is anything less than 100% theoretical load at 80C? I didn't see this other than information introduced by posters trying to criticize the idea.

Assuming anything about this is bad, because Intel have given us no indication of what this actually means. Until they tell us how SDP is defined, it is completely and utterly useless.

 

[NTMBK]

Okay, update from Intel! Have a read of this:

Today, Intel confirmed to us that though chips like the new Core i7-3689Y run at their advertised 1.5GHz in a system with 13 watts worth of cooling, and turbo up to 2.6GHz when things stay cool enough to do so, they're much slower at stock clocks. At 7 watts, Intel told us, the Core i7-3689Y runs at only 800MHz by default. Here's a crucial note, though: Intel tells us that the processor can still turbo up to that same 2.6GHz when the cooling system has available headroom.

http://www.theverge.com/2013/1/9/3856050/intel-candid-explains-misleading-7w-ivy-bridge-marketing (Emphasis mine)

So yeah, big surprise, they sacrifice a lot of performance to get it to 7W.

 

[Idontcare]

I'm NOT assuming that. It's others who are assuming that it necessarily can do 100% theoretical load (relative to 10W at Tjmax) at 80C Tj consuming 7W. I'm saying it isn't clear to me that this is being guaranteed by Intel.

I'm not assuming that either, that is why I keep mentioning this is like turbo-core/boost only in the other direction.

Today you have a base clockspeed and then turbo-bins above that provided you are fitting within a pre-defined TDP at TJmax.

Tomorrow you will have a base clockspeed and "deturbo-bins" below that for which the processor will underclock below the base clockspeed if the processor has been told to do so (the same as we tell our processor to let the clockspeed rise upwards when we enable turbo-core/boost at the platform level).

And just like with today's processors and 3rd party cooling solutions where you can spend more money, enable better cooling, and your turbo-core/boost remain enabled and active for far longer durations than if you simply relied on the stock HSF. So too for the "deturbo" settings, a superior cooling solution will enable the processor to run at less deturbo settings (closer to stock base clockspeed) than it might otherwise have to declock with other less effective coolers.

It is all two sides to the same coin.

 

[Exophase]

I still don't really get the implications of the comment that the performance can go up if there's headroom in the cooling system.

We know that throttling and turbo is temperature based, not power based. So yes, if you can remove heat more efficiently you can turbo higher/longer/better. But that doesn't mean you can do so at the same power consumption. Improving the transfer of heat generated by the core but reduce the power consumption, but it's not the same as removing the heat from the ambient area.

 

[NTMBK]

I still don't really get the implications of the comment that the performance can go up if there's headroom in the cooling system.

We know that throttling and turbo is temperature based, not power based. So yes, if you can remove heat more efficiently you can turbo higher/longer/better. But that doesn't mean you can do so at the same power consumption. Improving the transfer of heat generated by the core but reduce the power consumption, but it's not the same as removing the heat from the ambient area.

The Intel guy was saying that if there happens to be a good enough cooling system, then the chip will briefly turbo up to 2.6GHz. But given that the cooling system must be very weak as shown by the fact that the chip has to run at 800MHz, that turbo could surely only be for a very brief instant- if at all. And forget about getting that turbo when you're in a game and your GPU is eating up that thermal budget.

 

[Exophase]

I'm not assuming that either, that is why I keep mentioning this is like turbo-core/boost only in the other direction.

I didn't call you out in that comment, although you did say this:

Depends on what you are doing, obviously. It doesn't look like they are talking about superior binning, they are basically saying "look, if you can keep the temperature at 80C or below then the chip is gonna only use 7W or less, go above 80C and you are entering into 10W territory".

You seem to be saying that 7W vs 10W can be made purely a function of cooling and not workload.

But what I was really referring to is ShintaiDX's comment:

Its full load usage. It just only applies when the temperture is kept at 80C or below. So it puts an extra demand on the OEMs cooling solution.

If all Intel is saying here is that the system can be throttled so it doesn't use more than 7W under some "normal" cooling system (but can use more with better cooling) then that isn't anything especially new or interesting, in particularly because it doesn't say anything about perf/W. You were earlier talking about how reducing temperature reduces power consumption at the same frequency, so obviously there has some real perf/W improvement here, but it still has to be made very clear if that accounts for all or only some of the 3W savings down from the 10W cTDP.

The Intel guy was saying that if there happens to be a good enough cooling system, then the chip will briefly turbo up to 2.6GHz. But given that the cooling system must be very weak as shown by the fact that the chip has to run at 800MHz, that turbo could surely only be for a very brief instant- if at all. And forget about getting that turbo when you're in a game and your GPU is eating up that thermal budget.

I get that but this doesn't answer anything. Does it need to use more than 7W in that scenario or not? A better cooling system will allow a higher power dissipation.

 

[Idontcare]

I still don't really get the implications of the comment that the performance can go up if there's headroom in the cooling system.

We know that throttling and turbo is temperature based, not power based. So yes, if you can remove heat more efficiently you can turbo higher/longer/better. But that doesn't mean you can do so at the same power consumption. Improving the transfer of heat generated by the core but reduce the power consumption, but it's not the same as removing the heat from the ambient area.

I didn't call you out in that comment, although you did say this:

You seem to be saying that 7W vs 10W can be made purely a function of cooling and not workload.

I must be missing something because it sounds like you are not counting for the fact that for every watt you stop wasting on static leakage (temperature and voltage dependent) is one more watt you can allocate to dynamic power loss (boosting clockspeeds) which then boosts performance

If my chip uses 7W at 80C and 6W at 75C then I have some choices to make.

I can improve cooling such that for a given workload the processor is now running at 75C, using 6W. Now I can turbo-up a bin or two and get more performance at 80C all for the same wattage (7W, because now it is running at 80C).

7W vs 10W can be made purely a function of temperature. Look at the following graph, this is real data:

That power consumption is all from the same load, same performance. The variation in power use is 100% solely dependent on the junction temperature. In this specific case it varies by nearly 30% over a range of roughly 50C.

And here is an example that speaks to the performance increase you can create by improving the cooling (reducing static leakage) which then enables you to have higher clockspeeds for the same power consumption:

For example look at 4GHz stock versus 4.2GHz delidded. They both have the same power consumption, but the 4.2GHz configuration is clearly going to have higher performance at the same "TDP" because it has superior cooling (less static losses, lower operating temperature) than the 4GHz stock config.

 

[Exophase]

I must be missing something because it sounds like you are not counting for the fact that for every watt you stop wasting on static leakage (temperature and voltage dependent) is one more watt you can allocate to dynamic power loss (boosting clockspeeds) which then boosts performance

No, I am accounting for that. I just want to know if Intel or you is claiming that you can go from 10W to 7W by lowering the temperature of the SAME workload.

It's not a question of whether or not lowering the temperature reduces power consumption, I know it does. It's a question of whether or not it reduces it the entire amount specified.

7W vs 10W can be made purely a function of temperature. Look at the following graph, this is real data:

The graph doesn't show that this is the case for this CPU, even for an arbitrarily low temperature. There's an upper limit to power savings from this, and you can't extrapolate the curve you gave at a higher voltage to fit this processor at the voltage ranges it'd run at the given clocks vs temperatures.

And as I've said, there's a difference between improving cooling of the chip and improving cooling of the ambient environment. If you can't go below ambient there's a hard limit to how much cooling you can provide using heatsinks and fans.

 

[IntelUser2000]

We know that throttling and turbo is temperature based, not power based.

It does say in the Ivy Bridge Datasheets that the exponentially moving power average for long duration is factored in for TDP. So while it may be able to briefly exceed TDP for few seconds or so, if you look at it in longer term(like for measuring battery life), it'll be basically equal to TDP in power use anyway.

But given that the cooling system must be very weak as shown by the fact that the chip has to run at 800MHz, that turbo could surely only be for a very brief instant- if at all. And forget about getting that turbo when you're in a game and your GPU is eating up that thermal budget.

It has to use 7W over long periods too though. They are taking advantage of the fact that its of little likelihood that you'll run something demanding when in pure Tablet mode.

So except for small caveats, SDP is basically cTDPdownDOWN.

 

[Exophase]

It does say in the Ivy Bridge Datasheets that the exponentially moving power average for long duration is factored in for TDP. So while it may be able to briefly exceed TDP for few seconds or so, if you look at it in longer term(like for measuring battery life), it'll be basically equal to TDP in power use anyway.

So what you're saying is that basically it throttles based on power too, but applies a very low pass filter to the power measurement? I figured that thermal readings alone were supposed to approximate this effect (since thermal is slow), but that'd have made it possible to sustain a higher power limit. Do you know if this is done to reduce load on the power regulator, reduce wear, or just to bound operating cost?

I guess if I take my laptop and put it in liquid helium I won't be able to keep it at full turbo all the time afterall..

 

[IntelUser2000]

I figured that thermal readings alone were supposed to approximate this effect (since thermal is slow), but that'd have made it possible to sustain a higher power limit.

You do know that Sandy Bridge's Turbo Boost 2.0 relies on the fact that thermal is slower than power? The difference is what gets the greater Turbo over Nehalem/Westmere. But eventually that runs out and they settle at whatever the constant TDP limit is. I guess depending on the situation it can be some Turbo active, or even just Base.

Apparently you can set the duration for the exponentially weighed average power(probably firmware for the manufacturers). It goes from 32 seconds for the ULV to 64 seconds on the dual core and quad. You can go as little as 1 second.

 

[Exophase]

You do know that Sandy Bridge's Turbo Boost 2.0 relies on the fact that thermal is slower than power?

Isn't that what I was saying..?

But eventually that runs out and they settle at whatever the constant TDP limit is. I guess depending on the situation it can be some Turbo active, or even just Base.

Right, that's the part I didn't know. What does exponentially weighted average mean exactly?

 

[Lepton87]

That power consumption is all from the same load, same performance. The variation in power use is 100% solely dependent on the junction temperature. In this specific case it varies by nearly 30% over a range of roughly 50C.

.

That's why I don't think that 20C TJ drop could result in over 53% power consumption drop. They combine less strenuous workload with lower TJ to achieve that goal.
Think about it, at TJ85C it is 7W and at TJ105 it is 13W. So for 20c increase in temperature you get 85% increase in power consumption. Doesn't sound realistic.

 

[dastral]

No, I am accounting for that. I just want to know if Intel or you is claiming that you can go from 10W to 7W by lowering the temperature of the SAME workload

I thought that was a given, a fact...
If you need X watts to run a CPU at 60°, you will need more to run it at 90° and less at 30°.
Maybe SDP is Intel's way of saying it (but the numbers just don't add up)

 

[sm625]

Does anyone have any benchmarks of an ivy bridge i5 capped at 800MHz? I'm sure an i5 clocked at 800Mhz would still slaughter an atom or an E-450 but I would like to know by how much.

 

[Charles Kozierok]

Different article from a different time/market.

Yes, and no. The time is different, and the market is different. But the motivations and objections are identical.

Both server and mobile environments are power sensitive. Both represent situations where there are tradeoffs between power consumption and performance. And both ACP and SDP were created based on a claimed desire to better represent actual power usage rather than power limits intended to reflect cooling system design.

The main reason I posted that article is to show that not only was Intel opposed to this sort of thing in the past, they were rather condescending about it. Now, suddenly, it's a great idea!

 

[Abwx]

Technicaly , they integrated the power comsumption curve ,
that is , they took the average (arithmetic or exponential )
sliding mean over a period that is computed to take account
of the cooling and casing thermal inertias.

As such , it can be made to run at higher power drain than
if only average power is taken into account.

For instance , it can run at 13W for 2 minutes and then
being iddled at 3W for 4 minutes , the resultant SDP will be 7W.....

Depending of the integration period you can set the SDP at any
convenient value , let s say 5W if you re assumming that the CPU
run at 3W during 80% of the time , you ll end with a 13W CPU
specified for 5W SDP.....

 

[ShintaiDK]

Does anyone have any benchmarks of an ivy bridge i5 capped at 800MHz? I'm sure an i5 clocked at 800Mhz would still slaughter an atom or an E-450 but I would like to know by how much.

You need an mobile SB/IB. Since the lowest multiplier possible is 8x.

 

[Idontcare]

That's why I don't think that 20C TJ drop could result in over 53% power consumption drop. They combine less strenuous workload with lower TJ to achieve that goal.
Think about it, at TJ85C it is 7W and at TJ105 it is 13W. So for 20c increase in temperature you get 85% increase in power consumption. Doesn't sound realistic.

It all depends on how many of those 13W at 105C are due to static leakage versus dynamic power, as well as the activation energy (temperature dependence) of the leakage itself.

It is not impossible nor implausible that a processor at 105C uses 13W but at 80C it would use 7W, just means you have a rather leaky process tech at high temperatures and it is also sharply temperature dependent.

The physics are all there, its not rocket science but on the other hand it isn't magic either.

People are reacting to this with the attitude that it is impossible and that completely perplexes me because it is possible. What none of us can say is that it is what is actually going on, we can't positively confirm it because we don't have the silicon in hand to run the tests needed to determine the device physics parameters, but we also cannot positively rule it out because of the same reasoning.

But at the outset there is nothing about the device physics here that precludes the possibility that the 25C reduction in temperature solely accounts for the reduction in power footprint.

Take my i7-3770K for example, fully characterized across the entire multiplier range for static and dynamic power use properties.

It is a quad, so if we cut the power numbers in half to approximate the dynamic and static power of a dual core, dial in the clockspeed to 1.4GHz and compute power use I get the following:

1.4GHz, 0.760V, 105C = 6.92W static power and 6.03W dynamic power (total is 12.95W at 105C)

Lower the temperature and the necessary voltage accordingly and you get:
1.4GHz, 0.615V, 80C = 3.09W static power and 3.95W dynamic power (total is 7.04W at 80C)

Now this is just an example, obviously the operating voltages are picked such that the numbers work out but the voltages are also not unreasonable (my 3770k is LinX stable at 1.6GHz with just 0.636V at 36C). But the leakage parameters are real-world, for the same node and same microarchitecture.

So to say it is impossible or unlikely that Intel manages to reduce power by nearly half just by enforcing a 25C reduction in max operating temperature is really an unfounded assertion.

But...just because I can prove it is possible and plausible doesn't mean that is what Intel is doing, it just means we are not justified in concluding a priori that Intel couldn't possibly be doing it. It is very much within the realm of the possible.

 

[Nemesis 1]

So you guys got this all figured out . Hell now you changed what I waiting to see . Intels or Oems cooling solutions for a 7 watt chip . LOL

 

[Nemesis 1]

So on haswell A tdp 7 watts were looking at cores running at 4 watts is that not so.

 

[Blandge]

1.4GHz, 0.760V, 105C = 6.92W static power and 6.03W dynamic power (total is 12.95W at 105C)

Lower the temperature and the necessary voltage accordingly and you get:
1.4GHz, 0.615V, 80C = 3.09W static power and 3.95W dynamic power (total is 7.04W at 80C)

I can't believe that worked out so perfectly.