Haswell & the Reverse Turbo

[cytg111]

(Hold onto your knickers guys(s) not a troll thread, no need to report it just yet)

-So, I was looking into getting a surface pro 3, tablet/laptop AllInOne replacement, full x86, real windows 8 (not the locked down rt/arm stuff, ms is really botching that one).
So starting researching the SP3 a bit and found some reports

http://forums.wpcentral.com/microso...rottling-i7-i5-=-i3-speed-sustained-load.html

The Surface Pro 3 is not a good choice for professional software (i.e. DAW, encoding, etc) I bought it for or games that rely on sustained load capability. With sustained load, the i7 slows down to i3 speeds in short order. The Surface Pro 2 did not do this. Even simple Windows 8 store games run faster on my ASUS T100TA (Atom Z3740, 2GB RAM) than my Surface Pro 3 (i5 4300U, 8GB RAM) on battery due to aggressive throttling.

http://www.youtube.com/watch?v=UbWYQWY-vU8&feature=youtu.be
- Here it is struggling through minecraft (is that even hurting the IGP?)

http://answers.microsoft.com/en-us/...rottling/cdb99447-2f23-4ef1-9336-0c4e6b7de488

Surface Pro 3 is throttling CPU while playing games. I have confirmed this while attempting to play diablo 3 which runs fine at first, until the temperature get to a certain degree and fps start to drop. Not the only one experiencing this issue so please fix through a firmware upgrade to bring back "high performance" profile like my surface 1. Thanks

Anand reported on it too : http://www.anandtech.com/show/8077/microsoft-surface-pro-3-review/3

I first stumbled upon this behavior while trying to gather thermal data for Surface Pro 3. I noticed large run to run variance if I repeatedly ran 3DMark 11. I'm used to seeing this sort of behavior on smartphones that throttle quickly, but it was unique for a Surface Pro device.

But then leave it to microsoft right?
But a quick search reveals that this happens to other haswell tablets as well
Even Intels own NUC
http://www.silentpcreview.com/article1373-page3.html

With the graphics card torture test of Furmark, CPU temperatures jumped above 82°C, when the system would go into throttling, forcing the clock speed to drop as low as 800 MHz, then getting into a periodic cyling between higher clock speed (up to 2.3 GHz Turbo in multi-core load) and throttling over a 2 minute period.

I can understand the concept of getting a certain amount of work done faster->return to idle faster and cool down, it is problary good etiquette on a phone or even tablet, however, should you then advertise the chip as 2GHz? 3Ghz? If all you can guarantee is the 800MHz base clock then at least for a tablet device shouldnt you advertise that? 800MHz clock with a 2.6GHz turbo?
I am sure that the much improved IGP in haswell is especially targetting mobile, but what use is it if it throttles?
I just think that with the concept of "one chip to rule them all", from mobile to desktop to workstation - formfactors, the water is getting a little muddy. We have lots of desktop rigs throttling too, almost as pr. design.
So this is new(feature-throttling). Is it something we'll have to get used to or is things getting back to normal with Broadwell, Skylake etc.
(Oh, I dubbed this "feature-throttling" as Reverse Turbo, clever, right? )

 

[BrightCandle]

For several years now its been a common answer to peoples performance problems to check their temperatures, as throttling is happening on some peoples desktop PCs as well. The inconsistency of boost is causing problems with smooth constant delivery of frames for games and its worse in laptops and the more thermally constrained the laptop is the worse the situation gets.

I am not in the least bit surprised by this behaviour, we have been fighting it for a while.

 

[witeken]

Anand reported on it too : http://www.anandtech.com/show/8077/microsoft-surface-pro-3-review/3


But then leave it to microsoft right?
But a quick search reveals that this happens to other haswell tablets as well
Even Intels own NUC
http://www.silentpcreview.com/article1373-page3.html

No surprise. That's sort of what happens when you put a 15W chip in a (thin) tablet. Broadwell will fix / considerably improve Core tablet performance.

I can understand the concept of getting a certain amount of work done faster->return to idle faster and cool down, it is problary good etiquette on a phone or even tablet, however, should you then advertise the chip as 2GHz? 3Ghz? If all you can guarantee is the 800MHz base clock then at least for a tablet device shouldnt you advertise that? 800MHz clock with a 2.6GHz turbo?

Unlike Qualcomm and all other tablet/phone chip manufacturers, Intel does publish base and boost clocks, so I don't see the problem.

I am sure that the much improved IGP in haswell is especially targetting mobile, but what use is it if it throttles?

Do you mean Broadwell?

The key thing you really have to look for is performance per watt. All other things equal, a chip with a high(er) performance per watt will be able to go to higher clock speeds or use more EUs. 14nm and Gen8 will improve the performance per watt a lot, so the performance will be much higher (at highest sustained thermals).

I just think that with the concept of "one chip to rule them all", from mobile to desktop to workstation - formfactors, the water is getting a little muddy. We have lots of desktop rigs throttling too, almost as pr. design.
So this is new(feature-throttling).

Intel's Core architecture is great. Have you seen the benchmarks of Core M and Haswell E5?

Is it something we'll have to get used to or is things getting back to normal with Broadwell, Skylake etc.
(Oh, I dubbed this "feature-throttling" as Reverse Turbo, clever, right? )

Turbo is a great feature. It vastly improves performance for consumer workloads, so it is a good thing. You don't lose out on anything with turbo enabled; it doesn't harm performance.

 

[ShintaiDK]

If it for example throttles at 82C, its because the OEM decided so for various reasons. For example to protect other components.

Microsoft is working on multiple fixes for the Surface Pro 3. But it looks like a product where Microsoft simply wanted more than they should.
http://www.neowin.net/news/microsoft-is-aware-of-the-surface-pro-3-i7-heat-issues-fix-coming-soon

Not sure why you wish to blame Haswell yet again. We dont exactly lack examples of laptops running at lower speeds on battery due to OEM settings.

 

[kimmel]

(Oh, I dubbed this "feature-throttling" as Reverse Turbo, clever, right? )

Thermal throttling is throttling if it happens to Qcom, Apple or Intel. Reverse turbo is not really clever for something we already have a term for. During the Oculus Carmack keynote he noted that if you try to do anything somewhat taxing on the Qualcomm based phones you can get massive throttling in 20 seconds. Even apple's iphone 6 slide showing non-existent throttling was complete junk as independent testing showed.

With the graphics card torture test of Furmark, CPU temperatures jumped above 82°C, when the system would go into throttling, forcing the clock speed to drop as low as 800 MHz, then getting into a periodic cyling between higher clock speed (up to 2.3 GHz Turbo in multi-core load) and throttling over a 2 minute period.

Also, really? You want battery powered lightweight solutions optimized for non throttling furmark? Because that's just silly and this thread is silly if you think so.

If all you can guarantee is the 800MHz base clock then at least for a tablet device shouldnt you advertise that? 800MHz clock with a 2.6GHz turbo?

Intel advertises base frequencies much more prominantly than phone chip companies.

However, the market has spoken. Short bursts of high frequency with throttling later wins sales in the laptop/phone consumer market as shown by Apple and Qualcomm. With half decent cooling without overclocking I have yet to see a desktop throttle below base. Servers also with cooling commensurate with TDP ratings also can run pretty much flat-out as well without throttling.

Don't hate the player, hate the game.

 

[dave_the_nerd]

Pentium 4s would throttle if they got too hot. This isn't really new. (Remember this?)

 

[cytg111]

Intel's Core architecture is great. Have you seen the benchmarks of Core M and Haswell E5?

- Yes, eagerly awaiting a SP3 refresh or SP4.. I hope.. it throttles less .

Unlike Qualcomm and all other tablet/phone chip manufacturers, Intel does publish base and boost clocks, so I don't see the problem.

Sure, but in terms of a "laptop" you are buying an i5 with advertised x.yGHz speed. Compared to the ARM chips it is up against in its segment, I get it .. it is just not the only segment it competes in..

 

[cytg111]

Also, really? You want battery powered lightweight solutions optimized for non throttling furmark? Because that's just silly and this thread is silly if you think so.

- That was the NUC was it not? Also, we're not in tablet land here, we are in hybrid land.. As I perceive the market these tablets, x86 with fullblown windows, is sold as laptop AND tablet all in one. And yes, for a plugged in laptop i certainly expect it to run what I throw at it at full steam for as long as i want. (I guess it sounded too good to be true .. 2014 and still, no free lunch...)

Pentium 4s would throttle if they got too hot. This isn't really new

Yup, but haswells throttling mechanism/threshold is much much more agressive is it not? And it does it so well that you are sort of switching territory from "fail safe" to "feature". IMO anyway.


Also, stumbled on this

http://www.gamerheadlines.com/2014/...oost-performance-by-20-reduce-heat-and-noise/

Surface Pro 3 Undervolting Boost Performance by 20% reduce heat and noise

Undervolting to *gain* performance ....

 

[Dufus]

Undervolting to *gain* performance ....

This is fairly well known for users with mobile Haswell. Usually the performance increase is due to being able to run an increased multiplier where power throttling is happening due to TDP limits. In this case there will not be a huge saving in temperature at load though as still running the same power.

 

[Enigmoid]

Not sure why you wish to blame Haswell yet again. We dont exactly lack examples of laptops running at lower speeds on battery due to OEM settings.

I agree, this is not a haswell specific problem. MS messed it up and didn't design a proper cooling system or is implementing extremely conservative BIOS settings.

 

[VirtualLarry]

Don't forget, you don't want your tablet CPU to cook your Li-Ion battery, that could potentially be very dangerous!

 

[PPB]

Its not throttling. Its just running at non turbo speeds. This is what happens when you start to over rely on turbo to get performance. And it is a haswell problem as it is an inherently hot chip, blame 3d fets, the integrated voltage regulator, whatever.

That is why i laugh at whoever claims x86 is valuable at that form factor. Why would i want x86 if the worthwhile software that only run on that isa are intensive enough to make the cpu run at stock speeds most of the time? (and get horribad performance because of it). For a sku that can only handle light loads of bursty nature like web browsing, arm will always make more sense.

 

[Dufus]

The impression I got was that it might be an internal temperature sensor on the SP3 that was activating BDPROCHOT. An external signal to BDPROCHOT will cause the processor to throttle by limiting speed to the LFM. IOW not the processor getting too hot but the internal temperature of the SP3 triggering throttling.

 

[coffeejunkee]

Its not throttling. Its just running at non turbo speeds. This is what happens when you start to over rely on turbo to get performance. And it is a haswell problem as it is an inherently hot chip, blame 3d fets, the integrated voltage regulator, whatever.

In the video from the first link the cpu is dropping below the non-turbo speeds. Regular speed is 1.7 Ghz (http://ark.intel.com/nl/products/75114/Intel-Core-i7-4650U-Processor-4M-Cache-up-to-3_30-GHz) but it drops down to 1.2 Ghz. It's different from the throttling in the NUC though which goes down to the lowest possible 800MHz.

Don't know what to think of this. As a reader of tech forums I would say 'yeah you could have known' but you can't expect every customer to research stuff like this. I can imagine this being a nasty surprise for some people.

Also, can't compare this with regular desktop cpu's. For those to throttle there has to be something wrong, like the cooler's fan malfunctioning. Every normal desktop cpu can run Prime 95 without throttling.

 

[Dufus]

Here's a well done video showing the SP3 throttling at 0.9GHz with the CPU temperature in the 70's. Then with the use off an external fan the clock goes up to 2GHz and CPU temp up to 90C. Similar results were obtained with the use of a much smaller USB fan.

http://www.youtube.com/watch?v=I__UMfYWcI8

 

[cytg111]

Here's a well done video showing the SP3 throttling at 0.9GHz with the CPU temperature in the 70's. Then with the use off an external fan the clock goes up to 2GHz and CPU temp up to 90C. Similar results were obtained with the use of a much smaller USB fan.

http://www.youtube.com/watch?v=I__UMfYWcI8

A few lessons I guess, one being that haswell is premature in a tablet form, still too hot. I mean what good is unlocking the x86-windows software ecosystem if you cant run the software reliable. Hope CoreM and skylake fixes that.

Yet, even in desktop systems (and the NUC) haswell is overheating. Not all, but some, and usually stock cooled systems. As the chips get smaller so is the die area radiating the heat, it is not rocket science to figure out that for x watts to leave y mm^2, then for x watts to leave y/2 mm^2 you will need a cooling system twice as efficient pr mm^2 removing the heat.
I guess that if we want to stay within current paradigm of cooling solutions (metal block with fan on top) then for every time we halve the die area (typically a new node) then we need to halve the watts too, all other things being equal.
22nm vs 14nm, Haswell-H :47-57watts vs. Broadwell-H : 37-47watts
Halving the die size but not quite halving the watts...
Is that something moving against Core x86 in mobile formfactors (desktops as well i guess).. Are new cooling solutions coming up?

 

[Dufus]

Yes, power density increases with shrinkage and perhaps why TDP is being lowered but in the case of the Surface Pro 3 in the linked video it was not the CPU temperatures causing the throttling, at least not directly. This can be seen by cooling the whole device itself which led to higher CPU temperatures as the CPU was able to run over twice as fast.

Lot of reasons for throttling besides the CPU itself. For instance the SP3 BIOS also has temperature limits for the PCH, Memory, Ambient, Skin, Exhaust and VR.

 

[IntelUser2000]

The Surface Pro 2 did not do this.

BTW, based on that small sentence, and from reviews I never thought "Surfaces" were brilliant design on the hardware by any means.

It's not the thinnest Core Tablet, nor the lightest. And even then the thermals and battery life weren't excellent either. I guess the pricing is little more affordable which may be the reason for all this.

The Vaios and the Zenbooks and the Aspires do better, despite that some of them are as nearly as portable as the Surface Pro even as a clamshell form factor. Of course they are more pricey.

http://www.anandtech.com/show/8077/microsoft-surface-pro-3-review/3

But what it shows is that there's a limit to form factor putting a 15W chip. The SP3 is lot thinner than the SP2, and performs lower. The devices that moved to Haswell without changing the form factor much, like the XPS 12, and the Aspire S7 shows that it runs somewhat cooler than on Ivy Bridge, likely reflecting the reduced TDP.