Testing Turbo

The original MacBook Air had a 1.86GHz Core 2 Duo that pretty much never ran at 1.86GHz. Instead the chip typically ran at or below 1.4GHz, allowing the second generation 11-inch MacBook Air to outperform it. Since then I've been quite concerned about Apple playing clock speed games with its systems, but thankfully since then we haven't seen any similar issues.

Letting my guard down wasn't something I was interested in doing so I went about verifying Turbo on the new Air. First a quick refresher about what Turbo Boost does on Intel CPUs. All CPUs are designed to certain power and thermal limits. Those limits are shared with manufacturers who then spec power supplies, cooling systems and cases that can handle those CPUs. In the old days this was pretty simple, you had a single core and a single thermal design point (TDP) that you had to hit. Then multicore CPUs came around and made things more complicated. TDPs couldn't go up (laptops weren't going to double in size) but you now had multiple cores sharing the same TDP, so each core had to dissipate less heat. This meant that your dual-core CPU would run each core at a lower frequency than your single-core CPU, and your quad-core CPU would run at an even lower speed. When using all of your cores the tradeoff was worth it, but when only one or two were in use, it wasn't. We needed a way of shutting down cores to free up TDP and dynamically adjusting clock speed.

  Single Core Dual Core Quad Core
TDP
Tradeoff

Nehalem came along and introduced power gating to x86 CPUs. Unused cores could be power gated, effectively shutting them off, and the resultant available TDP could be used by turbo boosting active cores. All subsequent Intel architectures have supported Intel Turbo Boost, a feature we investigated extensively in our 2011 MacBook Pro review earlier this year.

Lion got rid of the 32-bit kernel so my previous trick to run MSR Tools to measure clock speed stopped working. I needed an alternative.

I offer you three pieces of evidence that support the theory that Apple is not artificially limiting clock speed on any of the MacBook Air systems I've tested.

The first is output from Lion's boot process, note the line that begins with AppleIntelCPUPowerManagement:

The screenshot above is from a 1.6GHz Core i5 on an 11-inch MacBook Air. This chip runs at 1.6GHz by default but can turbo up to 2GHz with two cores active and 2.3GHz with one core active. The line above lists turbo ratios 0047. This four digit number shows turbo ratios for 4C/3C/2C/1C active, with each digit corresponding to a different max turbo ratio. Each value is in hex to represent ratios above 9 with a single digit.

Take the example above: 0047. The first two numbers are 0s because the chip doesn't have more than two cores and thus doesn't support any turbo ratios when 4 or 3 cores are active. The third number tells us the maximum turbo boost with two cores active: 4. That's 4 bins, where each bin is 100MHz, or 400MHz above the stock 1.6GHz operating frequency (2.0GHz).

The fourth number gives us max turbo when only a single core is active: 7. Seven bins is 700MHz, which on top of the 1.6GHz base frequency gives us 2.3GHz.

What about the 1.7GHz Core i5 in the 13-inch MacBook Air? Its turbo ratios look like this:

The first two digits are 0 again since this is a dual-core processor. The third digit is a 7, representing a 700MHz max turbo boost with two cores active. The fourth digit is an A, which is 10 when you convert from hex to base 10, or a maximum 1GHz turbo boost with one core active. If you had a 1.8GHz Core i7 you'd see the string 008B.

The second bit of evidence supporting that Apple isn't mucking with clock speeds comes via Cinebench. The 2.0GHz Core i7 in the 15-inch MacBook Pro can turbo up to 2.9GHz with only one core active. Cinebench has a great single threaded test that ensures only one core is active. The chip in the 13-inch MacBook Air should be able to hit 2.7GHz with one core active. We already know the CPUs in the 2011 MacBook Pro have fully functioning turbo ratios based on our earlier investigations. While the i7 in the MacBook Pro has a larger L3 cache, we should still see around a 7% performance advantage in single threaded Cinebench for the 2011 MacBook Pro vs. the 13-inch MacBook Air (2.9GHz vs. 2.7GHz) if the Air is properly turboing up:

Clock Speed Scaling Comparison
  15-inch MacBook Pro (Core i7 2.0GHz) 13-inch MacBook Air (Core i5 1.7GHz) 11-inch MacBook Air (Core i5 1.6GHz)
Cinebench R10 - Single Thread 4060 3770 3154

The 2011 MacBook Pro is 7.7% faster in Cinebench's single threaded test. Perfect. We can't do the same comparison with multithreaded performance because the i7 in the MacBook Pro has twice the number of cores/threads as what's in the Air, but by this point I'm pretty confident that Apple is allowing these chips to turbo up.

If you need one more piece of evidence I present to you behavior of the 2011 11 and 13-inch MacBook Air under Windows 7, where we do have good turbo monitoring tools courtesy of Intel.


13-inch MacBook Air, Core i5 1.7GHz, max single-core turbo

Both systems had no problems hitting their specified turbo ratios, we are not artificially constrained here at all. In fact, the 11-inch MacBook Air was able to even exceed its dual-core turbo ratio by a small amount during some of my testing:


11-inch MacBook Air, Core i5 1.6GHz, temporarily exceeding max dual-core turbo

In the example above the 1.6GHz Core i5 is running a multithreaded workload, Cinebench once again. Both of its cores are pegged at 100% and it is able to turbo up to 2.0GHz, its max turbo frequency with two active cores. For a short period of time during the render however, we actually see numbers above 2GHz.


11-inch MacBook Air, Core i5 1.6GHz, max single-core turbo

This is a feature of Sandy Bridge. If your chip is idle then all of sudden gets a heavy workload it'll try to turbo up to its max ratio. The chip itself will take time to actually hit its max operating temperature though, it doesn't simply jump up to max TDP just because it's running at a higher frequency. During this TDP ramp, Sandy Bridge actually allows the chip to run even faster than its max turbo frequency for a short duration. This overboost gives you a little extra oomph in performance without ever violating the chip's TDP.

The CPUs A New Thunderbolt Implementation
POST A COMMENT

115 Comments

View All Comments

  • tipoo - Thursday, July 28, 2011 - link

    I'd like to know too, in fact I think a decibel reading for laptop reviews would be great. Reply
  • solatic - Thursday, July 28, 2011 - link

    "... [It's] easy to imagine a future where laptops become a lot more like the new Air and shift to a couple high bandwidth ports instead of numerous lower bandwidth connections."

    I agree in a sense, but I very much disagree with the manner of your idea.

    Getting rid of low-bandwidth ports on laptops is stupid because these low-bandwidth ports are industry standards. The standard display jack for projectors everywhere, from business rooms to classrooms etc., is VGA. Good old VGA from the 90's which was never displaced by HDMI or DVI despite their ubiquity and technological superiority. Why is irrelevant, but my point is that I can't tell you how many people I've seen with Apple laptops who time and again have asked to borrow my machine with its VGA port because they can't find their VGA dongle or forgot it at home/the office.

    VGA, RJ-45, USB - we don't use these jacks because of how much bandwidth they move but because we know we will encounter devices in the field that will use them.

    What you really want, Anand, is a docking station. Lenovo/IBM has/did made/make them for some time for the Thinkpad line. You come to the office, slide your laptop in, and boom - the docking connector is a high bandwidth connector that connects you to network, display, audio, interface, etc.

    The only real problem with docking stations is that they're either proprietary (the Thinkpad ones) or they're too slow for higher-bandwidth applications (USB docking stations). If you see a future in Thunderbolt docking station-type devices - like the Thunderbolt display - then this is a good thing. The Thunderbolt display can now be used by any Apple computer with a Thunderbolt port - whereas Lenovo has to manufacture different docking stations for the X and T series and these docking stations can't be used with Toshibas, Apples, Dells, etc.

    But to propose getting rid of VGA and RJ-45 ports now is not something I can agree with. Put Thunderbolt on new machines, make Thunderbolt projectors etc., wait for them to saturate the market - and then, only then, does it really make sense to get rid of these slower ports.
    Reply
  • repoman27 - Friday, July 29, 2011 - link

    It certainly makes sense to ditch VGA on a product like the MacBook Air, since a VGA port is too large to physically fit inside it. And it makes sense on all other laptops since most people would rather have 10 more mins of battery life all the time than a VGA port on the odd occasion that they need one. I propose that those who own archaic video devices lacking a digital interface buy a $5 adapter and leave it attached to the device, that way people with modern notebooks can connect to them without issue.

    The lack of a wired Ethernet port on the Airs is also due to its ultra slim profile, but it amazes me how many people I know that have no idea that they can plug their laptops into a wired network and get far better throughput. As long as WiFi offers more bandwidth than most people's ISPs, I think you can kiss that RJ45 port goodbye without upsetting too many folks.
    Reply
  • Wolfpup - Thursday, July 28, 2011 - link

    Great reviews like this are why I love this site. Super thorough and interesting on the tech, and, interesting thoughts on practical stuff too.

    I prefer Windows, but have been wanting a secondary Mac for a while for the heck of it. I'm SUPER torn on what to get...

    The 11.6" almost seems perfect, since I can throw it in my bag with my main notebook and be okay-stick it on my desk without too much issue. But...if I ever actually used it as my primary system, the 13.3" one would be a lot more usable. And at THAT point, the 13.3" Pro is a lot more usable, and at THAT point, the 15.4" Pro isn't much larger, and completely destroys it, and of course could be my main system...

    Sooooo you see my dilemma ;)

    Heh...maybe I should just go with the 11 since I'm not planning on using it as my primary.
    Reply
  • name99 - Thursday, July 28, 2011 - link

    "All three parts support Hyper Threading and Quick Sync, although the latter remains mostly unused in OS X. "

    Quick Sync is used by iChat HD, is it not?

    The other natural client for it would be AirPlay. My guess is that, come iOS5 in September, we will see AirPlay on SNB macs beefed up to be able to stream any content (not just h264) to AirPlay devices by doing the transcode transparently on Quick Sync.

    The third obvious sort of client would be a library that third party apps like HandBrake would get to. What's the currents situation now --- do you need to be root to get to QuickSync or can any app use it?

    My guess is that we are facing the constant problem of new "weird" hardware --- it never comes virtualizable in the first iteration, which means that there is ALWAYS the problem of how to mediate access. And we generally see the same pattern
    (a) A single app that is allowed access.
    (b) Some sort of library that provides its own mechanisms for mediating access.
    (c) The hardware (FINALLY) becomes virtualizable.

    Apple is currently at step (a). Getting to (b) is never completely trivial (in spite of the claims of no-nothings in blog comments), at least if you want to do the job properly. You have to consider questions like --- do you use a reservation model, or do you simply provide notifications when you want to grab the hardware away from a user? How easy is it juggle state and provide something that looks virtualized? etc etc.
    And there are ALWAYS, at least in the first gen, weird hardware interlocks that make life more difficult. I know nothing about QuickSync but I would not be surprised if, for example, using it has implications for using the main GPU, meaning one more thing that has to be balanced in the attempt to make it used more generally.

    Can someone from the Windows world (who understands these issues, and has something more useful to say than "Macs suck") tell us how QuickSync is used in the MS world? Does MS provide a general purpose library, and how does the mediation model for that library work?
    Reply
  • name99 - Thursday, July 28, 2011 - link

    Oops, my bad. The Macbook Airs apparently do not have an HD camera because it can't fit in the available depth of the thin screen. So no iChat HD on these models.

    I think the rest of what I said, especially about AirPlay, still stands.
    Reply
  • rootheday - Friday, July 29, 2011 - link

    The sharing/scheduling of the GPU enginers by multiple client applications on Windows is mediated by the OS as part of the WDDM driver model dating back to Windows Vista - QuickSync is no exception. This means that we are already at c) on your hierarchy on Windows with multiple client applications able to easily share the GPU for encoding.

    Moreover, Intel has a library already for this - see http://software.intel.com/en-us/articles/media/

    The media sdk library provides an API that applications can use to perform encoding and decoding. If run on a system with QuickSync hardware support and drivers, the work is routed to the GPU. If not, the library offers a CPU fall back path. This allows ISVs to write their application once - it will just run faster on Sandybridge systems.

    I don't know enough about Apple OS and graphics driver model to comment on how hard it would be for Apple to get to the same level.
    Reply
  • jvmxtra - Thursday, July 28, 2011 - link

    Thanks for the great review. I throughly enjoyed your review but I feel like one thing is missing.
    In fact, for a laptop review, I really want all the sites to start devoting some time and even creating a method to measure the heat that laptop brings on. We have to create some type of way to measure(benchmark?) the heat index as I feel like how hot laptop gets under certain circumstance is critical factor.

    In fact, I had to trade in my 15 inch mbp since it was just getting too hot.
    Reply
  • tipoo - Thursday, July 28, 2011 - link

    The thermals and power consumption page is a start. Reply
  • name99 - Thursday, July 28, 2011 - link

    "This is what Thunderbolt was meant to do. All we need now is widespread adoption, more accessories and a standard for external GPU form factors."

    AND device manufacturers who are not idiots. In particular, where are the TB hubs?
    The device the market obviously wants is a TB to USB3 hub ---
    two TB ports, four USB3 ports --- and yet we still have not seen this.
    I'm sorry, Sonnet, but this would be VASTLY more useful than your EN and FW800 bridges.

    WTF is going on?
    Reply

Log in

Don't have an account? Sign up now