Understanding Nehalem’s Turbo Mode

Modern day CPUs and GPUs are more power constrained than anything else. They could run faster, if they could get around pesky problems like power density. Intel and AMD have both figured out that the maximum power consumption for a single processor falls into one of the following ranges depending on the platform:

System Processor TDP Number of Cores
High End Desktop 80 - 130W 4
Mainstream Desktop 65W 2 - 4
Notebook 20 - 45W 2
Ultra Portable Notebook 10 - 20W 1 - 2
Netbook 2 - 5W 1

 

If we look at the bottom of the table we see that our limits to performance aren’t technology, but rather power; netbooks could be as fast as desktops if we could stick 130W processors in them.

Pay attention to the third column however. A high end desktop processor is designed to dissipate up to 130W of heat; you reach that value by running all four cores at full load. But what happens if you only have two active cores? The total power consumption and thermal dissipation of your processor is no longer 130W, it’s noticeably less.

I just finished saying that power was our fundamental limit to faster microprocessors, but if half of a 130W chip is idle - shouldn’t the working half be able to run faster? The answer is yes, but only with some clever technology.

The Nehalem CPU includes a fairly complex hardware monitoring microprocessor on-die. This processor is called the Power Control Unit (engineers r awesome). It monitors the temperature, current and power consumption of each core independently. The PCU also the part of the chip that handles OS requests to drop the cores down to lower power states. Now get this; if there’s room in the power envelope, and the OS requests a high performance state, the PCU will actually increase the clock speed of the active cores beyond their shipping frequency.

It all boils down to the TDP of the chip, or its Thermal Design Point. The more TDP constrained a platform is, the more you stand to gain from Intel’s Turbo mode. Let me put it another way; in order to fit four cores into a 130W TDP, each core has to run at a lower clock speed than if we only had one core at that same TDP.

At higher TDPs, there’s usually enough thermal headroom to run the individual cores pretty high. At lower TDPs, CPU manufacturers have to make a tradeoff between the number of cores and their clock speeds - that’s where we can have some fun.

The Other Difference Between the Quad and Eight Core Models

Apple sells two versions of the new Mac Pro, a quad-core and an eight-core system. The motherboard is the same in both machines, but the processor board is different. The quad-core processor board has a single LGA-1366 socket and four DIMM slots, while the eight-core processor board has two sockets and eight DIMM slots. They also use significantly different CPUs, although Apple doesn’t tell you this.

Below you’ll find the standard and upgraded options for each system:

Apple Mac Pro (2009) Quad Core Model Eight Core Model
Default CPU Xeon W3520 (2.66GHz) Xeon E5520 (2.26GHz)
CPU Upgrade Options Xeon W3540 (2.93GHz) Xeon X5550 (2.66GHz)
Xeon X5570 (2.93GHz)

 

Although Apple offers a 2.93GHz CPU in both systems, it’s actually a different chip that’s used in each model. The clock speeds, core counts and cache sizes are the same, the difference is in the TDP.

The quad-core Mac Pro uses 130W TDP Xeon uniprocessor workstation processors, the eight core Mac Pro however uses an 80W (2.26GHz) or 95W chip (2.66/2.93GHz). There are more CPUs in the eight-core model, so Intel offers chips with lower TDPs to keep total platform power under control. While the eight-core Mac Pro uses more power than the quad-core Mac Pro, each chip individually should use less power. And remember what we discussed earlier: lower TDPs mean higher turbo frequencies.

The table below shows the maximum turbo frequency available for each chip depending on the number of cores currently in use:

System (Processor) Default Clock Max Turbo w/ 4-cores active Max Turbo w/ 3-cores active Max Turbo w/ 2-cores active Max Turbo w/ 1-core active
8-core Mac Pro (Xeon X5570) 2.93GHz 3.20GHz 3.20GHz 3.33GHz 3.33GHz
8-core Mac Pro (Xeon X5550) 2.66GHz 2.93GHz 2.93GHz 3.06GHz 3.06GHz
8-core Mac Pro (Xeon E5520) 2.26GHz 2.40GHz 2.40GHz 2.53GHz 2.53GHz
4-core Mac Pro (Xeon W3540) 2.93GHz 3.06GHz 3.06GHz 3.06GHz 3.20GHz
4-core Mac Pro (Xeon W3520) 2.66GHz 2.80GHz 2.80GHz 2.80GHz 2.93GHz

 

What the table above tells us is that while the quad-core Mac Pro can turbo up by 133MHz if more than one core is active, and 266MHz if only one core is active, the processors in the eight-core Mac Pro can do better. The Xeons in the eight-core Mac Pro can turbo up by 266MHz or 333MHz, depending on the number of cores active. The 333MHz turbo mode is available even if two cores are active.

Apple isn’t big on specs like these so we don’t see any mention of them in Apple’s Mac Pro sales literature, the only clue you get is in the form of the model numbers Apple lists on its spec sheets:

Although it’s a pricey upgrade, you do get better processors with the eight-core Mac Pro than you do with the quad-core version. If you don’t need more than four cores however, you’ll still be better off with a 2.66GHz quad-core Mac Pro than a 2.26GHz eight-core model.

The Crossroads of Simplicity and Sophistication Performance
POST A COMMENT

58 Comments

View All Comments

  • jamesst - Tuesday, July 14, 2009 - link

    "The Lexar reader is FireWire 800 (woo!) and the iSight is FireWire 400; I can’t use the iSight on the new Mac Pro."

    You can still use your Firewire 400 iSight camera on the Mac Pro's Firewire 800 ports. All you need is a Firewire 400 to Firewire 800 cable. I know that Belkin makes just such a cable and I even purchased one at my local Apple Store here in Raleigh, NC.
    Reply
  • joelypolly - Monday, July 13, 2009 - link

    I have actually had something similar happen to a socket I was working on. It was a matter of finding a sewing needle and moving each "pin" back to the original position. Reply
  • HilbertSpace - Monday, July 13, 2009 - link

    It would be interesting to try swapping the 2-socket tray with a 1-socket Mac Pro, and see if it works(?) Would be cheaper to buy the 2-socket board and upgrade yourself, no? Reply
  • MonkeyPaw - Monday, July 13, 2009 - link

    Are FB-DIMMs going to disappear from the market? While at first it doesn't sound Mac-related, original MacPro owners might soon be running out of memory upgrade options (though I doubt they've held out this long to upgrade). It wasn't cheap to start with, but it seems like it was Band-Aid technology. The IMC was the answer, but FB-DIMMs were a stop-gap until Nehalem-Xeons could arrive. Perhaps a memorial article for the technology is needed? Reply
  • JimmiG - Monday, July 13, 2009 - link

    Ok so I get it, even the "cheap" Mac Pro uses a Xeon, not an i7... But for all intents and purposes, it's an i7 920.

    Who in their right mind would pay $2,500 for a i7 920 system with 3GB of RAM, 640GB HDD and a rebranded Geforce 9500 GT? You can build a similar PC (or hackintosh) with the same specifications for the a fraction of the price - in fact you could also bump the RAM to 6GB and throw in a 1TB drive and a 4870 1GB or 4890 if you wanted and still stay *well* below that price point, even if using quality components and case.

    The Mac Pro isn't even shiny!
    Reply
  • plonk420 - Tuesday, July 14, 2009 - link

    did you read page 10? Reply
  • MrDiSante - Thursday, July 16, 2009 - link

    Did you read his comment? Reply
  • ltcommanderdata - Monday, July 13, 2009 - link

    Another great in depth review. Your experiences with upgrading the processors were particularly interesting although I don't think it'd be something I would try.

    I just wanted to suggest you Boot Camp the Mac Pro and run the benchmarks needed to add 2x2.26GHz Gainestown and 2x2.93GHz Gainestown results to the Anandtech Bench. It might also be interesting to get a sample of the new nVidia GTX285 Mac Edition. It would certainly address the 1GB of VRAM concerns and would be cheaper than getting the HD4870 if you need 2 dual-link DVI ports since you don't need to buy that finicky adapter. There really aught to be DVI to mini-DP adapters though for people who still want to use the 24" LED display.

    http://www.glbenchmark.com/result.jsp">http://www.glbenchmark.com/result.jsp

    Oh and for interest, there turns out to be a 3D benchmark comparing the various iPhones to other cell phones. It's called GLBenchmark and needless to say, the iPhone 3G S is a screamer. They are also detecting the iPhone 3G S GPU as a PowerVR SGX 535.
    Reply
  • ddobrigk - Monday, July 13, 2009 - link

    Actually, the Nehalem-EX's octo-core possibility is a no-go for now. It is a future product and has not been launched yet.

    Also, a little bit of nitpicking, but it won't use LGA1366 like these Xeons, it'll use LGA1567, because each CPU will sport a 4-channel memory controller.

    In addition, it'll sport 4 QPI links, and its intended target are 4-way and 8-way systems, not really 2-way systems. A few rumors exist about some integrators being interested in 2-socket systems, though we're still a few months from actually seeing any LGA1567 motherboard on display, AFAIK. All we saw was an Intel Demo about it.

    Don't know if Apple intends to go with 2-socket nehalem-exs, anyway, because when Nehalem-EX really hits the market, there'll also be the 6-core westmeres, I think. In any case, we're way beyond a reasonable number of cores for the typical user. :D
    Reply
  • BrianMCan - Friday, July 17, 2009 - link

    MacPro's really aren't meant for typical users ;)
    Scientific, Video/Movies, 3D, and advanced users who may do many things including the already mentioned, or many things at once. Always other things I can be doing while some video is rendering, including playing some Civ 4, or starting the next video project, researching upgrades & repairs for customers, stuff like that.

    Although I personally may wait for the 2nd gen Nehalem MacPro's before I upgrade from my first gen MacPro, other than raw processing power, it does most of what I need efficiently enough.
    Reply

Log in

Don't have an account? Sign up now