3

Let's say that we have a set workload doing a set amount of work.

If I have two identical servers, one equipped with a E5-2660 and one with a E5-2670 and otherwise identical hardware, will the total amount of power consumed by the system while processing this workload (regarded as work units/second, not a total amount of work to be done) differ?

If we compare these very similar processors we see the following differences:

E5-2660: Clock speed 2.2/3GHz, TDP 95W

E5-2670: Clock speed 2.6/3.3GHz, TDP 115W

It certainly makes sense that we can load up the E5-2670 with more load and it will thus have a higher TDP, but if we're only loading it up to:

  • 0%
  • 1000 transactions/second (arbitrary workload value mentioned above)
  • 100% of the capacity of a E5-2660

then what should I expect the actual wattage draw of the system to be (when compared with the E5-2660 system)?


Available frequencies:

  • E5-2660: 2201000 2200000 2100000 2000000 1900000 1800000 1700000 1600000 1500000 1400000 1300000 1200000
  • E5-2670: 2601000 2600000 2500000 2400000 2300000 2200000 2100000 2000000 1900000 1800000 1700000 1600000 1500000 1400000 1300000 1200000

Turbo limits:

  • E5-2660: Multiplier 99.96MHz
    Maximum turbo limit with 1/2/3/4/5/6 cores active is 30/30/29/29/28/28
  • E5-2670: Multiplier 99.73MHz
    Maximum turbo limit with 1/2/3/4/5/6 cores active is 33/33/32/32/31/31

Running idle with the performance profile and the ondemand governor means both types of processor (on RHEL6) clock down to 1.2MHz and spend 50% of their time in C1.

4

I think that will depend on so many different things that the only reasonable answer is "try it". Some things that could make a difference:

  • Is the branch prediction significantly better in one of them?
  • Is the layout of the pipeline different?
  • Is the size of the L1 cache the same?
  • Do various internal paths and paths to buses differ?
  • Does the OS do power management better for one of them for some reason?

The pages you linked to mention no such details, so set up a lab and try it out :)

  • Well, the pages at least say that the L1 cache is the same. I'd lay good money on the rest of them being the same given that they are very nearly the same processor in the same family, just with different multipliers. – MikeyB Jul 17 '13 at 14:07
4

Even within batches of the same processor, there will be differences in power load.

In the overclocking community it's widely known that you can get "good" CPUs and "bad" CPUs. The bad ones generate a lot more heat under load (even at stock speeds) and thus will fail more quickly, thus limiting how much you can overclock them.

This is generic, standard CPU binning:

All processors start out from the same line. Unlike what some may think, chip makers do not have dedicated lines for processors of different speed grades. Instead, they have a single line from which all processors of the same model are made. Even then, these chips will not turn out exactly the same.

At the end of the fabrication process, some chips will invariably be non-functioning or malfunctioning. Those that work though will not have the same characteristics. Some will be able to run faster while others can only run at a lower clock speed. However, it would take an extraordinarily long time to test each and every chip and discover their maximum potential.

To save time, chip makers use a method called "speed-binning". Instead of testing and selling chips at their maximum clock speed, they test and sell chips at certain speed grades - like 1GHz, 1.2GHz, etc. This is one of the reasons why overclocking is possible.

Suppose a chip maker sells a particular processor at 3 different speed grades - 1GHz, 1.2GHz and 1.4GHz. After every processor is made, it is first tested at 1.4GHz. If it passes, then it's labelled and sold as a 1.4GHz processor. If it fails at that clock speed, it's retested at 1.2GHz and labelled as a 1.2GHz processor if it passes. Otherwise, it's tested at 1GHz. If the processor fails at that speed, then it's discarded.

In addition to that, as I mentioned above, now that all the voltage regulation stuff is on-die, CPU manufacturers can easily tell CPUs "hey, you're not quite working at 1.2 Ghz but we really want you to. Turn up your voltage a little."

Even at the same clocks, Intel may set the on-die voltage to a higher level to compensate for minor manufacturing defects. So if you get a CPU that has a relatively high factory default voltage, it's going to produce more heat and use more power to get to the same clock speeds.

You might be able to view the factory voltage settings of your CPUs via something like CPU-Z.

  • Noted. In this particular case, the load would be across 20,000 processors so variations within a batch should even out, but statistical small differences could potentially have a large impact. Another consideration brought up by someone I asked about this is that if the E5-2670 is clocking up higher yet running for a shorter period of time, the consumption may follow a square law resulting in a net power consumption increase. – MikeyB Jul 19 '13 at 13:31

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.