Core i7 vs. Core i5: Understanding the Power Story

Between generations Apple constantly struggles between squeezing every last ounce of max performance out of silicon and reducing system temperatures. I believe Apple's philosophy here is that most of the time your CPU should be running at relatively low utilization and as a result offering the full dynamic range of CPU performance is preferred to clamping max performance in order to preserve lower thermals. The problem is that in some cases, lazy background task management (e.g. keeping too many Safari windows open with Flash active) can drive CPU usage and thermals up even if you're actively doing nothing on the machine. This scenario coupled with Haswell ULT's excellent idle power consumption I believe are primary motivators for Mavericks' App Nap and occluded window slumber features.

 

 

To understand the impact on thermals (and battery life) of the Core i7-4650U on the 13-inch MacBook Air you need to understand what's going on under the hood. To hit higher frequencies, the i7-4650U generally requires a higher voltage. Power consumption (and thus thermal dissipation) can scale linearly with frequency, but it scales quadratically with voltage. The combination of the two is quite possibly the worst case scenario from a power consumption standpoint. This is why it's generally always best to increase performance via process shrinks or architectural enhancements vs. simply scaling frequency. In the case of the i7-4650U we're not talking about huge frequency/voltage scaling here, but rather a tradeoff between added performance and increased power consumption. In the table below I noted typical CPU core voltages for a couple of different operating modes on my i5-4250U and i7-4650U samples. Several years ago Intel introduced voltage binning even at a given frequency, so the voltages you see in the table below are only applicable to my parts (or other similar parts) - you could see a range of acceptable voltages in other binned parts even carrying the same model number. The values in parantheses indicate the CPU frequency (or frequencies) observed during the workload.

13-inch MacBook Air (Mid 2013) CPU Comparison - Observed Voltages
  Idle Cinebench 11.5 (1 thread) Cinebench 11.5 (4 threads)
Intel Core i5-4250U 0.665V
(800MHz)
0.852V - 0.904V
(2.3GHz - 2.6GHz*)
0.842V
(2.3GHz)
Intel Core i7-4650U 0.655V
(800MHz)
0.949V - 1.041V
(2.9GHz - 3.3GHz*)
0.786V - 0.949V
(2.8GHz - 2.9GHz*)

There are a bunch of observations here. First off, the two parts are very comparable at idle - this is how Apple can quote all implementations of the MacBook Air as being capable of up to 12 hours of battery life. At idle large parts of the silicon are clock gated if not fully power gated. Idle voltages are extremely low (even compared to what you find in modern smartphones) and both parts run at the same 800MHz frequency at idle, so power consumption is comparable between the two at idle.

Using Cinebench 11.5, I ramped up a FP intensive single threaded workload. FP workloads tend to force a bunch of large units into switching making this a great test for voltage scaling. Here we see that the i5-4250U is capable of hitting its max turbo frequency but for the most part it hangs out around 2.3GHz. The same is true for the i7-4650U, 3.3GHz is possible but most of the time it's sitting down at 2.9GHz. The i7-4650U needs higher voltages all around to hit these higher frequencies.

Next, I cranked up the number of threads. First you'll notice a reduction in clock speeds and voltages. This is where multithreading can actually be good for power consumption. Running more cores at a lower voltage for a shorter period of time can reduce total energy consumed while performing a task. The i5-4250U has no issues running at its max DC turbo frequency (2.3GHz), while the i7-4650U mostly sticks to 2.8GHz with occasional bursts up to 2.9GHz. Note that the 4650U's min voltage at 2.8GHz is actually lower than the 4250U's here. In order to hit these higher frequencies within the same TDP, Intel does have to bin for parts that do a bit better at higher frequencies whereas to make the cut for a 4250U the leakage requirements aren't as severe.

There shouldn't be any surprises thus far, but this data should give us an indication of what we can expect in terms of battery life and thermals. Where the i7 vs i5 comparison becomes tricky is if you look at workloads that can complete quick enough due to the faster performance in order to offset any additional power consumption.

CPU Performance Battery Life & Thermals
POST A COMMENT

127 Comments

View All Comments

  • josephandrews222 - Thursday, July 04, 2013 - link

    Anand--wonderful post. Care to speculate on the battery life and thermals for the Surface Pro 2? Reply
  • teiglin - Thursday, July 04, 2013 - link

    Derp in last paragraph, first page: "The max clocks increase by almost 30%, as does the <s>increase in</s> L3 cache."

    Nothing really surprising here, but good to see it confirmed anyway!
    Reply
  • Subyman - Thursday, July 04, 2013 - link

    I wouldn't call the lower battery life a detriment, remember you have 20% less battery time, but you have 20% more processing speed, therefore you simply get the same amount of work done, but faster. For general surfing, there is no difference between the two but when working its faster. Reply
  • TheinsanegamerN - Thursday, July 04, 2013 - link

    am I the only one more than a little disturbed to see temperatures of over 90c? thats awfully hot.... Reply
  • MrSpadge - Thursday, July 04, 2013 - link

    That's the price you have to pay for a flat laptop, without totally killing your ears. Reply
  • Paapaa125 - Thursday, July 04, 2013 - link

    It doesn't matter what the core temperature is if it is within specs. And I guess 90C is. Reply
  • FwFred - Thursday, July 04, 2013 - link

    This should only matter if the CPU was throttling due to temperature. As a consumer, you should be more concerned with skin temperature and fan noise. Reply
  • solmaker - Thursday, July 04, 2013 - link

    Fantastic job on a fascinating topic. Two follow-up questions:

    1) Would i7 compare any different to i5 on the 11" model, given tighter thermals?

    2) How will the comparison change under Mavericks?

    My own thinking is that if Mavericks improves battery life by (say) another 25% across the board, 2013 MBA battery life will be so terrific (perhaps 12 hours even for the 11" i7 MBA) that 1/2 to 1 hour differences between models will seem like small change. At that point, processing speed would be the more important bottleneck, making the i7 look like an even better choice.
    Reply
  • helloworldv2 - Thursday, July 04, 2013 - link

    My guess is that Maverics increases battery life by 5% max and will inject a ton of new annoyances into OS X (like having to caffeinate everything that you want to keep on running even if you dare to take some other program into foreground). I read somewhere (maybe Gizmodo) that they didn't notice any battery life increase with Mavericks beta though.. Reply
  • KitsuneKnight - Thursday, July 04, 2013 - link

    A 25% increase across the board I'd say is incredibly unlikely. What seems far more likely, though, is making more workloads closer to being a 'light' workloads (so hitting closer to the advertised numbers in more cases).

    Timer Coalescing and AppNap will make it less likely for noisy apps to drain significant battery life when merely sitting in the background doing nothing (including background tabs of Safari, I believe... which'll hopefully include Flash applets on background tabs). This will help improve the likelihood of the CPU being able to hit lower idle states, saving more power.

    In addition, Apple has put effort into making it easier to identify power hungry apps (which aren't necessarily CPU hogs!), both for end users and developers (the battery menu, Activity Monitor, Xcode, and Instruments all have an increased focus on identifying and addressing battery hogs).

    But don't expect miracles like a 25% increase in battery life (I'd be *very* shocked if Apple was able to pull that off). I'd imagine it would be much closer to 5% in practice, but varying drastically depending on the actual workloads (anywhere from 0% if the system can't sleep at all to 50%+ in the extreme case of a horribly buggy app that AppNap neuters... with most clustering around the low end).
    Reply

Log in

Don't have an account? Sign up now