Best performance per watt

[gorion]

Hi everyone.
I'm curious: right now what are the desktop chipsets with the best performance per watt ratio?

Is Intel still ahead in this respect?

 

[SiliconWars]

The Intel i7's I guess, or an i5 if you don't want to spend the extra cash on hyperthreading. i3's are ok for most tasks as well. Intel is pretty far ahead yeah.

 

[BSim500]

Hi everyone.
I'm curious: right now what are the desktop chipsets with the best performance per watt ratio?

Is Intel still ahead in this respect?

Intel is still way ahead. Up to 90w less TDP i5 3570K vs FX 8350 at same full load...

The best performance-per-watt chipset is probably still the Z77 boards (or Z87 Haswell as long as you don't run AVX benchmarks), then buy an i5 / i7, then undervolt it.

I have an i5-3570 running at 4.2GHz and it still has enough "headroom" for a stable -0.075v undervolt at that speed resulting in a ridiculously cool 51w full 4T load under Prime (nearer 40-45w for regular gaming) that never hits 55c under any condition even with very slow spinning 700rpm silent fans. :thumbsup:

 

[ShintaiDK]

The best performance-per-watt chipset is probably still the Z77 boards (or Z87 Haswell as long as you don't run AVX benchmarks), then buy an i5 / i7, then undervolt it.

Performance/watt, Haswell is just even better in AVX.

In Linpack, 3570 gives around 100Gflop on stock. 4670 around 175Gflop.

 

[Qwertilot]

Presumably that 35W i7 that got mentioned in the CPU round up on the main site must be doing rather well here? (i7 4765T).

It doesn't obviously lose much except being clocked down to 2 GHz base/3 turbo and costing a bit more. Not sure if there is any detailed benchmarking out there, but the simple figures here: http://www.quietpc.com/haswell-i7 seem to have it comfortably ahead of its i7 brethren.

 

[BallaTheFeared]

As IDK pointed out to me and I confirmed myself through my own testing, at a certain point you lose perf/w due to fixed platform power with slower chips.

My chip in particular is actually peaking in efficiency around 4.4GHz.

 

[AtenRa]

I believe that Watt per Performance is a much better metric. It is way better to know the power consumption at the same performance than the performance per Watt.

Measuring performance per watt only tells you hafe the story. Core i3 has higher performance per watt but im sure it will have way worst Watt per Performance against Core i7 in multithreaded workloads.
For example take the performance in x264, Core i3 will need a substantial raise of its frequency to get to the Core i7 performance level. To do that it will need more voltage + Frequency and thus it will completely destroy its performance per watt.
On the other hand, Core i7 will significantly lower its frequency and voltage to get to the same Core i3 level of performance and thus will raise its Performance per Watt significantly.

For me, measuring performance per watt is meaningless for real world applications. The same CPU will exhibit different performance per watt depended on the performance it will output every time. So, in order to know which CPU is the more efficient and by how far, measuring the power consumption at the same performance or measuring the performance at the same power consumption is the only way.

 

[gorion]

AtenRa, i must confess that in my mind performance per watt means what you wrote.

For a given FPS level and a given resolution which GPU eats less power?

Same for CPU related tasks..

 

[Qwertilot]

Well, measured like that, it'll depend on what your task is

For tiny tasks a tiny processor but they obviously won't scale up.

For a medium/mixed task like typical desktop use and a bit of moderate gaming, then I guess it gets interesting. All that turbo/scaling down etc on intel's stuff.

Intriguing if those T series i5/i7's actually make a real difference in practice for that sort of thing. From the little I could see on the 'net the answer was mostly no for Sandy Bridge ones, but maybe that has changed a bit with Haswell being so focused on low power operation/notebooks etc.

 

[frozentundra123456]

I was confused about AtenRa's post too, but I guess it makes sense. He said it in a kind of way that to me was hard to understand. I would still look at the metric as performance per watt, but that would vary with the workload. I think that is what Aten was saying as well.

 

[ShintaiDK]

I believe that Watt per Performance is a much better metric. It is way better to know the power consumption at the same performance than the performance per Watt.

Measuring performance per watt only tells you hafe the story. Core i3 has higher performance per watt but im sure it will have way worst Watt per Performance against Core i7 in multithreaded workloads.
For example take the performance in x264, Core i3 will need a substantial raise of its frequency to get to the Core i7 performance level. To do that it will need more voltage + Frequency and thus it will completely destroy its performance per watt.
On the other hand, Core i7 will significantly lower its frequency and voltage to get to the same Core i3 level of performance and thus will raise its Performance per Watt significantly.

For me, measuring performance per watt is meaningless for real world applications. The same CPU will exhibit different performance per watt depended on the performance it will output every time. So, in order to know which CPU is the more efficient and by how far, measuring the power consumption at the same performance or measuring the performance at the same power consumption is the only way.

That just nonsense, because an i7 wouldnt lower its frequency.

You can measure 2 things. Performance/Watt for the CPU or Performance/Watt for the entire system. The performance would be a workload case. For example encoding a movie or running a game. When task is done, you look at energy used over the timeframe.

This is also why people get fooled by T/S CPUs, because they dont understand how they work.

 

[AtenRa]

http://www.anandtech.com/show/7318/intel-demos-14nm-broadwell-up-to-30-lower-power-than-haswell

Intel just demonstrated 14nm Broadwell ULT (Y-series) silicon, normalized for performance against 22nm Haswell ULT (Y-series) silicon running a multithreaded Cinebench test. Intel was monitoring SoC power during the benchmark and demonstrated a ~30% reduction in power, at the same performance level.

Also, Broadwell Cinebench Demo. Normalized performance shows 30% power reduction.
http://www.youtube.com/watch?v=KeDtXucTwRI

 

[ShintaiDK]

http://www.anandtech.com/show/7318/intel-demos-14nm-broadwell-up-to-30-lower-power-than-haswell



Also, Broadwell Cinebench Demo. Normalized performance shows 30% power reduction.
http://www.youtube.com/watch?v=KeDtXucTwRI

Yes, 14nm vs 22nm.

 

[Blue_Max]

Surprised no one's mentioned Intel's T or S series.

These processors drop the speed a little, but the wattage a lot!

Mind you, AMD has a few half-decent options, like the same speed but no overclocking for 65W instead of using overclocking and 100W+

 

[Torn Mind]

Surprised no one's mentioned Intel's T or S series.

These processors drop the speed a little, but the wattage a lot!

Mind you, AMD has a few half-decent options, like the same speed but no overclocking for 65W instead of using overclocking and 100W+

Maybe for the i7 or i5 "T" and "S" chips, but not really for the lower end chips like Celerons and Pentiums.
All current chips have Speedstep to downclock the CPU to the lowest multiplier. The T and S chips might be run at a lower voltage straight out of the factory though. And undervolting is available on many boards.

 

[sniffin]

Yeah undervolting makes T and S chips moot. I know Ivy Bridge stock voltages were ridiculously high (I think my 3770k stock voltage was 1.15v, when it could run stock frequencies at under 0.95v), I'm assuming the same is true with Haswell.

 

[Puppies04]

I think I get what AtenRa is saying.

Assuming all chips are unlocked I can either want to use as little energy as possible to encode a 2 hour HD movie (per/watt) or I can say I need this operation done in 20 minutes so how much power does each chip need to be overclocked/underclocked to hit this target (watt/per)

 

[BallaTheFeared]

That's how it should be looked at, that is what perf/w is...

Perhaps the thought was lost in translation?

Confuses me since I'm not sure how else to draw a perf/w based conclusion other than through a fixed operation where you measure time to completion, and watts used to formulate your joules and thus figure out the actual performance per watt.

2.8GHz = 75 Joules /s over 659 seconds, so 75 x 659 = 49425 Joules to complete the task


4.3GHz = 116 Joules /s over 429 seconds, so 116 x 429 = 49764 Joules to complete the task


4.8GHz = 178 Joules /s over 381 seconds, so 178 x 381 = 67818 Joules to complete the task

Looking at simply performance and then figuring out perf/w at the same performance is what AtanRa is saying, it is a different discussion than perf/w.


Perf/w is separate from raw performance, because you are attempting to optimize for the lowest power consumed to complete a task, without defining the condition based on how long it takes to complete said task (this would be performance/raw performance).

 

[AtenRa]

I think I get what AtenRa is saying.

Assuming all chips are unlocked I can either want to use as little energy as possible to encode a 2 hour HD movie (per/watt) or I can say I need this operation done in 20 minutes so how much power does each chip need to be overclocked/underclocked to hit this target (watt/per)

Yeap, and the highest the performance the lower your power efficiency becomes.

Example (numbers are not real) :

Lets say that Quad Core ATOM Baytrail takes 20 hours to encode a video using 2Watts per hour. That means it will use 40Watts to finish the job.

Core i7 4770 Haswell needs 2 Hours to finish the same encode, that means it will consume 2x 87Watts = 174Watts.

Now, if you calculate the Performance per Watt you will find that ATOM has higher performance per Watt than Haswell Core i7. That is true but, ATOM Baytrail is also 10x slower than Haswell.
Now if you try to make the ATOM decode the same Video in 2 hours, same time as Core i7, it may have lower Performance per Watt than Haswell and it may even be impossible to reach the same performance due to technical limitations.
Also, if you make the Core i7 decode the same video in 20 hours(lower frequency, lower Voltage etc) it may even have higher Performance per Watt than ATOM.

 

[Torn Mind]

Perf/w is a ratio and many combinations can produce the same number. It is a common mistake to interpret ratios in a manner in which the amount of its constituent variables always do not matter.

 

[Torn Mind]

Yeap, and the highest the performance the lower your power efficiency becomes.

Example (numbers are not real) :

Lets say that Quad Core ATOM Baytrail takes 20 hours to encode a video using 2Watts per hour. That means it will use 40Watts to finish the job.

Core i7 4770 Haswell needs 2 Hours to finish the same encode, that means it will consume 2x 87Watts = 174Watts.

Now, if you calculate the Performance per Watt you will find that ATOM has higher performance per Watt than Haswell Core i7. That is true but, ATOM Baytrail is also 10x slower than Haswell.
Now if you try to make the ATOM decode the same Video in 2 hours, same time as Core i7, it may have lower Performance per Watt than Haswell and it may even be impossible to reach the same performance due to technical limitations.
Also, if you make the Core i7 decode the same video in 20 hours(lower frequency, lower Voltage etc) it may even have higher Performance per Watt than ATOM.

Actually, any measure in which decreasing values signify an increase in performance requires a caveat when interpreting them: A LOWER ratio means a "better" ratio value. In fact, it is better to use negative numbers with these ratios because a "bigger" negative number value means the value is smaller.

If frames per second were used instead, Haswell would come out on top. If Bay Trail can encode 200 fps and Haswell 600 frames per second, Haswell will have a much better performance/watt ratio. (In addition, the units make sense if we divide f/s by J/s, giving us frames per joule)

 

[Essence_of_War]

That means it will use 40Watts to finish the job.

Strictly speaking, 40 Watt*hrs, or for more grok-able units: .04 kW*hrs.

But your point is clear

 

[gorion]

Thanks everyone for the interesting points made so far.

If i get it right the best CPU for daily usage and power conscious users might be an undervolted i5. Am I correct?

 

[Abwx]

What about the perf/watt if one use his IGP for a few gaming.?.

 

[aigomorla]

Thanks everyone for the interesting points made so far.

If i get it right the best CPU for daily usage and power conscious users might be an undervolted i5. Am I correct?

undervolting vs stock power differences...

So lets look at the differences shall we...

Stock processor 1.275V = 84W TDP ~ 66 Amps

Were gonna hold the Amps constant while changing the voltage.

You lower the voltage to like 1.1 x 66Amps ~ 72.6W

84W - 72.6 = 11.4W savings... in 100 hours, u will hit the 1kw metric for rates. I believe which is arround ($.05-$.15) savings every 100 hours the machine is on..

:T

Leave it at stock... leave the power saving options native to the cpu's... as they do downvolt automatically when CPU is not required by OS.
And those values are probably lower then your downvolting settings.
Also playing with overclocking settings can disable these power saving features on the CPU, or cause your CPU to have problems when trying to downvolt the overclocked settings.
So if ur really conscience about power.... leave it at stock.. leave power profile on conservative, and then forget it....

I dont recommend downvolting any 3rd or 4th or 5th generation Core.

Yeap, and the highest the performance the lower your power efficiency becomes.

thats because i believe it becomes multiplicative due to the higher loss as HEAT at the faster speed...
Also the IC's are more efficient and faster at lower temperatures.

You guys forget... the faster you push your CPU, what happens?? it becomes hotter.
Heat is a waste product not a performance indicator.