Intel's Fastest Processor Ever.....

[Hans de Vries]

Intel's Fastest Processor Ever.....and the Burst Mode Trick.

The tiny "4.5W" Core M seemingly humiliates the previous record holders,
including the Core i7 4790 Devil's Canyon running at a burst speed of 4.4GHz.

source

source


I guess we will see many claim that Intel's new 14nm process is now 20
times more power efficient as Intel's 22nm process or similar wild claims.
(Even though Intel itself never made such outrageous claims of course)


Those who know how Intel's Burst Mode work will understand how you
can produce these kind of miraculous benchmark result. They will wonder
instead if the new F-stepping can now officially be boosted to 4+GHz
for very short periods of time.

Here's explained how it works:

source

The figure at the bottom right with the three blue spikes explains it all.
The blue spikes mean that the processor is running at maximum frequency
while during the intervals the processor is basically halted.

Say with a single active core Broadwell at 14nm, running a 4+GHz requires
~18 Watt (would be very good!).

A power dissipation of 4.5 Watt would mean that the processor is bursting
at maximum frequency during 25% of the time and idle during 75% of the
time to cool down.

You can't measure this at the outside of the package because the on-
package voltage regulators and capacitors take care of the large power
and current spikes. From the outside you only see a processor using
4.5 Watt.

The benchmark can't see this either. It measures the "process-time"
when the processor is active and running at maximum frequency.
The benchmark doesn't measure the time when the processor is idle
and cooling down. Therefor you get a benchmark result as if the
processor is running always at maximum burst frequency.


Hans.

 

[monstercameron]

a benchmark god...

 

[Enigmoid]

I'm not sure if you are serious. Its because the Broadwell tablet is running IE.

On my 3630qm laptop

Chrome 37 - 180 ms
Firefox 31 - 178 ms
IE 9 - 103.7 ms

 

[monstercameron]

I'm not sure if you are serious. Its because the Broadwell tablet is running IE.

On my 3630qm laptop

Chrome 37 - 180 ms
Firefox 31 - 178 ms
IE 9 - 103.7 ms

there is an order of magnitude lower tdp between them, 4.5 v 45W

 

[Abwx]

The tiny "4.5" Core M seemingly humiliates the previous record holders,
including the Core i7 4790 Devil's Canyon running at a burst speed of 4.4GHz.

They didnt use the standard chrome version that is used in other tests, the comparison in a same graph is like mixing scores of CB R15 with CB 10 scores, and yet this is what a lot of sites did...

The Llama Mountain tablet is likely using a newer Chrome release than any of the other systems, which could give it an edge over the rest of the field.

http://techreport.com/news/27039/preliminary-core-m-benchmarks-hint-at-broadwell-potential

Other than that the perf/watt is about the one of Mullins or BTray.

 

[Enigmoid]

there is an order of magnitude lower tdp between them, 4.5 v 45W

And sunspider runs a single thread.

 

[Phynaz]

Take off the tin foil hat Hans. That slide does not show what you talking about at all.

 

[Hans de Vries]

What is the real frequency improvement, at equal power, going from
22nm to 14nm. Intel states this at ~30% or so at the transistor level.
Indeed this is confirmed by recent slides:

For CPU applications 19%, 12% and 11%

Intel told to journalists that the 22nm Core i5 had been down clocked to
lower frequencies to make it operating at 4.5 Watt instead of it's normal
TDP of 11.5 Watt in order to compare at equal TDP (4.5W)

Intel confirms this in the backup sheet (last lines)

Hans

 

[mavere]

Non-sustainable boosts in, say, gaming or video encoding could be a concern, but most Javascript workloads are inherently bursty. For web-browsing, I'm 100% in support of high-frequency microbursts.

 

[IntelUser2000]

What is the real frequency improvement, at equal power, going from
22nm to 14nm. Intel states this at ~30% or so at the transistor level.
Indeed this is confirmed by recent slides:

For CPU applications 19%, 12% and 11%

Nice try. You need the near top Turbo frequency to achieve 19% gain over Core i5 4302Y.

 

[Abwx]

Nice try. You need the near top Turbo frequency to achieve 19% gain over Core i5 4302Y.

It is written "up to 19%", so it s a best case scenario, this suggest that it s generaly lower if we follow your logic.

For the time i dont see where the progress is if we except the smaller dies, it looks like their supposed current advantage that they have at 22nm in respect of the competitors 28nm is going to be seriously entailed once they get to just 20nm.

 

[Hans de Vries]

I'm not sure if you are serious. Its because the Broadwell tablet is running IE.

On my 3630qm laptop

Chrome 37 - 180 ms
Firefox 31 - 178 ms
IE 9 - 103.7 ms

That would make sense because we'll probably stay with a burst frequency
of 2.6GHz then (Of course with a way higher power dissipation during the
burst than 4.5W)

Not so good to throw incomparable results from different browsers
in one table ...

It would also mean that it would score ~200ms and higher on the other
browsers and 470ms and higher on its base frequency of 1.1GHz which
is much more representable of sustained operation.



On my 3630QM laptop (3.4GHz burst) I get this with the Browser's on my
system.

_99.9 ms Internet Explorer 11.011
184.0 ms Firefox 32
180.0 ms Chrome 37
187.0 ms Opera
193.0 ms Safari 5.1.7
274.1 ms Chrome 39 (64 bit Canary)

Hans

 

[Hans de Vries]

Non-sustainable boosts in, say, gaming or video encoding could be a concern, but most Javascript workloads are inherently bursty. For web-browsing, I'm 100% in support of high-frequency microbursts.

Yes, I do agree with this.

 

[Khato]

That would make sense because we'll probably stay with a burst frequencyof 2.6GHz then (Of course with a way higher power dissipation during the
burst than 4.5W)

While there's no question that the performance Core M achieves is due to its usage of turbo modes to exceed the 4.5W TDP for short durations I have to disagree that a frequency of 2.6GHz on a single threaded load would result in way higher power dissipation than 4.5W and that high turbo speeds on CPU-only loads aren't sustainable. It's quite easy even with just a desktop Haswell system to see how power efficient the CPU cores can be at these frequencies - reducing frequency and voltage of my i7-4770k to 2.6 GHz and 0.875V has it going from 38W idle to 64W running cinebench at the socket, which works out to all of 6.5W per core. Then we know that Core M shows at least a 30% reduction in power compared to Haswell-Y, so that would cut it down to 5W. Or if I clock down to 2.4 GHz and 0.8V it drops to 37W idle and 57W running cinebench, so 5W per core while still getting 5.26 points.

All that said, we really need to more information to know exactly where Core M power consumption stands. It's somewhat telling that Intel doesn't have some of their reference tablets wired up to provide power readings on external equipment like they did with Baytrail last year. I wouldn't be surprised if it settles somewhere in the 2-2.2 GHz range for sustained (aka 4.5 watt SoC power dissipation) multi-threaded CPU-only workloads. Then drops down to the base frequency or below when it comes to sustained gaming workloads where the GPU gets the majority of the power budget.

 

[Blue_Max]

The new ultraportable Intel chip was supposed to introduce new graphics too... supposed to be above Iris?

 

[Burpo]

In an Entertainment Dock w/fan
"able to be overclocked.. up to 40%"
https://www.youtube.com/watch?v=znKgvPk3-uc

 

[Dufus]

You can't measure this at the outside of the package because the on-
package voltage regulators and capacitors take care of the large power
and current spikes. From the outside you only see a processor using
4.5 Watt.

You can measure unhalted and halted clocks with the performance counters and apply them accordingly.

Intel told to journalists that the 22nm Core i5 had been down clocked to
lower frequencies to make it operating at 4.5 Watt instead of it's normal
TDP of 11.5 Watt in order to compare at equal TDP (4.5W)

IMO it would have been better if they were both compared at their defaults to emphasize the power savings of the newer processor.

For instance a car manufacturer making an economical car with 1BHP engine would possibly show how good mileage is, not take a Ferrari and set it to also run at 1BHP so it runs like a dog then possibly even say the economical car runs up to 50% faster than a "comparable" Ferrari.

Advertise it for what it is, an excellent low powered CPU with great battery life. Wonder how many people are going to miss the fine print and expect it to run 50% faster than the i5-4302Y it was compared to.

 

[TurtleCrusher]

This is definitely something to look forward to. Pure speculation, but I foresee a netbook revolution 2.0 if this if the case.

 

[USER8000]

AFAIK,the Intel reference tablet is also a 13" one compared to much smaller volume and TDP limited competitors.

Hopefully we will have some actual retail ones in the next few months,so we see some more in-depth tech site and user reviews.

This is definitely something to look forward to. Pure speculation, but I foresee a netbook revolution 2.0 if this if the case.

Atom is the Intel netbook and low cost SOC. This is more orientated towards the performance and higher cost area.

Its basically Intel doing what other ARM based SOC makers are doing with high and low end options. Makes sense.

 

[Nothingness]

I'm not sure if you are serious. Its because the Broadwell tablet is running IE.

On my 3630qm laptop

Chrome 37 - 180 ms
Firefox 31 - 178 ms
IE 9 - 103.7 ms

That's exactly why Javascript benchmarks are useless to compare CPU, and Sunspider is the worst of all.

 

[Piroko]

Intel told to journalists that the 22nm Core i5 had been down clocked to
lower frequencies to make it operating at 4.5 Watt instead of it's normal
TDP of 11.5 Watt in order to compare at equal TDP (4.5W)

That made me listen up and do a small experiment for fun:

I limited my i7-3517u to 5W TDP at the SOC level via Throttlestop and ran Cinebench R15. The CPU ran at full speed for the first 3~4 seconds until it limited itself to 800 MHz and a fairly stable 5.4W SOC TDP.
Its normal CB score is 240~245, anyone take a guess on the throttled score?

 

[SPBHM]

That's exactly why Javascript benchmarks are useless to compare CPU, and Sunspider is the worst of all.

if you use exactly the same software (which is possible for the Core M and other x86 CPUs) it's OK, but yes, with different softwares it's hard to see it as a CPU benchmark.


here:
Firefox 31 188ms
IE 10 117ms

 

[Enigmoid]

That made me listen up and do a small experiment for fun:

I limited my i7-3517u to 5W TDP at the SOC level via Throttlestop and ran Cinebench R15. The CPU ran at full speed for the first 3~4 seconds until it limited itself to 800 MHz and a fairly stable 5.4W SOC TDP.
Its normal CB score is 240~245, anyone take a guess on the throttled score?

40% of that.

 

[sm625]

There seems to be more to sunspider than just the cpu. I have a haswell i7-4800EQ here in my lab and it scores around 85mS on sunspider. I cannot explain how it scores so well given its rather modest clock speed. All I can say is it is a custom design, it is a COM Express SBC.

 

[JoeRambo]

The benchmark can't see this either. It measures the "process-time"
when the processor is active and running at maximum frequency.
The benchmark doesn't measure the time when the processor is idle
and cooling down. Therefor you get a benchmark result as if the
processor is running always at maximum burst frequency.

Great conspiracy theory, but in reality to complete benchmark you need to execute all code involved. Benchmark measures absolute time @ start and @ end ( or same for some internal partitions of benchmark that get summed/weight averaged etc to produce final magic score). And absolute time is independant of CPU clock or idleness.

To get to the bottom of this problem you need to compare two approaches:

1) Done by morons from vendor marketing - detect that benchmark code is running, know from previuos experience that it will take X seconds and require Y instructions, take your thermal budget of Z joules and rise clocks for benchmark since you know you won't bust your thermals executing Y in X seconds. Great for benchmark scores, no impact in real world.

2) Done by Intel - know your thermal budget of Z joules, know your total battery power cap of Y, dynamically rise clocks for any load by turbo ( hoping of course that it is a race to sleep or some short benchmark) to stay within Y as long as your long term thermal load is not busting Z. By watching temps and managing clocks in real time they benefit all loads. In fact it can be argued that it benefits light loads the most and punishes conspiracy theorist loads of Linpack the most


So in the end TDP is just that, power level that cooling solution has to dissipate in long term, there is no magic requirement for it to limit power consumption to certain value.



P.S. And all those browser benchmarks are a joke, only same browser version on same OS is comparable as it can have order of magnitude of impact on same CPU ( IE8 vs Chrome 64bit)