CES 2014: more details on Mullins/Beema

[Lonyo]

Android works on x86 just fine.

Oh yeah? Prove it.
Link to the x86 mobile phones and tablets running Android.

Oh wait, there are quite a few, and the number is growing. Oops. Guess someone missed the last 12 months and CES.
And yes, they are even shipping and available to buy. And some are even made by people with their own fabs, like Samsung.

Crazy. What is the world coming to?

 

[Abwx]

Easy. Perf/TDP.
Most useless metric ever.

First time i read that perf/watt is a useless metric but seems
that for AMD bashing purposes some people are quietly
re definying or dismissing metrics used by the very same
people that design the devices discussed on this thread.

Indeed that s just hollow word with no explanation to explain
why it is good or bad , actualy we are supposed to take at face value unsubstancied claims motivated firstly by emotionaly
unfavourable prejudices....

 

[NTMBK]

Oh yeah? Prove it.
Link to the x86 mobile phones and tablets running Android.

Oh wait, there are quite a few, and the number is growing. Oops. Guess someone missed the last 12 months and CES.
And yes, they are even shipping and available to buy. And some are even made by people with their own fabs, like Samsung.

Crazy. What is the world coming to?

I'm confused.

 

[lefty2]

AMD already have a few design wins for both Beema and Mullins, so they might be on the market sooner than you think.

 

[inf64]

Actually the new 2 slides state in no uncertain terms it's performance (Mullins and Beema vs BT-T/M). Perf./watt is stated below the slides in reference to previous generation which is Kabini/Temash.

So Mullins 4C (TDP 4.5W/SDP 2.5W) scores ~20% higher in PCmark8 and 2.5x higher in 3dmark11 (100% is BT-T z3770 in both cases).
Similar goes for (15W?) Beema part: 25-30% better (unclear what is the exact #) in Pcmark8 and 3.6x better in 3dmark11 versus BT-M N3510.

 

[Abwx]

Similar goes for (15W!) Beema part: 25-30% better (unclear what is the exact #) in Pcmark8 and 3.6x better in 3dmark11 versus BT-M N3510.

That s actualy the 15W Beema compaired to the 25W Kabini
according to the slide on your first post for the claimed
2x perf/watt improvements.

 

[inf64]

I already said that in my previous post

Perf./watt is stated below the slides in reference to previous generation which is Kabini/Temash.

 

[Abwx]

I already said that in my previous post

Corrected your quote accordingly in my post.

 

[mikk]

AMD already have a few design wins for both Beema and Mullins, so they might be on the market sooner than you think.

They are telling this every time. When AMD launches a mobile product it needs months until we really can buy the notebooks or tablets.

 

[ams23]

First time i read that perf/watt is a useless metric

AMD is showing measured performance divided by a rated TDP. That is a pretty silly way to show perf. per watt. The only proper way to show perf. per watt is to measure both performance and power consumed to get that performance.

 

[inf64]

AMD is showing measured performance divided by a rated TDP. That is a pretty silly way to show perf. per watt. The only proper way to show perf. per watt is to measure both performance and power consumed to get that performance.

In the slide where they compare Beema/Mullins with BT-T/M it's just straight performance.
Perf./watt is in regards to previous gen. (Kabini/Temash).

 

[frozentundra123456]

In the slide where they compare Beema/Mullins with BT-T/M it's just straight performance.
Perf./watt is in regards to previous gen. (Kabini/Temash).

No, look carefully at the axis. It says performance "normalized". Doesn't say by what, but I assume it is by tdp.

 

[Abwx]

AMD is showing measured performance divided by a rated TDP. That is a pretty silly way to show perf. per watt. The only proper way to show perf. per watt is to measure both performance and power consumed to get that performance.

Do you even read what you re confidently writing.??.

They re saying that the performance is measured at max TDP ,
how can it be otherwise since they are specifying the MAX PERFS.?

So read that the scores are for 15 watts thermal dissipation , actualy
it s very clear , dont brand something silly because you didnt understand
the obvious.

 

[Abwx]

No, look carefully at the axis. It says performance "normalized". Doesn't say by what, but I assume it is by tdp.

Normalized means that the reference is set as 100% ,
that is as an indice, for perfs comparisons....

Same as the other post , in short...

 

[Enigmoid]

First time i read that perf/watt is a useless metric but seems

How many times does it have to be stated on this forum that TDP != power consumption.

Actually the new 2 slides state in no uncertain terms it's performance (Mullins and Beema vs BT-T/M). Perf./watt is stated below the slides in reference to previous generation which is Kabini/Temash.

So Mullins 4C (TDP 4.5W/SDP 2.5W) scores ~20% higher in PCmark8 and 2.5x higher in 3dmark11 (100% is BT-T z3770 in both cases).
Similar goes for (15W?) Beema part: 25-30% better (unclear what is the exact #) in Pcmark8 and 3.6x better in 3dmark11 versus BT-M N3510.

Because they are comparing the absolute worst Kabini SOC in terms of perf/tdp? A6-5200 on the CPU for instance gains 500 mhz (33%) for +10W (67%) more power vs the 1.5 ghz A4-5000 model. GPU gain is even smaller.

Other things to watch out for.

Beema and Temash RAM speeds are different.

Different drivers (september driver really?).

And they are stating perf/TDP not perf/watt.

 

[frozentundra123456]

Normalized means that the reference is set as 100% ,
that is as an indice, for perfs comparisons....

Same as the other post , in short...

Read the footnote. It clearly states that the performance value is divided by TDP.

 

[Abwx]

How many times does it have to be stated on this forum that TDP != power consumption.

Because they are comparing the absolute worst Kabini SOC in terms of perf/tdp? A6-5200 on the CPU for instance gains 500 mhz (33%) for +10W (67%) more power vs the 1.5 ghz A4-5000 model. GPU gain is even smaller.

Other things to watch out for.

Beema and Temash RAM speeds are different.

Different drivers (september driver really?).

And they are stating perf/TDP not perf/watt.

Beema has nothing to do with Temash , this latter is replaced
by Mullins for fanless design , Beema is Kabini s replacement.

10W delta for 500MHz increased frequency is not that incredible.

Let s start with 15W for 1.5GHz , 33% higher frequency will
increase TDP by the same amount up to 19.95W , from there
if the circuit need just 12% more supply voltage to sustain
this higher frequency then thermal dissipation will increase
as the square of this voltage variation to end increasing the
thermals by 5 more watts up to the final 25W.

Add the necessary voltage margin to make sure all chips
will run at higher specs if ever it s not enough....

 

[inf64]

No, look carefully at the axis. It says performance "normalized". Doesn't say by what, but I assume it is by tdp.

Yeah normalized for BT-T and BT-M as these are at 100%. Really, is that so hard to see?
Lisa Su even explains it in the video (full time video is available at YT). She says 20% higher x86 performance than that xx intel product and 2.5x the GPU performance of that same intel product.

 

[Abwx]

Read the footnote. It clearly states that the performance value is divided by TDP.

Yes , and.?.

How do you measure perf/watt, if thoses words have
a meaning for you.?.

Doesnt this expression explicitely say that perf/watt
is to divide the performance by the thermal dissipation.?.

 

[Enigmoid]

Beema has nothing to do with Temash , this latter is replaced
by Mullins for fanless design , Beema is Kabini s replacement.

10W delta for 500MHz increased frequency is not that incredible.

Let s start with 15W for 1.5GHz , 33% higher frequency will
increase TDP by the same amount up to 19.95W , from there
if the circuit need just 12% more supply voltage to sustain
this higher frequency then thermal dissipation will increase
as the square of this voltage variation to end increasing the
thermals by 5 more watts up to the final 25W.

Add the necessary voltage margin to make sure all chips
will run at higher specs if ever it s not enough....

Sorry about the typo. I meant Mullins and Temash.

Just saying that comparisons must be at the same power level (approximation with TDP is much less accurate but doable). Compare Intel mobile SKU's. Dual 35W part vs. Quad 35W part vs. Quad 45W part. Where do you think the greatest efficiency is even though they are all on the arch?

(Offtopic the relatively high power use of 128 GCN cores at 600 mhz is why I'd argue that nvidia has no chance of getting 192 kepler cores @ 950 mhz in a 4W window).

AMD is comparing against TDP (though the same core config). Its a poor way to do unless you are looking at the same TDP levels (really CPU power draw should be used).

Just compare the A10-6800k to the a10-6700. Nowhere near the same efficiency.

 

[frozentundra123456]

Yes , and.?.

How do you measure perf/watt, if thoses words have
a meaning for you.?.

Doesnt this expression explicitely say that perf/watt
is to divide the performance by the thermal dissipation.?.

And... it is not measured watts, it is an arbitrary TDP number. The slide even says the performance is "simulated".

 

[inf64]

The last slide I have linked is from APU13 conference that was back in November... They comapred old AMD stuff with new AMD stuff, they used perf./watt metric. They still state this in the bottom of the 1st slide (old AMD vs new AMD low power APU).

The 1st slide is from AMD's CES 2014 press event they held today. The metric is pure performance and the slide can be seen in the video in which Lisa Su states it's pure performance. BT is the one they normalized to, it sits at 100% in x86 and GPU workloads.

 

[sefsefsefsef]

I think we should start using slide numbers in our discussion.

So we can say that slide 27 from CES 2014 clearly states that the AMD systems have better *performance* than Intel systems, and do not bring any notion of power or perf/power into the discussion.

 

[Abwx]

And... it is not measured watts, it is an arbitrary TDP number. The slide even says the performance is "simulated".

Max perf is obtained at max TDP, isnt it...?

Why would they limit TDP when measuring the perfs of their chips.?.

You mean that a 15W device perfs should be the measured
with a 10W comsumption.?.

btw , the performance is on real silicon , if you had really read
the slide you would have noticed the set up...

 

[inf64]

I think we should start using slide numbers in our discussion.

So we can say that slide 27 from CES 2014 clearly states that the AMD systems have better *performance* than Intel systems, and do not bring any notion of power or perf/power into the discussion.

That is correct with regards to intel systems. The slide also mentions Beema/Mullins having 2x more perf./watt than previous gen. of AMD products.