AMD Bulldozer Dual-Interlagos Benchmarks On Linux

[Schmide]

Since this is a cubic function , i assume that it is composed
of a product of three first degree functions.
First is the current needed to charge/discharge the parasistic
capacitance, wich increase linearly with frequency.
Second should be the cross conduction losses of the push pull
cmos pairs wich increase also linearly with freq.
And third, i can only see the necessary voltage increase as the third factor that raise the final degree to 3...
Is that plausible ?..

I'm sure IDC will describe this better than me but transistors are non-linear in their current to voltage ratio. I believe they follow more of an inverse logarithmic curve, also known as the exponential function e^x. I'm kind of surprised they modeled it with a cubic function, then again the power series for e^x is well approximated by its 4th iteration. (e^x = 1 + x +x^2/!2 + x^3/!3...) which in turn models well to the above equation.

 

[Abwx]

. I'm kind of surprised they modeled it with a cubic function, then again the power series for e^x is well approximated by its 4th iteration. (e^x = 1 + x +x^2/!2 + x^3/!3...) which in turn models well to the above equation.

Agree, but then , there are the coefficients that would
seriously complicate the exponents of the exp function
if it was to interpolate the polynomial form correctly.


On a side note, and back to topic, about RWT BD/MC comparison.

Strangely, he (D. Kanter) goes to compare them using also Himeno bench,
and then conclude that BD is 28% faster than MC in this test,
but then ; when digging , we see that a 2.6 X3 beat them
both by a large margin in the very same test, which is a clue
that this test is simply irreleveant even for a cpu/cpu comparison.

 

[Idontcare]

Since this is a cubic function , i assume that it is composed
of a product of three first degree functions.
First is the current needed to charge/discharge the parasistic
capacitance, wich increase linearly with frequency.
Second should be the cross conduction losses of the push pull
cmos pairs wich increase also linearly with freq.
And third, i can only see the necessary voltage increase as the third factor that raise the final degree to 3...
Is that plausible ?..

I'm sure IDC will describe this better than me but transistors are non-linear in their current to voltage ratio. I believe they follow more of an inverse logarithmic curve, also known as the exponential function e^x. I'm kind of surprised they modeled it with a cubic function, then again the power series for e^x is well approximated by its 4th iteration. (e^x = 1 + x +x^2/!2 + x^3/!3...) which in turn models well to the above equation.

Power consumption is comprised of a static component (idle leakage) and a dynamic component (active leakage).

The dynamic leakage scales to the square of the operating voltage for traditional CMOS on 90nm and older nodes.

P = C * f * V^2
where C is capacitance of the chip (the thing that scales with die-size) and f is the frequency.

CMOS on nodes 65nm and newer have been shown to actually scale power consumption to the cube of the voltage. So we should be using: P = C * f * V^3

The data in the graph was for a 65nm based CPU and I generated the data to determine whether or not the initial reports of voltage scaling as a cubic instead of a quadratic were true.

It's part of a longer thread here somewhere in the forums regarding the topic.

Since modern IC's are composed of a wide range of transistors, both in type of transistor (variable gate thickness and/or implant conditions) as well as varying lengths and widths from circuit to circuit, it is impossible to actually derive or define an analytical expression describing the power-consumption of any given chip.

It has to be modeled, even by the guys who design them. Of course the power-consumption of a transistor does not go as the square of Vcc, nor the cube, but the system of circuits and transistors when treated as an ensemble, can be easily modeled as such and it is a surprisingly robust model given how overly simplistic it is.

 

[Morg.]

Even if Bulldozer can do 3.5Ghz, its very unlikely to have it at launch. AMD needs some headroom for future updates/revisions, they cannot just reveal all their cards at once.

I respectfully disagree :

65nm process e6700 C2D : base clock 2.66 Ghz, Average stable OC 3.5 Ghz
45nm process e6700 PDC : base clock 3.20 Ghz, Average stable OC 4.4 Ghz

45nm process Thuban : base clock 3.3 Ghz
32nm process Thuban : fat chance it's over 4.2 Ghz

So why would a Bulldozer be UNDER 3.5Ghz while it will have huge benefits from the shrinking die compared to a "normal" 8-core + the process shrinking.

 

[Abwx]

P = C * f * V^2
where C is capacitance of the chip (the thing that scales with die-size) and f is the frequency.

CMOS on nodes 65nm and newer have been shown to actually scale power consumption to the cube of the voltage. So we should be using: P = C * f * V^3

.

According to the first equation , if we keep C and F as constant,
Pd increase classically like in the ohm s law.

The second equation say that with reduced nodes, there s a third
factor that come to play and it has to do with voltage dependant
caracteristics of the Cmos trannies.

I m not familiar with micro electronics circuit design, but could
this be caused by the Fet s gate width modulation that would
start to being noticeable at such nodes?

Is this effect, wich is well known from electronic circuits designers,
still existent at such low supply voltage, and that the fet s size
reduction increase the influence of this effect?..

Anyway thanks for your explanations that are concise and easily
understable...

 

[SickBeast]

I've gotta say, I think a 3.5ghz Bulldozer is gonna PWN. 8 cores at that speed should crush intel.

 

[Dice144]

I've gotta say, I think a 3.5ghz Bulldozer is gonna PWN. 8 cores at that speed should crush intel.

I sure hope the above is true. But double the "real" cores it better! I have been saving for the top of the line CPU.

 

[PreferLinux]

I've gotta say, I think a 3.5ghz Bulldozer is gonna PWN. 8 cores at that speed should crush intel.

What though? How about a >3 GHz 8 core 16 thread Intel?

 

[IntelUser2000]

What though? How about a >3 GHz 8 core 16 thread Intel?

Nobody asked, but I'll say it anyway. The 8 core HPC optimized variant of the Westmere EX has a 2.66GHz clock. Now that's with dual memory controllers, all the RAS features, and on an older architecture. Optimizing it for the PC segment, the 8 core Sandy Bridge EP/EX should be able to do 3GHz at 130W TDP with fair ease.

 

[HW2050Plus]

Nobody asked, but I'll say it anyway. The 8 core HPC optimized variant of the Westmere EX has a 2.66GHz clock. Now that's with dual memory controllers, all the RAS features, and on an older architecture. Optimizing it for the PC segment, the 8 core Sandy Bridge EP/EX should be able to do 3GHz at 130W TDP with fair ease.

I think that in the server market this Sandy Bridge Beast will have difficulties as the Interlagos will be available there. So in this market segment I expect the Interlagos to be ahead regarding performance. It will as well feature 4 memory channels, a bit slower per core but faster overall with the CMT vs. SMT advantage.

In desktop where Sandy Bridge EN will come as Core i7 Extreme the picture will be different, since the top AMD competing will be the Zambezi 4M/8C. So it has half logic cores and slightly slower cores - no chance. It is simply that AMD will not offer an Interlagos like part in desktop market.

What I expect for 1H/2012:
Top performance/Enthusiast: Core i7 Extreme (i9 ?)
Performance Mainstream: AMD Zambezi 8 Core
Mainstream: Core i7 2xxx and AMD Zambezi 6 Core
Low cost: Core i5/i3, AMD Zambezi 4 Core, Llano

 

[IntelUser2000]

So in this market segment I expect the Interlagos to be ahead regarding performance. It will as well feature 4 memory channels, a bit slower per core but faster overall with the CMT vs. SMT advantage.

My guess:

SpecInt/FP_Rate2006

Core i7 2600 3.4GHz: 144/109
2S Opteron 4176HE 2.4GHz(6 cores/2 socket, fairly close mimic of 12 core Opteron): 185/156

(Typical gain by cores has 85% scaling in SpecInt. FP varies too much to make a statement, but the scaling lies between 50-70%)

Straight doubling of cores for the i7 2600: 266/??
50% increase in Spec number quoted by AMD for Interlagos: 278/??

Preliminary conclusion: They will be close in SpecInt, but Interlagos will have 10-20% lead on SpecFP.

What I expect for 1H/2012:
Top performance/Enthusiast: Core i7 Extreme (i9 ?)

On naming, I think they will call it 28xx/29xx. It's reasonable. 2nd generation 8xx/9xx. xx is SKU number.

 

[Abwx]

My guess:

SpecInt/FP_Rate2006

Core i7 2600 3.4GHz: 144/109
2S Opteron 4176HE 2.4GHz(6 cores/2 socket, fairly close mimic of 12 core Opteron): 185/156

(Typical gain by cores has 85% scaling in SpecInt. FP varies too much to make a statement, but the scaling lies between 50-70%)

Straight doubling of cores for the i7 2600: 266/??
50% increase in Spec number quoted by AMD for Interlagos: 278/??

Preliminary conclusion: They will be close in SpecInt, but Interlagos will have 10-20% lead on SpecFP.

.

An april 2009 AMD slide give some numbers.

SPECint : 415
SPECfp : 510

Given the values, we can assume that it s a 2S system.
Of course, the slide is quite old, but the perfs should be close
to these ones, and scaling down using your scaling factors
numbers, the results are relatively close to your estimations..

 

[JFAMD]

We never gave spec int or FP numbers

 

[Abwx]

We never gave spec int or FP numbers

So the slide is surely photoshoped...
Anyway, there s really big expectations from the general public,
so we just hope that you firm will be up to the task...

 

[996GT2]

So the slide is surely photoshoped...
Anyway, there s really big expectations from the general public,
so we just hope that you firm will be up to the task...

And rightly so. Intel's Sandy Bridge based CPUs will have been out for half a year by the time Bulldozer launches. If BD can't match the performance of Sandy Bridge at an equivalent price point, then it will be a failure.

 

[itsmydamnation]

And rightly so. Intel's Sandy Bridge based CPUs will have been out for half a year by the time Bulldozer launches. If BD can't match the performance of Sandy Bridge at an equivalent price point, then it will be a failure.

in which market? its not like server margins are massively higher or anything and bobcat is almost the prefect 3rd world chip.

So if bulldozer cant beat SB in light threaded workloads but does really well at 100% load across all cores (ie virtual server,DB servers, data miners etc), is bulldozer a failure?

if i was AMD and i could pick one of two markets to make a high performance chip its obivious which market to choose. the difference between AMD and intel is that AMD has like a 10th of the fab capacity they cant make as much money as intel in the consumer space no matter what performance they deliver.

if bulldozer can beat my I7 920 (C0) steping (shouldn't be hard with how shit an OC'er it is) i'll get a bulldozer just because im australian and that means we support the underdog by default

 

[JFAMD]

You hit the nail on the head. Everyone keeps looking at the top bin speed parts, but the reality is that 99.9% of the market is buying something down the stack.

If you want to see if a part is successful, you have to look at how it does in the places where people actually spend their money.

 

[996GT2]

You hit the nail on the head. Everyone keeps looking at the top bin speed parts, but the reality is that 99.9% of the market is buying something down the stack.

If you want to see if a part is successful, you have to look at how it does in the places where people actually spend their money.

Who said we were only looking at the top-spec parts? LGA 1155 Sandy Bridge isn't even Intel's "high end." It's their low-end to midrange lineup.

I don't care too much about how the top of the line, $1000 CPUs perform. What I care about is priceerformance. If Bulldozer can match or exceed the performance of an equivalently priced Intel part at the same price point, then I'll consider it a success. If not, I won't. Simple as that.

 

[996GT2]

You hit the nail on the head. Everyone keeps looking at the top bin speed parts, but the reality is that 99.9% of the market is buying something down the stack.

If you want to see if a part is successful, you have to look at how it does in the places where people actually spend their money.

Who said we were only looking at the top-spec parts? LGA 1155 Sandy Bridge isn't even Intel's "high end". It's their low-end to midrange setup.

I don't care too much about how the top of the line, $1000 CPUs perform. What I care about is priceerformance. If a Bulldozer based CPU can match or exceed the performance of an equivalently priced Intel part at the same price point, then I'll consider it a success. If not, I won't. Simple as that.

 

[996GT2]

You hit the nail on the head. Everyone keeps looking at the top bin speed parts, but the reality is that 99.9% of the market is buying something down the stack.

If you want to see if a part is successful, you have to look at how it does in the places where people actually spend their money.

Who said we were only looking at the top-spec parts? LGA 1155 Sandy Bridge isn't even Intel's "high end". It's their low-end to midrange setup.

I don't care too much about how the top of the line, $1000 CPUs perform. What I care about is priceerformance. If a Bulldozer based CPU can match or exceed the performance of an equivalently priced Intel part at the same price point while showing similar power draw, then I'll consider it a success. If not, I won't. Simple as that.

 

[JFAMD]

Top bin sandy bridge is probably pretty low on the list as well. mid bin is where the product "make or break" happens.

 

[Abwx]

Top bin sandy bridge is probably pretty low on the list as well. mid bin is where the product "make or break" happens.

I m not sure i did understand well.
Top bin SB will be mid bin in a few months, so either
BD can compete with it or else, we can expect a somber
outcome for AMD..

 

[JFAMD]

BD will be plenty competitive.

However, my point is that you need to look at where people are buying to judge whether something will be attractive. And I don't believe that SB will rev silicon in a few months. Cycles are longer than that.

 

[996GT2]

BD will be plenty competitive.

However, my point is that you need to look at where people are buying to judge whether something will be attractive. And I don't believe that SB will rev silicon in a few months. Cycles are longer than that.

Top bin Sandy Bridge (2500K and 2600K) are far from high-end. They occupy the $200-300 price point, which is a very common price range for consumers. Yet the performance of the 2500K and 2600K is excellent...matching $1000 CPUs from just a year ago. Thus, Sandy Bridge has set a high standard in terms of price : performance ratio, and that's something Bulldozer will need to match in order to be at all competitive

If you are suggesting that we need to look below the 2500K and 2600K, then we're looking at the low-end segment in the market...which are the < $200 CPUs. This is the only market segment in which AMD has been competitive for the last few years, and I think that for AMD's sake this situation really needs to change with Bulldozer.

 

[Topweasel]

Top bin Sandy Bridge (2500K and 2600K) are far from high-end. They occupy the $200-300 price point, which is a very common price range for consumers. Yet the performance of the 2500K and 2600K is excellent...matching $1000 CPUs from just a year ago. Thus, Sandy Bridge has set a high standard in terms of price : performance ratio, and that's something Bulldozer will need to match in order to be at all competitive

If you are suggesting that we need to look below the 2500K and 2600K, then we're looking at the very low-end segment in the market...which are the < $200 CPUs. This is the only market segment in which AMD has been competitive for the last few years, and I think that for AMD's sake this situation really needs to change with Bulldozer.

Actually the amount customers are actively spending on CPU's have gone down. Same on the video card end. Hardware performance is so far ahead of software requirements that you have geeks running on recent C2D and Athlon II parts.

In a down economy even more so. People are holding off computer purchases, even companies are. When users are purchasing them, they are not going for super machines they are going for machines that do the job they need at a price they can afford.

You can't use enthusiast retail CPU sales from an enthusiast forum to prove a trend. I would tell you the most none netbook CPU sales are in the avenue of $80-$125. The $150-$300 price range is of the enthusiast crowd which probably accounts for 10% of consumer sales. With the $500+ being that of the Overkill top of the line user or super enthusiast and maybe 1% of the market.

What's killing AMD is the lack of consistent laptop CPU options and Net book and other super cheap lower power solutions. They need Llano and Zacate more then BD. In fact the module design points more toward them hoping that its a force in the Server business. BD for desktops and its performance matters because they need to at least try to keep up in the desktop space so that they can stay in our minds when it comes to recommendations we make to the non technical crowd. But its not the end all be all.