How much is AMD behind.

[Abwx]

You can't really base how far they are behind by time. They could come out with their own core 2 efficiency/performance jump and blow Intel away.


I'd say they're a little over 100% behind efficiency wise, which is a very significant metric to measure by.

I saw these and couldn't believe BD didn't improve at all on efficiency against older 45nm designs. Shows me just how bad this first implementation is.

What are those THGs "multithreaded" benchs where a 975 is only
10% slower than a 980X despite the latter s 50% more cores..??...

 

[Idontcare]

Put llano next to it and a 4core phenom and you'll notice the biggest hurdle is the 32nm which only gave a higher density but nothing on power or frequency spectrum.

Which is exactly what everyone under the sun who knew anything about process tech was saying was going to happen with AMD/GloFo's choice to go gate-first for their HKMG process. (not talking here in the laymen/public space, am talking about in the industry itself, from chemical suppliers to tool makers to process engineers themselves, the AMD PR machine was fooling no one about the "benefits" of gate-first...then again, IPC doesn't go down either, by their math, so no wonder gate-first made sense )

 

[AtenRa]

Which is exactly what everyone under the sun who knew anything about process tech was saying was going to happen with AMD/GloFo's choice to go gate-first for their HKMG process. (not talking here in the laymen/public space, am talking about in the industry itself, from chemical suppliers to tool makers to process engineers themselves, the AMD PR machine was fooling no one about the "benefits" of gate-first...then again, IPC doesn't go down either, by their math, so no wonder gate-first made sense )

You cant just compare two different micro architecture designs manufactured in two different litho processes and say that process A(Gate Last) is better than process B(Gate First). You would have to manufacture the same CPU with both processes and make your final conclusion.

Since PileDriver is going to have ~15% better performance per wattage than Bulldozer im betting that PileDriver (32nm SOI HKMG Gate First) CPUs coming in the next 3-4 months will be very close in performance per wattage in multithreaded apps compared to Core i7 2600K SandyBridge (32nm HKMG Bulk Gate Last). And all that with double the L cache (16MB vs 8MB).

Im not saying that Gate First is better but i believe that each process has its positives and negatives. There are so many technical and economical parameters we dont know for both processes that i believe we really canot say if one is better over the other in general.

Also, Gate First may work well for a company like AMD but not for Intel which has much larger production capacity.

 

[Idontcare]

Im not saying that Gate First is better but i believe that each process has its positives and negatives. There are so many technical and economical parameters we dont know for both processes that i believe we really canot say if one is better over the other in general.

The differences are rather stark, and specific. One gives you lower performance but higher density, the other gives you higher performance but lower density. Literally everyone who does this for a living in the industry understands the basic tradeoff that goes into the decision table.

The only folks who seemed to not understand that AMD was not going to have its cake and be able to eat it too was the PR machine that spit out all those nice GloFo pdf files touting gate-first superiority.

You cant just compare two different micro architecture designs manufactured in two different litho processes and say that process A(Gate Last) is better than process B(Gate First). You would have to manufacture the same CPU with both processes and make your final conclusion.

If you fundamentally understand the cause then you can project the effects in a relatively straightforward fashion. This is the difference between guessing, and being surprised when your guess is wrong, versus making an educated estimate. It is the difference between the doctor who can tell you the difference between one type of cough and another versus your mom who says everything will be better if you eat chicken soup.

Sometimes mom is right, other times you have cancer and that cough just isn't going to get better with a little soup, and the doctor is probably the one you want to be seeking an educated opinion from.

When IBM and AMD announced they were doing gate-first HKMG there was no one left in the audience who bought the AMD PR line that it was going to deliver as promised for very basic fundamental reasons relating to the stress engineering aspects on the channel that come about when you replace the disposable gate.

If process tech could not be so readily compared, and understood on the basis of universally applicable metrics of analysis, then professional conferences like IEDM and IITC would not have been held for 60 yrs now.

Whether you look at Llano's or Bulldozer's performance/watt, either chip paints the same story, as expected, when comparing to GloFo's non-HKMG 45nm or Intel's HKMG 32nm - it delivered on areal density entitlement and not much else.

It reduces production costs but does not enable higher ASP because it does not enable higher performance. And AMD's numbers prove this out, to the surprise of no one.

 

[alyarb]

Would gate last have secured more contracts other than from AMD? Would it have improved yields?

 

[Charles Kozierok]

It's a double standard, and there should be IMO. David vs Goliath.

Valid argument. Yes, if AMD were able to beat Intel by 22%, that would be more of a big deal than Intel beating AMD by 22%.

The flip side, though, is that specifically because AMD is the challenger, it must beat Intel to be competitive. AMD needs to give people a reason to consider an alternative to the industry standard.

And so, to a large practical extent, they really need to be beating Intel by a considerable margin to be successful. And yet they are trailing. That is why they are no longer competitive.

Saying "well, they're only X% behind" really misses the point. If they aren't X% ahead, why would anyone bother with them?

 

[Ferzerp]

So you guys would buy cars 22% worse (in whatever metric is most important to you) just because it comes from a smaller company?

How about televisions? Food that is 22% worse for you but comes from a slightly less large corporation?

Really?

 

[richaron]

...And AMD's numbers prove this out, to the surprise of no one.

Yes... Surprise of no one... ~twiddles thumbs waiting for AM3+ chip worthy to replace $40 ebay 555~

 

[taltamir]

What are those THGs "multithreaded" benchs where a 975 is only
10% slower than a 980X despite the latter s 50% more cores..??...

Scaling to more core is rarely done linearly, hence only a partial gain from having more cores rather then the full 50%.

 

[sm625]

Well I would be perfectly happy with an A10. It can handle everything I need it to just fine, including LoL and the upcoming Smite. I guess it all comes down to how much intel is going to charge for the cheapest HD4000 chip. Right now it is way way too expensive. And when the i3 releases there's no telling how much an HD4000 version will cost, if there is one at all.

The problem for AMD is that an A10 costs about the same as a G620 and a 7750 class card.

 

[taltamir]

So, nobody has anything to say about my contention that AMD is not "years" behind but has feature parity with inferior "budget" feature and that redoing things right will require them to spend more then intel originally did AND put them years behind?

 

[AtenRa]

The differences are rather stark, and specific. One gives you lower performance but higher density, the other gives you higher performance but lower density. Literally everyone who does this for a living in the industry understands the basic tradeoff that goes into the decision table.

The only folks who seemed to not understand that AMD was not going to have its cake and be able to eat it too was the PR machine that spit out all those nice GloFo pdf files touting gate-first superiority.

I know and i perfectly understand the general differences between those two processes. But you clearly cannot take Llano or Bulldozer and compare them to SandyBridge because the architectures are different.
We cannot even compare Llano and Bulldozer although they both are manufactured with the same 32nm process.

Whether you look at Llano's or Bulldozer's performance/watt, either chip paints the same story, as expected, when comparing to GloFo's non-HKMG 45nm or Intel's HKMG 32nm - it delivered on areal density entitlement and not much else.

It reduces production costs but does not enable higher ASP because it does not enable higher performance. And AMD's numbers prove this out, to the surprise of no one.

Although Llano is a Deneb derivative design we really dont know what type of transistors they used for the CPU cores. Did they used bigger but faster transistors or smaller for greater density ?? Remember that Llano is an APU with a large iGPU, they may needed to keep the die as small as possible and they may sacrificed CPU performance in order to make the die smaller. Does the Llano CPU core architecture allows it to operate in higher frequencies or was it designed for up to 3-3.5GHz.

On the other hand Bulldozer was designed for higher operational frequencies and if you compare it to 6 Core Deneb you will see that AMD was able to double the Cache, put extra logic and at the end they managed to have the same Performance per watt.
Now, PileDriver (32nm SOI HKMG Gate First) will have 15% more performance per watt than Bulldozer and that means better performance per watt than Deneb (45nm SOI Gate Last).
So it seams that GloFos 32nm SOI HKMG Gate First gave them both area density AND lower power usage. You see what i did here, i used the same logic you and others have and came to a conclusion that GloFos 32nm gave them both. But because Bulldozer and PileDriver are different microarchitectures than Deneb, we cant be 100% sure if the better performance per watt came from the microarchitectural design or it came from the manufacturing process. I believe both contributed.


If we had 6 Core Deneb manufactured at GloFos 32nm process then we would be able to see how much area density and power usage they gained from 45nm.

ASPs will go higher, starting with Trinity. A10 APUs will go against the Core i5 in the Mobile.

If you fundamentally understand the cause then you can project the effects in a relatively straightforward fashion. This is the difference between guessing, and being surprised when your guess is wrong, versus making an educated estimate. It is the difference between the doctor who can tell you the difference between one type of cough and another versus your mom who says everything will be better if you eat chicken soup.

Sometimes mom is right, other times you have cancer and that cough just isn't going to get better with a little soup, and the doctor is probably the one you want to be seeking an educated opinion from.

Because my mother is a doctor (Pneumonologist : doctor specializing in lung care) I can tell you that she would need a lot of exams (Blood exams, x-rays etc) to diagnose a cancer. She will not be able to tell you only by hearing you cough if you have a Cancer or a cold.

Same with Litho process, we need to know all the technical specs to say with certainty if 32nm only gave them higher density and not performance or lower power usage.

 

[taltamir]

Where do you people find doctors that actually perform tests and find what the problem is?
Every doctor I have been to looks at the symptoms, guesses, and prescribes an antibiotic. Then when the symptoms don't go away they try a different one, then a different one.

And if you really insist they actually figure out WHAT is wrong by the 3rd time they assure its really not necessary and insist you try a 4th one. Then you just give up and learn to live with the issue. (assuming you do live).

The only things doctors seem willing to test is a blood test and an xray/mri scan. Oh, and cancer biopsies.
I think a big deal is how fucking expensive and difficult labs tests are. If you want one you basically gotta volunteer for a research clinic.

 

[inf64]

So you guys would buy cars 22% worse (in whatever metric is most important to you) just because it comes from a smaller company?

How about televisions? Food that is 22% worse for you but comes from a slightly less large corporation?

Really?

Intel's chip that is 1.22x faster than FX8150 (3770K) is 1.75x more expensive...So it gives the buyer an "amazing" 1.43x worse perf./dollar ratio.
It's not about AMD being the smaller company it's about them having solid product. 22% slower in real world workloads than top of the line 8T IB is not that much actually. At the same time you pay MUCH less for this 22% perf. deficit and in the end you get a product with better perf./dollar (stock vs stock).

 

[Yuriman]

Intel's chip that is 1.22x faster than FX8150 (3770K) is 1.75x more expensive...So it gives the buyer an "amazing" 1.43x worse perf./dollar ratio.
It's not about AMD being the smaller company it's about them having solid product. 22% slower in real world workloads than top of the line 8T IB is not that much actually. At the same time you pay MUCH less for this 22% perf. deficit and in the end you get a product with better perf./dollar (stock vs stock).

Bulldozer does indeed look pretty good from that perspective. How does it stack up against a 3570K (and non-K) though? Also, where would Intel's performance be if they aimed for the same TDP?

I could do some testing to see how high my 3570K will ramp up when limited to 125w.

Part of that increased cost is paying for the compromise Intel made to lower power usage and reduce noise levels, which carries over into cost savings on cheaper motherboards and smaller heatsinks.

 

[ultimatebob]

I don't think that AMD has plans to "catch up". Instead, they'll probably focus on low cost and mobile processors, and let Intel have the high-end performance crown.

 

[inf64]

There is 3570 in the hardware.fr chart. It's slightly slower than FX8150 while costing 7.5% more. It does have lower TDP rating though.

 

[moonbogg]

There is 3570 in the hardware.fr chart. It's slightly slower than FX8150 while costing 7.5% more. It does have lower TDP rating though.

Slightly slower with half the cores...

AMD = low power, low cost, low performance, low desirability, low FPS, low IPC, low value, low benefit, low dreams, low hope, low chance, low prayers, low chance of miracle, low low low low, but high comedy value.

 

[ShintaiDK]

I don't think that AMD has plans to "catch up". Instead, they'll probably focus on low cost and mobile processors, and let Intel have the high-end performance crown.

AMD is on the VIA route and passed the point of no return.

 

[Phynaz]

Intel's chip that is 1.22x faster than FX8150 (3770K) is 1.75x more expensive...So it gives the buyer an "amazing" 1.43x worse perf./dollar ratio.
It's not about AMD being the smaller company it's about them having solid product. 22% slower in real world workloads than top of the line 8T IB is not that much actually. At the same time you pay MUCH less for this 22% perf. deficit and in the end you get a product with better perf./dollar (stock vs stock).

Your comparing the cost of one ingredient and not the whole meal.

How much cheaper is a comparably built BD system compared to an IB system. It's a couple of percent.

Now which has the better perf / dollar?

 

[Sweepr]

How is Gulftown a fat core compared to BD? Die size is 248mm2 and 1.2B transistors. Amd BD? 315mm2 and also 1.2B. Just look at the performance gap in games, it is HUGE.

According to the Hardware.fr chart FX8150 cant even beat the nearly 5-year old Core 2 Quad QX9770 in games. Core i5 3570K is 62% faster than the FX8150 in their list of CPU-intensive games at 1080p, a chip with nearly twice the die size (without GPU) and much higher load power comsumption (not even talking about OC). And then there's Haswell next year, which should bring not only great iGPU performance but hopefully another healthy ~10% IPC boost (+AVX2) and higher clocks.

http://www.hardware.fr/articles/863-22/moyennes.html

 

[Martimus]

I know and i perfectly understand the general differences between those two processes. But you clearly cannot take Llano or Bulldozer and compare them to SandyBridge because the architectures are different.

We cannot even compare Llano and Bulldozer although they both are manufactured with the same 32nm process.

Although Llano is a Deneb derivative design we really dont know what type of transistors they used for the CPU cores. Did they used bigger but faster transistors or smaller for greater density ?? Remember that Llano is an APU with a large iGPU, they may needed to keep the die as small as possible and they may sacrificed CPU performance in order to make the die smaller. Does the Llano CPU core architecture allows it to operate in higher frequencies or was it designed for up to 3-3.5GHz.

On the other hand Bulldozer was designed for higher operational frequencies and if you compare it to 6 Core Deneb you will see that AMD was able to double the Cache, put extra logic and at the end they managed to have the same Performance per watt.
Now, PileDriver (32nm SOI HKMG Gate First) will have 15% more performance per watt than Bulldozer and that means better performance per watt than Deneb (45nm SOI Gate Last).
So it seams that GloFos 32nm SOI HKMG Gate First gave them both area density AND lower power usage. You see what i did here, i used the same logic you and others have and came to a conclusion that GloFos 32nm gave them both. But because Bulldozer and PileDriver are different microarchitectures than Deneb, we cant be 100% sure if the better performance per watt came from the microarchitectural design or it came from the manufacturing process. I believe both contributed.

If we had 6 Core Deneb manufactured at GloFos 32nm process then we would be able to see how much area density and power usage they gained from 45nm.

ASPs will go higher, starting with Trinity. A10 APUs will go against the Core i5 in the Mobile.

Because my mother is a doctor (Pneumonologist : doctor specializing in lung care) I can tell you that she would need a lot of exams (Blood exams, x-rays etc) to diagnose a cancer. She will not be able to tell you only by hearing you cough if you have a Cancer or a cold.

Same with Litho process, we need to know all the technical specs to say with certainty if 32nm only gave them higher density and not performance or lower power usage.

Global Foundries 32nm SOI process has nearly the same low-K process as their 45nm process. The difference is in a few layers where an even lower-K material is used. So, the leakage should be reduced slightly based on that, but there are other issues that may have actually raised leakage over 45nm levels.
As for the high-K material, the reason gate last is better for performance is simply that you can pick and choose the material to use for each transistor, and optimize both N-Type and P-Type transistors. Gate first requires that you use the same material for all transistors. Beyond this, gate last also allows you to use more exotic materials that would not survive the entire process, since these materials are added at the end of the process; gate first materials need to be able to withstand the entire process as the High-K material is applied first. Gate last requires more space, since you need to be able to cut out the gates, and replace them with the new material. Gate last also costs more, for the same reason (extra steps).
Anyway, IDontCare knows more about all of this than I do by a long shot. In fact, he is the one who taught me all of what I just said above. I just figured I would try to write out the process as I understand it, for two reasons: One, it may help you understand the process at GloFo; Two, if I misunderstand the process, someone may correct me and I will be better informed for it.

 

[inf64]

Slightly slower with half the cores...

AMD = low power, low cost, low performance, low desirability, low FPS, low IPC, low value, low benefit, low dreams, low hope, low chance, low prayers, low chance of miracle, low low low low, but high comedy value.

Actually the part that does most of the calculations in real world workloads (integer SIMD) is equivalent to intel i5/i7. Bulldozer "8 core" has 4 floating point units,just as Nehalem/SB/IB.

But you wouldn't know this,right?

 

[taltamir]

Intel's chip that is 1.22x faster than FX8150 (3770K) is 1.75x more expensive...So it gives the buyer an "amazing" 1.43x worse perf./dollar ratio.

Electricity costs money and once you factor that in intel chips are cheaper.

 

[inf64]

Electricity costs money and once you factor that in intel chips are cheaper.

By the time you get the "savings" from lower wattage you will be uprgading anyway. The difference is 150 dollars. How much you think you can save up? Most of the time you won't stress all cores on both CPUs ,on the contrary. Chips will sit idle in lower power state mode 90% of the time. Bulldozer does have really good idle power results due to good power gating so savings won't show up soon enough to justify this "metric". The only way the higher power draw will show up sooner on the power bill is if you OC both to the max . IB is not power savvy either in max. OC situations but BD's power draw is quite a bit higher in this case. In this situation power bill may justify the higher starting price,after some time.