AMD: What happened?

[blckgrffn]

Turn that 1.3 ghz duron into a smoothwall box , that's what I did with mine.. complete overkill.

Haha, I think it actually was in a router box for a while, but I couldn't stomach the power consumption

I suppose it is worth considering that AMD is much more resource constrained than Intel, perhaps it is incredible that they have been so close/competitive for so long.

Maybe it is Intel we should be disappointed in

 

[Vesku]

The CMT implementation of Bulldozer seems quite good actually, doesn't seem the engineers working on that sucked. The cores that are CMTed together certainly need tweaking but it's the power curve that is Bulldozer's Achilles' heel. The unlocked Llanos point to current GF 32nm not being much better than mature 45nm in terms of power consumption at 3+GHz. Wonder if we will see AMD CPUs being made at TSMC late 2013 early 2014.

 

[LOL_Wut_Axel]

That could definitely be part of it, but an engineer cannot just make a decision to go against the whole CPU arch they are told to build. They can make the best of it, mitigate as many of the deficiencies as possible (ie Intel worked wonders with the fsb before going IMC).

In the end, if the 'top brass' dictate a certain direction, the engineer can only do so much. I don't know the inner-workings of what happened at AMD, but I also would love to get a 'behind the blow' perspective of what went down.

I doubt it was simple, and I am sure it was wrought with bad decisions by folks who were either in over their heads or just plain stubborn to change their tune.

It could have something to do with that story (not saying it's completely accurate, but it does make sense) that was published some months ago saying that, sometime after the Athlon 64 X2 was made and the ATI acquisition, AMD decided to fire their small group of talented CPU engineers and instead decided to hire a bigger, but much less experienced/talented group.

Contrast that to ATI, which ever since AMD bought has made for the most part good (HD 3000 series) or great products (HD 4000, 5000, 6000 series).

I think AMD is being wise in admitting they can't compete with Intel when it comes to the Performance and Enthusiast CPU market and they're gonna slowly pull out of it, or at least not focus nearly as much on it. They need to keep improving on the good things they've made in the recent past and present (graphics cards), and the APU division they started around a year ago that ended up being a success.

 

[Ferzerp]

The CMT implementation of Bulldozer seems quite good actually, doesn't seem the engineers working on that sucked. The cores that are CMTed together certainly need tweaking but it's the power curve that is Bulldozer's Achilles' heel. The unlocked Llanos point to current GF 32nm not being much better than mature 45nm in terms of power consumption at 3+GHz. Wonder if we will see AMD CPUs being made at TSMC late 2013 early 2014.

The problem with what you state is that it assumes the "cores" are at fault. We don't have the inside knowledge to say if that is true, or if the truth is that the implementation of CMT is the reason the "cores" are deficient.

 

[LOL_Wut_Axel]

The problem with what you state is that it assumes the "cores" are at fault. We don't have the inside knowledge to say if that is true, or if the truth is that the implementation of CMT is the reason the "cores" are deficient.

The real problem with it is that it didn't work out the way the CPU engineers thought it would. While the CMT implementation itself is handled just fine, to be competitive with Intel they needed to make the integer cores stronger, not weaker, than Stars.

You can have twice the number of integer cores, but if they're 60% weaker than each of your competitor's, you're not going anywhere. They forgot the fundamental fact that applies to all CPU workloads, whether it be desktop or server: core scaling decreases as you add cores. So Bulldozer was a doubly bad decision (one was adding more cores than the X6 had, and two was making each core weaker).

 

[pelov]

The CMT implementation of Bulldozer seems quite good actually, doesn't seem the engineers working on that sucked. The cores that are CMTed together certainly need tweaking but it's the power curve that is Bulldozer's Achilles' heel. The unlocked Llanos point to current GF 32nm not being much better than mature 45nm in terms of power consumption at 3+GHz. Wonder if we will see AMD CPUs being made at TSMC late 2013 early 2014.

It's good in theory, but I dunno about execution. There's a ~20% CMT tax that hurts its multi-threaded performance when both integer cores within a module are tasked. Whether this has to do with fetch/decoder, nobody's really sure. Granted, that 20% wouldn't look so bad if the individual modules/cores had higher IPC or if the savings in die size, space and transistors is worth it (doubt this it due to BD's price still staying the same). Until we can see just what a non-gimped BD module can do, nobody is quite sure how well CMT will fare. It does look good in theory, but until the processor as a whole can perform competitively then it doesn't matter =P

Trinity is clocked ~30% higher and

According to performance benchmarks conducted by AMD, the Trinity 35W APU with Piledriver-class x86 cores will provide 25% better x86 performance compared to Llano 35W (with K10.5+ "Husky" x86 cores) based on results obtained in PC Mark Vantage Productivity benchmark

so I'm thinking maybe they've caught up on IPC, but I highly doubt it's much higher than Llano, if it's any higher at all.

http://www.xbitlabs.com/news/cpu/di...ity_Desktop_APUs_to_Hit_4GHz_Clock_Speed.html

Seeing high clock speeds makes me instinctively shudder and cringe

 

[Ferzerp]

Yes, but you missed my point.

Are the "cores" deficient by themselves, or do they become deficient because of the tradeoffs of gluing two of them together along with the logic to try to share as many resources as they do.

 

[Ferzerp]

AMD has shown time and time again that we cannot place any stock in any benchmark that they provide, so you will just have to wait and see, but I gather that you realize that from your post.

 

[LOL_Wut_Axel]

Yes, but you missed my point.

Are the "cores" deficient by themselves, or do they become deficient because of the tradeoffs of gluing two of them together along with the logic to try to share as many resources as they do.

They're already deficient by themselves, AND you have the problem of sharing logic of those weak cores.

 

[Vesku]

While we don't know every intricacy of Bulldozer, there has been a lot of experimentation with giving one core all module resources and comparing it to the full module. The CMT penalty seems to be inline with AMD's whitepapers as I haven't seen anyone demonstrating much worse than a 20% performance hit.

Whether CMT is worth the performance hit over 2 fully separate cores would need more information from AMD on the die area differences CMT introduces. Which dovetails into that whole 2B to 1.2B transistor count issue.

 

[pelov]

While we don't know every intricacy of Bulldozer, there has been a lot of experimentation with giving one core all module resources and comparing it to the full module. The CMT penalty seems to be inline with AMD's whitepapers as I haven't seen anyone demonstrating much worse than a 20% performance hit.

Whether CMT is worth the performance hit over 2 fully separate cores would need more information from AMD on the die area differences CMT introduces. Which dovetails into that whole 2B to 1.2B transistor count issue.

Oh yea, absolutely. And then there's the fact that software hasn't been advancing toward multi-threading as quickly as AMD would like, and certainly not to the extent AMD had hoped. It's why Intel's approach to beefier cores that can handle lightly threaded workloads and extra threads in case you need em makes more sense and currently fits much better. I'm assuming AMD went with their CMT approach because it saves space/cost when scaling up, but considering they cancelled their 5-module PDs it's safe to say they know they messed up there too. I guess in the sense that if they were planning on scaling core count up, then no, CMT doesn't work. At least not yet and not for the desktop. But I hear the server crowd likes their IPC too I guess you can say I'm not sold on CMT.

 

[Ferzerp]

I'm assuming AMD went with their CMT approach because it saves space/cost when scaling up

The current modules are nearly the exact same transistor count as 2 of their previous generation (faster) cores, so it really didn't save any space at all.

Their 6 core procs had 3/4 of the transistors as their 4 module systems do now.

 

[Vesku]

Not so much IPC as whatever combination of IPC and Hz gets them the most performance under their wattage thresholds. So in a situation like Bulldozer where you have combine a core IPC miss with worse than planned Hz-Power curve, you get a mediocre product.

I have yet to read any reasonable arguments that CMT is the culprit in the Bulldozer issues. Seems to me the reason for no 5 module PD is most likely it wouldn't fit in the highest TDP tier.

The current modules are nearly the exact same transistor count as 2 of their previous generation (faster) cores, so it really didn't save any space at all.

Their 6 core procs had 3/4 of the transistors as their 4 module systems do now.

That gets murky when you take into account the new features Bulldozer supports (AVX, AES Acceleration, XOP, FMA4, etc.) A refined Thuban core with those features is bound to add to the transistor count.

 

[pelov]

The current modules are nearly the exact same transistor count as 2 of their previous generation (faster) cores, so it really didn't save any space at all.

Their 6 core procs had 3/4 of the transistors as their 4 module systems do now.

I think you're forgetting die-size, 45nm>32nm and the atrociously large amounts of L2 and L3 cache BD has. They definitely saved some space and transistors to cram 4modules/8integer cores in there, but I think that had marketing gotten the right number the first time and a cache diet it would be more visible. It definitely is quite porky in its current state

I have yet to read any reasonable arguments that CMT is the culprit in the Bulldozer issues. Seems to me the reason for no 5 module PD is most likely it wouldn't fit in the highest TDP tier.



That gets murky when you take into account the new features Bulldozer supports (AVX, AES Acceleration, XOP, FMA4, etc.)

While that's true and CMT looks good in theory, the theory behind it is to save space/transistors by sharing resources, thus making it easier and cheaper to scale up or down than it would be with entire separate cores. AMD needed to overcome that 20% performance hit with either more cores, higher clock speed or higher IPC. They offered more integer cores, but the other two didn't work out as they'd hoped.

The jury is out on CMT and the only way AMD can possibly sway them, and me, is to prove that they can embrace the CMT approach and also deliver beefier and better performing modules. Otherwise, the only way that design can hope to succeed is to keep scaling up and selectively favor highly-threaded workloads.

 

[Ferzerp]

I'm not forgetting die size at all because there's no reason to mention it because of the process shrink. The die size for 1.2B on 32nm had better be less than the die size for 1.2B on 45nm, but, there is no reason that a 1.2B transistor die that started on a 45nm process, but was moved to a 32nm process, would be significantly larger than the 32nm 1.2B transistor die. Cache sizes (not cache performance) are nothing special and pretty in line with their competition's cache sizes.


Yes, new features add murk, but new features combined with a performance decrease in all but the most contrived of situations balances that out quite a bit, wouldn't you say? It likely would have been wiser to add the new features to a core derivative of the previous core.

 

[pelov]

Cache. Cache is the most dense part of a CPU as far as transistor count goes and BD has a LOT of it. Too much, actually. Where the Thubans had 9MB of L2+L3, BD's have 16MB.

 

[Vesku]

Merging architecture with fabrication node there. We'll have a much better picture once tangible GF 32nm improvements show up in retail chips. According to the recent AMD Q&A they have GF, AMD, and IBM engineers all working at tweaking the 32nm process.

AMD needed to overcome that 20% performance hit with either more cores, higher clock speed or higher IPC. They offered more integer cores, but the other two didn't work out as they'd hoped.

 

[Ferzerp]

The last thing AMD needs to do is throw "moar cores" at their problems.

 

[Ferzerp]

Cache. Cache is the most dense part of a CPU as far as transistor count goes and BD has a LOT of it. Too much, actually. Where the Thubans had 9MB of L2+L3, BD's have 16MB.

I see what you mean. I hadn't noticed it was 8+8 and was working under the assumption that L2 was a tiny amount compared to L3

 

[N4g4rok]

Not really. Phenom II has always been the second choice, Intel made sure of it.

With the right overclock, it's done very well for what i paid for it. Had i understood the longevity value of the 2500k at the time, i certainly would have changed my mind on that.

With how late and lackluster the response to core was you'd think AMD only heard about it the day it was released for public purchase.

What continues to confuse me is how on earth decisions like that continue to be made. It's one thing to march to be beat of your own drum, but it's another to remain completely blind to the quickening tempo of the industry.

In the end, if the 'top brass' dictate a certain direction, the engineer can only do so much. I don't know the inner-workings of what happened at AMD, but I also would love to get a 'behind the blow' perspective of what went down.

Who would have been the one to make that call? Was it a situation of having a CEO that had no idea what was going on? Or were the engineers really that unaware of how it was going to turn out. I can't pick out where it's based off poor management or poor testing.

 

[amenx]

The last thing AMD needs to do is throw "moar cores" at their problems.

Yes, not least of all the PR disaster moar cores can be when the competitions less cores outdoes it.

 

[piesquared]

Only recently have i looked into the specifics of AMD's newer activities and product launches, and as an admittedly overzealous AMD fan, i have to say i'm a little bit confused.

When i built my first system a year ago, the Phenom II was an obvious choice. Price was great in comparison to the performance i expected from it. It kept up with all the games i played (always with vsync enabled), and managed to work with most of adobe's products and a few other software development tools at reasonably close to the same levels as Intel's somewhat expensive offerings at the time.

That was fine back then, but the end of the K10/10.5 production has forced me to look into BD, Interlagos, and the APU line. I saw the excessive market hype, and figured it may have paid off just a little bit. But, like most others, i just don't see it.

Even after remote success of their graphics devision, they seem to be cutting or otherwise loosing key players from that department. I hear rumors of a BD engineer expecting this somewhat poor performance. What ever happened with that? And has there been any news of what the new CEO plans to do about some of those problems?

Even though most of this is purely conjecture, i wouldn't mind laying out what's happening with what they've said they would do about it. It's difficult to support any company when it seems like they're actively making bad decisions.

These thread titles are amuzing. It's classic manipulation, just like in the main stream media, politics, etc. They introduce a premise, and then the puppets propagate the message to suit their agenda.

What the question really oughta be, is how pathetic it is that intel with all their resources, 10x the size of it's competitors, proven illegal tactics and cohersion with the defacto software provider of 95% of worldwide PC's and largest OEM's, taken 25 years and still haven't managed to outperform the competition by more than 10-20% in the traditional CPU space. Graphics of course is another matter. But hey, everyone on this forum already knows all about it, and as long as legitimate enthusiast continue to bend over to marketing and investor interests that benefit from your manipulation they'll laugh all the way to the bank (not that I believe there are many enthusiasts actually left). The message is clear and i'm sure many see it, regardless of whether or not they choose to express their opinion.

 

[bononos]

Is AMD suffering from inferior process tech vs Intel to the point where it is having to make big jumps in hopes of catching up I wonder.

These thread titles are amuzing. It's classic manipulation, just like in the main stream media, politics, etc. They introduce a premise, and then the puppets propagate the message to suit their agenda.

What the question really oughta be, is how pathetic it is that intel with all their resources, 10x the size of it's competitors, proven illegal tactics and cohersion with the defacto software provider of 95% of worldwide PC's and largest OEM's, taken 25 years and still haven't managed to outperform the competition by more than 10-20% in the traditional CPU space........

I doubt the TS was being a manipulative viral marketer. Its just an interesting look back at the glory days of AMD.

You bring up a valid point but nothing you say explains why AMD fired some key personnel from the gfx/cpu divisions or why BD was so lackluster. In the early days Intel didn't have a monopoly over PC processors so it wasn't like they had a 25yr head start. AMD and others were oeming IBM cpus too.

 

[DominionSeraph]

Not really. Phenom II has always been the second choice, Intel made sure of it.

No, in the generation preceding SB, Intel was only the choice once you got above the price range where there were any AMD chips. Over on General Hardware we were throwing out 955/555 builds all the time. Lynnfield i3's were dead to us -- for the same total build cost of that locked dual-core you could get an unlocked AMD quad. And dual-core i5's were a COMPLETE joke. We didn't spec Intel until the i5 750. AMD didn't have anything that could match its performance at any price.

Things changed with SB because not only did Intel increase the performance by 15-20% for $0 increase in price, but decent motherboards were released in the $50-60 range. Now you can't beat Intel at any budget.

 

[N4g4rok]

What the question really oughta be, is how pathetic it is that intel with all their resources, 10x the size of it's competitors, proven illegal tactics and cohersion with the defacto software provider of 95% of worldwide PC's and largest OEM's, taken 25 years and still haven't managed to outperform the competition by more than 10-20% in the traditional CPU space.

I agree with that completely. I feel like it's a separate issue, but it's still there. Given the resources that Intel has, i think they should be a lot further than they are now, but having not used one of their products, i can't say too much.

And up until the release of BD, that's where i felt an AMD purchase within the K10/10.5 architecture was still valid. Had i gone with Intel, a few of the things i do would have benefited a little, but not so much as to justify an additional $100-$150. For me, anyway.