AMD to manufacture console APU at GLF

[NTMBK]

Bad, but Bulldozer isn't a significant improvement either, is it? And killing Bulldozer in 2009 would mean that AMD would have a brand new chip by 2013, and depending on the scope of this new chip that would have left the company with a far better cash position, especially if they started to make the SG&A cuts at that time. Instead AMD went full speed ahead with the R&D, didn't adjust the cost structure and launched Bulldozer.

Dirk was staring at a failure in 2009, and instead of canning it and move on he proceeded ahead. This is the decision that killed his job, not speaking against the WSA. WSA is killing AMD because the CPU business is becoming a shadow of its former self. If they weren't to pay charges, the WSA would be just bad.

They would have four years without a CPU that isn't even remotely competitive, and no guarantee that their "brand new chip" would turn out any better than Bulldozer would! And it took them seven years to get their last "brand new chip" finished, you really think that they would have one finished by 2013?

 

[NTMBK]

Sure but that's also unfair on BD/PD as the traditional architectural reasons for BD/PD was more cores. Trinity is nothing but a hack together of a good iGP and half a server chip. If I recall correctly, the BD architecture doesn't start coming into it's own until the 5th core and onwards.

It's a pretty interesting comparison with Llano, actually. In a comparable die area on the same process, they managed to get slightly better multithreaded performance and significantly higher multithreaded performance. Take a look at the numbers for Cinebench:

A10-5800k:
-single threaded 4083
-multithreaded 13119

A8-3850:
-single threaded 3387
-multithreaded 12137

That's the tradeoff you get from CMT. Compared to Intel, you can fit more weaker cores into the same area, yes; but when you compare it to what AMD has, you fit stronger single threaded performance into the same area.

There's obviously far more to it than that, but you don't need to go to 5+ cores to see the difference. You just need to compare 2 Husky cores with one Piledriver module.

 

[erunion]

In all fairness, that's a bad comparison. Piledriver and Thuban were on different process nodes. Llano and Trinity is the closest comparison we have- same market segment, same process node.

Llano was the first 32nm product while Trinity came out a year later on a more mature process. Llano should have been bulldozer but obviously AMD decided that Stars would perform better than bulldozer would.

Trinity has a full ghz advantage and is barely faster.

 

[erunion]

:

A10-5800k:
-single threaded 4083
-multithreaded 13119

A8-3850:
-single threaded 3387
-multithreaded 12137

..

That's a 2.9 GHz llano vs 3.8ghz (4.2 ghz turbo) trinity.

 

[NTMBK]

Llano was the first 32nm product while Trinity came out a year later on a more mature process. Llano should have been bulldozer but obviously AMD decided that Stars would perform better than bulldozer would.

Trinity has a full ghz advantage and is barely faster.

AMD obviously knew about GloFo process yields, clocks and maturities better than any of us forum posters do. If they think that a higher binned Llano would have performed better than Trinity, then they would have delayed Trinity and released a higher clocked Llano- same way that they delayed Kaveri and released Richland (higher clocked Trinity).

And "a full GHz advantage"? Oh come on, you must know better than that. Different architectures clock differently on the exact same process. It's overall performance/W, or performance/mm^2, that actually matter, not the choices that architects make to get there.

 

[NTMBK]

That's a 2.9 GHz llano vs 3.8ghz (4.2 ghz turbo) trinity.

It's a 32nm, 100W Llano vs a 32nm, 100W Trinity.

 

[ShintaiDK]

Add turbo+1 year more mature node. And Stars would perform the same.

And remember there already was a 3870K too. 100Mhz more.

 

[erunion]

It's a 32nm, 100W Llano vs a 32nm, 100W Trinity.

Right, but a year apart. The 32nm process Llano was built on was immature. Could stars have clocked higher if it was built a year later? Probably.

AMD probably chose Stars for Llano because the process couldn't clock high enough to take advantage of bulldozer's pipeline.

 

[NTMBK]

Add turbo+1 year more mature node. And Stars would perform the same.

They already had Llano on that node. If they could have shipped a better-than-Trinity new Llano bin at the time that they shipped Trinity, then they would have done. Come on, even you have to admit that AMD isn't that intentionally suicidal. The performance delta between Llano and Trinity is pretty big, and the only higher clocked Llano bin we ever saw was the 100MHz bump to the 3870k. No way that process improvements could have closed that gap. If GloFo fixed anything, it wasn't performance- it was yields. There's a reason that AMD had massive Llano shortages in their first 6 months, and it wasn't overwhelming demand.

 

[NTMBK]

Right, but a year apart. The 32nm process Llano was built on was immature. Could stars have clocked higher if it was built a year later? Probably.

AMD probably chose Stars for Llano because the process couldn't clock high enough to take advantage of bulldozer's pipeline.

They already had a design ready to go on 32nm, which based on the latest process numbers would have outperformed Trinity, and decided to ditch it and launch Trinity instead? Why the hell would they do that?

If Llano + 1 year of process improvements would have outperformed Trinity, then that is what they would have sold- it would have cost almost nothing to just release some higher binned Llano chips. And their motherboard partners got pretty badly burned by AMD ditching socket FM1, leaving them with a lot of unsold stock- if AMD could have avoided that situation, then they would have.

 

[ShintaiDK]

They already had Llano on that node. If they could have shipped a better-than-Trinity new Llano bin at the time that they shipped Trinity, then they would have done. Come on, even you have to admit that AMD isn't that intentionally suicidal. The performance delta between Llano and Trinity is pretty big, and the only higher clocked Llano bin we ever saw was the 100MHz bump to the 3870k. No way that process improvements could have closed that gap. If GloFo fixed anything, it wasn't performance- it was yields. There's a reason that AMD had massive Llano shortages in their first 6 months, and it wasn't overwhelming demand.

AMD hoped for the speedracer to success. Never gonna happen. And it costed them the server segment, the mobile segment, the enthusiast desktop and the performance desktop segment. Not to mention the problems even Celerons give them.

And talking about suicidal. To start and restart a project that Intel had to abandon for the exact same problems AMD got now is suicidal.

 

[NTMBK]

AMD hoped for the speedracer to success. Never gonna happen. And it costed them the server segment, the enthusiast and the performance segment.

And talking about suicidal. To start and restart a project that Intel had to abandon for the exact same problems AMD got now is suicidal.

Despite the fact that there are some pretty striking similarities between the Netburst and Bulldozer stories, you do know that they're not actually the same architecture, right?

 

[ShintaiDK]

Despite the fact that there are some pretty striking similarities between the Netburst and Bulldozer stories, you do know that they're not actually the same architecture, right?

Its completely irrelevant. They share the same design mentality. Brute force frequency on the expense of everything else. And it simply doesnt work. That AMD also had to cripple themselves further with the CMT design and shared FPU is another matter. But again, it was an expense they had to pay for due to their design decision of the speedracer.

 

[NTMBK]

Its completely irrelevant. They share the same design mentality. Brute force frequency on the expense of everything else. And it simply doesnt work. That AMD also had to cripple themselves further with the CMT design and shared FPU is another matter. but again, it was an expense they had to pay for.

Higher frequency != frequency at the expense of all else. And in laptop parts, the parts that really matter in terms of sales, those frequencies really aren't that high. Heck, even the desktop parts don't have frequencies that much higher than what Prescott hit! Intel were aiming for 5GHz back then! Trying to hit the same frequency 8 years later isn't that insane by comparison. CMT is the more innovative focus of the architecture. (Whether you like the results of CMT is another matter, of course.)

 

[NTMBK]

Aww shit, I did it again. I got sucked into an argument defending Piledriver/Bulldozer, instead of actually talking about the thread topic. I'll not be pursuing these arguments any more I'm afraid, guys.

 

[AtenRa]

For the millionth time, Bulldozer was designed for two things. Server throughput workloads (dies with L3 Cache) and Fusion. In both cases it was a big gain from the previous Stars architecture.

In servers the first Bulldozer Opterons(62xx series) have more than 33% higher Integer Throughput than Stars Opterons(61xx series) at the same power. Vishera Opterons (63xx series) have 45% more Integer Throughput than Stars (61xx) at the same power.

Opteron 6180SE 12core 140W TDP SPECint_rate_base = 371

Opteron 6282SE 16core 140W TDP SPECint_rate_base = 499

Opteron 6386SE 16core 140W TDP SPECint_rate_base = 536

Also, as was mentioned before, Trinity and Ritchland is way faster than Llano with lower power at the same 32nm process. Not only that, but Kaveri and Carrizo will continue to have Bulldozer architecture CMT cores.

The Bulldozer architecture is more efficient both in performance and die area with more features and higher scalability than Stars. It just that the original Bulldozer was not ready for Desktop use, Vishera was.

If the desktop market would not have been collapsed, a 22nm SteamRoller 8 Core CPU would be a very nice alternative against Intel's Core i7. But things have changed and AMD have change since then.

 

[ShintaiDK]

Higher frequency != frequency at the expense of all else. And in laptop parts, the parts that really matter in terms of sales, those frequencies really aren't that high. Heck, even the desktop parts don't have frequencies that much higher than what Prescott hit! Intel were aiming for 5GHz back then! Trying to hit the same frequency 8 years later isn't that insane by comparison. CMT is the more innovative focus of the architecture. (Whether you like the results of CMT is another matter, of course.)

Innovative? The performance is not there, the diesize savings are not there. So are we talking innovative for the sake of innovative?

AMD is already at 220W for 4.4 and 4.7Ghz. 4Ghz is hitting 140W+ and AMD wont even release specs for it.

And thats from the same company that made slides with jokes about the P4 in 2005 on the same issues.

 

[mrmt]

AMD obviously knew about GloFo process yields, clocks and maturities better than any of us forum posters do. If they think that a higher binned Llano would have performed better than Trinity, then they would have delayed Trinity and released a higher clocked Llano- same way that they delayed Kaveri and released Richland (higher clocked Trinity).

How much did they spend on Bulldozer and how much did they spend on Stars? Instead of cutting resources like mad, they could have made a much softer transition to their next chip.

What I'm saying is that Bulldozer was a bad business decision that killed Dirk et all career at AMD. I'm saying that Bulldozer wasn't a good investment.

 

[SiliconWars]

You can say that about any failed chip. There has been plenty of failures that didn't cost people their jobs, though in this case I probably agree somewhat that Bulldozer cost Dirk his. It's the wrong chip or at best it was the wrong time for it.

 

[monstercameron]

well isn't this design made to be a 1st step for a fusion processor? maybe at worst we can say it was too early...

 

[mrmt]

You can say that about any failed chip. There has been plenty of failures that didn't cost people their jobs, though in this case I probably agree somewhat that Bulldozer cost Dirk his. It's the wrong chip or at best it was the wrong time for it.

This. What NTMBK was saying is that Bulldozer and subsequent iterations were not a failed chips, when it is clear that it is. The thing was clearly aimed at servers, and AMD market share crashed big time after Bulldozer was launched, and on top of that, Piledriver could not reverse the tide to the point that AMD isn't even considering steamroller for servers. AMD new server strategy does not pass through Bulldozer, it will be ARM as mainstay focus. On the consumer market AMD also lost market share everywhere, and today they are basically tied to desktops, the worst place for margins and growth.

AMD lost share in every single market CMT was fielded. If this is not a failure, I don't know what it is.

Sure, it takes more than a failed chip to take down a CEO, but bet your entire farm in a failed chip after the first iteration of it could not beat your then 6 years old workhorse is the kind of thing that will cost the CEO a job.

well isn't this design made to be a 1st step for a fusion processor? maybe at worst we can say it was too early...

Fusion is not dependent on Bulldozer or CMT. Kabini is fusion, and does not leverage on CMT. AMD on ARM will also be Fusion.

 

[NTMBK]

This. What NTMBK was saying is that Bulldozer and subsequent iterations were not a failed chips, when it is clear that it is. The thing was clearly aimed at servers, and AMD market share crashed big time after Bulldozer was launched, and on top of that, Piledriver could not reverse the tide to the point that AMD isn't even considering steamroller for servers. AMD new server strategy does not pass through Bulldozer, it will be ARM as mainstay focus. On the consumer market AMD also lost market share everywhere, and today they are basically tied to desktops, the worst place for margins and growth.

Wow, way to entirely misrepresent what I said...

What I was saying is that following through with Bulldozer at 32nm was better than the alternative, which was some sort of 32nm Husky based product. Bulldozer (and in particular Piledriver, after they fixed some of the bugs in initial Bulldozer) was superior to the AMD parts which preceded it.

What it wasn't is competitive with Sandy Bridge in performance/W, or raw performance, which is why AMD continued to bleed market share. It was an improvement, but nowhere near enough of an improvement to catch up with Intel. The only way that the 6300 series Opterons could compete with the Xeon E5 was on performance/$, and providing a considerably cheaper way to get quad-socket servers than the Xeon E7.

(Also, AMD is actually releasing Steamroller products for the single socket server market- Berlin, i.e. Kaveri with a different name. It's aimed at the same market as the Xeon E3, but with more of a GPGPU focus.)

 

[NTMBK]

How much did they spend on Bulldozer and how much did they spend on Stars? Instead of cutting resources like mad, they could have made a much softer transition to their next chip.

What I'm saying is that Bulldozer was a bad business decision that killed Dirk et all career at AMD. I'm saying that Bulldozer wasn't a good investment.

No, it probably wasn't a good investment. 7 years of development and a completely canned 45nm design isn't cheap! That I agree with. But by the time the 45nm Bulldozer part was canned, the majority of that development cost was already sunk. Getting 32nm Bulldozer finished was the "softer transition" by that stage, as opposed to abandoning the project entirely and restarting from scratch.

AMD definitely blew their predictions on Bulldozer- I'm not denying that. If instead of doing Bulldozer at all they had plowed the money into a) building on Stars and b) improving their process technology, they probably would be in a much better position. But by the time 45nm Bulldozer was canned, that ship had already sailed, and they had to make the decision based on where they were at the time.

 

[mrmt]

What I was saying is that following through with Bulldozer at 32nm was better than the alternative, which was some sort of 32nm Husky based product.

This. A better alternative or a better chip? If you say a better chip, I can understand. Bulldozer and especially Piledriver are faster than Husky and Stars.

But better alternative? Husky and Stars were already sunk cost, AMD had to spend basically 0 R&D on them. Any money investing on these cores would just mean better cores, and quite frankly, Piledriver is not that far from Husky. Do you think that an updated Husky (not a rush job like Stars) would be too far from Bulldozer?

Bulldozer OTOH ate 7 years of AMD R&D, and all they got in exchange is a chip a bit faster, but not by the magnitude they needed in order to stay competitive. That's not a good alternative at all.

AMD definitely blew their predictions on Bulldozer- I'm not denying that. If instead of doing Bulldozer at all they had plowed the money into a) building on Stars and b) improving their process technology, they probably would be in a much better position. But by the time 45nm Bulldozer was canned, that ship had already sailed, and they had to make the decision based on where they were at the time,

And they picked the wrong decision. I don't think they would be much worse than they are now if they have picked a Husky derivative to field circa 2013. More important, if they were to pursue different venues like they are now, conditions would be far more favorable in 2009 than in 2011 and 2012. Instead AMD went from a bad situation to a situation where they would not be able to pay their SG&A bill and shipments crashed to the point it triggered the minimum commitment clause of the WSA.

If pursuing Bulldozer was the best alternative to AMD, or at least the BoD was unanimously intending to do, Dirk and the old management team would not have lost their jobs. Quite the opposite, the BoD would have backed them up. That Dirk was fired as soon as AMD got definitive Bulldozer silicon from the foundry and the new management team got carte blanche to move/fire whoever they want speaks volumes about who was the ultimate backer of the Bulldozer team, and that captain had to go with his ship.

 

[AtenRa]

But better alternative? Husky and Stars were already sunk cost, AMD had to spend basically 0 R&D on them. Any money investing on these cores would just mean better cores, and quite frankly, Piledriver is not that far from Husky. Do you think that an updated Husky (not a rush job like Stars) would be too far from Bulldozer?

Either you are clueless or.....you are clueless.
In order to get STARs Cores faster, you had to spend money on R&D for the Front End, then you had to change the execution units in order to have AVX, FMA etc and that would cost you again. Then you would have to invest R&D money on Retire and L caches. And the end result would be a new Micro-architecture and tons of money spend on R&D and lost time.

On the other hand, AMD spend money and Time to get the CMT architecture, it was only logical to make a new strategy around what they were investing all that time and create the Bulldozer micro-architecture. That architecture was created to last 5-6 years, Stars architecture would not be able to scale as far as Bulldozer and AMD would have to change to a new architecture sooner again.