Can AMD "rescue" the Bulldozer?

[pelov]

AMD already has a supply issue with Llano, they got to bin real wide. That they have to do this is just as telling as the GHz and clockspeed issue in the first place.

The yield issue has more to do with the density of the APU design, namely the graphical portion of it, and how demanding it's been for GloFo to produce them at high yields. Given the low # of cache I'm not sure how big this Ghz and clockspeed issue is for a product like the Llano. Frankly, I think it's a great little chip that's a decent performer given the TDP, hefty graphical performance and low clockspeed.

It's quite telling that AMD got pissy after they (GloFo) couldn't provide enough Llanos for OEMs and that they potentially lost the chance to provide crApple with Llanos to fill their lappies due to it. Has it impacted Bulldozer as well? Maybe. The large amounts of cache probably isn't making it any easier on GloFo.

 

[wlee15]

Bulldozer isn't a server oriented CPU, SpecInt and FP Interlagos outperforms the Xeon X5690 by 14% and 30% respectively.

The 6 core Core i7 3960X, gets 34% and 65% better in the same metric over Core i7 990X, which is essentially the same chip as X5690. Then there's the 8 core Xeon E5's, which should add 10-15% on top of that because of 33% more cores but lower clocks.

What doesn't work in one area doesn't work anywhere else.

A 4-way 6282 SE matches a 4-way E7-4870 quite well in Specint and FP. Even through Sandy Bridge-EP is going cover a bit of the ground with IPC and cpu frequency increases, it'll still not match Westmere-EX 2 extra cores.

 

[wlee15]

The yield issue has more to do with the density of the APU design, namely the graphical portion of it, and how demanding it's been for GloFo to produce them at high yields. Given the low # of cache I'm not sure how big this Ghz and clockspeed issue is for a product like the Llano. Frankly, I think it's a great little chip that's a decent performer given the TDP, hefty graphical performance and low clockspeed.

It's quite telling that AMD got pissy after they (GloFo) couldn't provide enough Llanos for OEMs and that they potentially lost the chance to provide crApple with Llanos to fill their lappies due to it. Has it impacted Bulldozer as well? Maybe. The large amounts of cache probably isn't making it any easier on GloFo.

For what it's worth Fudzilla is reporting that Trinity yields are good.

http://www.fudzilla.com/home/item/25099-trinity-yields-are-good

 

[pelov]

For what it's worth Fudzilla is reporting that Trinity yields are good.

http://www.fudzilla.com/home/item/25099-trinity-yields-are-good

If that's true that's great news. Unfortunately perhaps too little too late

http://www.techspot.com/news/46389-report-amd-cancels-28nm-apus-from-globalfoundries.html

 

[Olikan]

The yield issue has more to do with the density of the APU design, namely the graphical portion of it, and how demanding it's been for GloFo to produce them at high yields. Given the low # of cache I'm not sure how big this Ghz and clockspeed issue is for a product like the Llano. Frankly, I think it's a great little chip that's a decent performer given the TDP, hefty graphical performance and low clockspeed.

actually, nop. it's the cpu. it was designed to give >3.0Ghz

here is another proof: http://www.fudzilla.com/processors/item/22390-bulldozer-yields-are-good

 

[pelov]

I'm not sure you realize that you're just proving that it's an APU thing and not a CPU thing. You may want to reread my post and then that link again.

AMD didn't ditch GloFo's 28nm to fab their APUs because yields were awesome. And Apple certainly didn't choose to go back to Intel because yields were awesome.

 

[denev2004]

A 4-way 6282 SE matches a 4-way E7-4870 quite well in Specint and FP. Even through Sandy Bridge-EP is going cover a bit of the ground with IPC and cpu frequency increases, it'll still not match Westmere-EX 2 extra cores.

A 4-way 6282 SE with AVX matches 4-way E7-4870 without AVX in FP performance?

That's not really "quite well" to some extend.

But I'd like to say if that's ture, AMD has reached their goal of providing enough INT performance for certain kinds of server, database for instance.

 

[Vesku]

It all points to the biggest hurdle being the 32nm fabrication. It's almost Phenom I deja vu minus TLB bug. Except CMT is a bit less disappointing than "native quad" out in the wild, imo.

 

[piesquared]

Speaking of TLB bug, how's intel coming along with their broken virtualization? Just going to keep selling the broken silicon and let marketing gloss it over?

 

[denev2004]

Speaking of TLB bug, how's intel coming along with their broken virtualization? Just going to keep selling the broken silicon and let marketing gloss it over?

Sandy Bridge-E VT-d Broken In C1 Stepping, Fixed In C2 Stepping, Shortly After Launch
Actually It's just like how AMD is dealing with TLB problem.
Only one thing different is AMD set up a new name for the bug-fixed version while Intel didn't.

 

[wlee15]

A 4-way 6282 SE with AVX matches 4-way E7-4870 without AVX in FP performance?

That's not really "quite well" to some extend.

But I'd like to say if that's ture, AMD has reached their goal of providing enough INT performance for certain kinds of server, database for instance.

Well for Bulldozer it's going to be FMA4 that's beneficial not AVX, and we're comparing AMD's top dog to Intel's top dog(at least for another 9 months to a year)

 

[piesquared]

Sandy Bridge-E VT-d Broken In C1 Stepping, Fixed In C2 Stepping, Shortly After Launch
Actually It's just like how AMD is dealing with TLB problem.
Only one thing different is AMD set up a new name for the bug-fixed version while Intel didn't.

There's another difference. AMD stopped shipment of their faulty products, while intel's market position allows them to continue selling theirs, regardless of bugs.

 

[Duvie]

According to the owner of hardware canucks:



So intel's AMD killer is selling at a measly 3:1 ratio. There's nothing there to be impressed about for intel considering every review on the net has proclaimed AMD's death due to Bulldozer's 'performance'. Gotta love the propaganda, keep up the good work it seems to be working in AMD's favor!

And this



http://www.xtremesystems.org/forums...d-is-also-successfully-flying-off-the-shelves.....


Also just as a reminder, intel's virtualization is broken in sb.

Well consider the 2500K and 2600K have been out for a better part of the year and they still pull 3:1 ratio on a new product.....AMD numbers already lower will take a nose dive based on the reviews and natural progression..I would say that sucks...Could be 4:1 next week and then 5:1 after that...Sound like an AMD fanboy to spin that into Intel propaganda.

Consider also the fact that the only area I see sell well for AMD in box stores is budget PCs...and the FX is no budget CPU...and even there more people will buy the Intel boxes by a 6:1 margin like INtels reported market share in the desktop arena....


The fact that AMD seems to be backing away from "competing against Intel"....I would say the AMD Killer moniker may be true. Which is unfortunate, because I think we need them around to keep the cpu market staying innovative and reasonably price.

 

[denev2004]

Well for Bulldozer it's going to be FMA4 that's beneficial not AVX, and we're comparing AMD's top dog to Intel's top dog(at least for another 9 months to a year)

Yeah I forgot that = = But still some FP tests have been optimized for FMA

There's another difference. AMD stopped shipment of their faulty products, while intel's market position allows them to continue selling theirs, regardless of bugs.

Oh really? That sounds so bad.


But anyway, this has nothing to do with me because P67/H67/Z68/H61 all do not support VT-d and I don't know where can I buy Q67/Q65/B65.

 

[Duvie]

Speaking of TLB bug, how's intel coming along with their broken virtualization? Just going to keep selling the broken silicon and let marketing gloss it over?

I think you look like a bit of a troll here....

The OPs thread has nothing to do with INtel cpus but AMD's issue with the bulldozer chip....I would say stick with the topic instead of deflecting criticism to SB shortcomings in an area that a vast majority of SB users could care less about....

 

[piesquared]

Well consider the 2500K and 2600K have been out for a better part of the year and they still pull 3:1 ratio on a new product.....AMD numbers already lower will take a nose dive based on the reviews and natural progression..I would say that sucks...Could be 4:1 next week and then 5:1 after that...Sound like an AMD fanboy to spin that into Intel propaganda.

Consider also the fact that the only area I see sell well for AMD in box stores is budget PCs...and the FX is no budget CPU...and even there more people will buy the Intel boxes by a 6:1 margin like INtels reported market share in the desktop arena....


The fact that AMD seems to be backing away from "competing against Intel"....I would say the AMD Killer moniker may be true. Which is unfortunate, because I think we need them around to keep the cpu market staying innovative and reasonably price.

That's one way to look at it; the most pessimistic way. Another way is to say that the product is still ramping and with the 32nm issues that so many like to point out (i'd be very surprised if you haven't heard of them), sharing that node with server and APU's out of 1 Fab, it's much more likely that sales will increase not decrease.

 

[cbn]

Two of the improved Stars core that are in Llano, the one's that get 3-6% better IPC than the cores already in Thuban, are roughly the same size as one bulldozer module when you include the L2$.

One can make the argument they could swap out 4 bulldozer modules in exchange for 8 Llano cores and Zambezi would remain nearly the same diesize.

1 BD Module + 2MB L2$ = 30.9 mm^2

2 Llano Cores + 2MB L2$ = 35.4 mm^2

The problem with this analysis is that for whatever reasons those Llano cores suck in terms of clockspeed and power-consumption. A mere 2.9GHz and the quadcore gulps down the juice when you do CPU intensive stuff.

So what would an 8-core Llano look like in terms of clockspeed and power-consumption? It would not have been pretty. Something is not right with GloFo's process and I think it shows in both Llano and Bulldozer.

So the Bulldozer concept kinda works if we make a comparison to Llano's stars core? Would it be safe to Bulldozer is a design that essentially is trying to reverse Pollack's rule by going narrower, but kinda wrecks that with the high clock speed targets?

Or maybe the high clock speed target with 32nm is intentional, allowing AMD to scale back clock speed even more when the design is built on 22nm planar xtors?

I know the idea of lowering Bulldozer clockspeeds on 22nm planar xtor sounds pitiful, but I am really beginning to wonder how much heat density is factoring into the engineers strategies?

Maintaining single thread performance sounds like a tough job when every node jump reduces silicon surface area by 50%, but power consumption only decreases 25%? (Intel's 22nm trigate 3d xtor being the exception. With its 50% power consumption reduction, a 50% die size reduction shouldn't have negative repercussions on heat density. In fact, maybe the 3d shape of the trigate xtor even helps increase surface area to help heat dissipation compared to planar?)

AMD should have avoided the whole CMT thing, reduced the core count from 8 to 6 but made them real cores.

Of course we are talking about people who couldn't count transistors correctly, nor realize in advance that their microarchitecture was going to suck if they didn't get MS to tune their scheduler before the chip was launched.

So we can't really hold their feet to the fire, its obvious they put their A-team on Brazos, their B-team on Llano, and the leftovers got shoved onto operation derpdozer.

Maybe AMD was afraid Stars wouldn't be able to maintain single threaded performance with the transition to 22nm planar xtors?

 

[Duvie]

That's one way to look at it; the most pessimistic way. Another way is to say that the product is still ramping and with the 32nm issues that so many like to point out (i'd be very surprised if you haven't heard of them), sharing that node with server and APU's out of 1 Fab, it's much more likely that sales will increase not decrease.

I dont think anyone would doubt that AMD can fine tune this...The issue is most can agree that AMD will not be able to fine tune it enough to live up to the prebilling of this chip....

The price point is high and the performance isnt in line. I think the FX at its current price point, cost of supporting chipset, etc is too high to move these chips against much cheaper intel setups that it doesn't outperform. I think the current line of FX chips is DOA.

trust me i want AMD to survive...I hae owned many of their chips...Thunderbird, Athlon64, and X2's and opterons 265's....

 

[Idontcare]

Sound like an AMD fanboy to spin that into Intel propaganda.

piesquared is not what I would call an AMD fanboy. Going by the general content of his contributions in this subforum I would be more apt to characterize him as an "anti-Intel fanboy".

He is a big fan of espousing anything that is anti-Intel, that just happens to coincidentally make it seem like he is pro-AMD, but if you check his posts you'll see he rarely posts to ooze over AMD. Mostly posts to state the shortcomings of Intel.

And nothing wrong with that, if there weren't any shortcomings to talk about then he wouldn't have the opportunity to bring them up.

That's Intels problem, hopefully they will address their shortcomings. Until they do it is folks like piesquared that do all of us a service by keeping their feet to the fire.

So the Bulldozer concept kinda works if we make a comparison to Llano's stars core? Would it be safe to Bulldozer is a design that essentially is trying to reverse Pollack's rule by going narrower, but kinda wrecks that with the high clock speed targets?

Or maybe the high clock speed target with 32nm is intentional, allowing AMD to scale back clock speed even more when the design is built on 22nm planar xtors?

I know the idea of lowering Bulldozer clockspeeds on 22nm planar xtor sounds pitiful, but I am really beginning to wonder how much heat density is factoring into the engineers strategies?

Maintaining single thread performance sounds like a tough job when every node jump reduces silicon surface area by 50%, but power consumption only decreases 25%? (Intel's 22nm trigate 3d xtor being the exception. With its 50% power consumption reduction, a 50% die size reduction shouldn't have negative repercussions on heat density. In fact, maybe the 3d shape of the trigate xtor even helps increase surface area to help heat dissipation compared to planar?)



Maybe AMD was afraid Stars wouldn't be able to maintain single threaded performance with the transition to 22nm planar xtors?

I think comparing Zambezi to Llano at the core level for operating voltage (speaks to the design for manufacturing, or DFM, efficiency), for upper-end clockspeeds, and for power-consumption at those clockspeeds is the best way to attempt to divine whether or not there is a 32nm issue holding bulldozer clockspeeds back.

Actually we (not you and me per se, but the community at large) had these same conversations for the months and months leading up to Zambezi's launch.

Noting the apparent step backwards that Llano took compared to Thuban in terms of normalized power-consumption and GHz, the argument was easy to make that bulldozer was not going to hit the 4.5GHz to 5Ghz clocks that were said to be needed for Zambezi to best a stock 2600K. (yes, even before launch, those were the clocks we were talking about it needing)

And so I think you are right in asserting that bulldozer's clocks, while falling slightly short of what was needed, are a testament to the capability of the microarchitecture in light of the canary in the coal-mine observation that Llano is in regards to the capability of GloFo's 32nm process tech.

It really is 65nm dejavu all over again. Remember 65nm Athlon X2's had problems just getting to the same clockspeeds of their 90nm siblings, and before Phenom even launched the biggest concern among enthusiasts was that the K10 would not be able to hit the intended clockspeeds because 65nm seemed to be a miss.

 

[pelov]

It really is 65nm dejavu all over again. Remember 65nm Athlon X2's had problems just getting to the same clockspeeds of their 90nm siblings, and before Phenom even launched the biggest concern among enthusiasts was that the K10 would not be able to hit the intended clockspeeds because 65nm seemed to be a miss.

This time around there isn't a die-shrink in the near future, at least for the enthusiast CPUs. This puts more emphasis on AMD really getting things right with Piledriver and GloFo vastly improving their 32nm so AMD won't shift their entire operations over to TSMC.

There a couple stark contrasts between the 65nm and 32nm fiascos. For one, AMD seems to have hyped Bulldozer to be the be-all-end-all of the enthusiast market. It was a disappointment even if we put all the hype aside. Secondly, there was a considerable decrease in IPC going to 32nm BD from the Phenom II 45nm. That's not good at all.

I really hope they get things sorted out, the both of them. But they've got quite a long road ahead of them.

You can throw me into the "Anti-Intel" category. I just hate the way they operate their business; both illegal and legal means but often it just feels so dirty...

 

[Duvie]

piesquared is not what I would call an AMD fanboy. Going by the general content of his contributions in this subforum I would be more apt to characterize him as an "anti-Intel fanboy".

He is a big fan of espousing anything that is anti-Intel, that just happens to coincidentally make it seem like he is pro-AMD, but if you check his posts you'll see he rarely posts to ooze over AMD. Mostly posts to state the shortcomings of Intel.

And nothing wrong with that, if there weren't any shortcomings to talk about then he wouldn't have the opportunity to bring them up.

That's Intels problem, hopefully they will address their shortcomings. Until they do it is folks like piesquared that do all of us a service by keeping their feet to the fire.



I think comparing Zambezi to Llano at the core level for operating voltage (speaks to the design for manufacturing, or DFM, efficiency), for upper-end clockspeeds, and for power-consumption at those clockspeeds is the best way to attempt to divine whether or not there is a 32nm issue holding bulldozer clockspeeds back.

Actually we (not you and me per se, but the community at large) had these same conversations for the months and months leading up to Zambezi's launch.

Noting the apparent step backwards that Llano took compared to Thuban in terms of normalized power-consumption and GHz, the argument was easy to make that bulldozer was not going to hit the 4.5GHz to 5Ghz clocks that were said to be needed for Zambezi to best a stock 2600K. (yes, even before launch, those were the clocks we were talking about it needing)

And so I think you are right in asserting that bulldozer's clocks, while falling slightly short of what was needed, are a testament to the capability of the microarchitecture in light of the canary in the coal-mine observation that Llano is in regards to the capability of GloFo's 32nm process tech.

It really is 65nm dejavu all over again. Remember 65nm Athlon X2's had problems just getting to the same clockspeeds of their 90nm siblings, and before Phenom even launched the biggest concern among enthusiasts was that the K10 would not be able to hit the intended clockspeeds because 65nm seemed to be a miss.

Maybe not....However I still dont see what that does to contribnute to the OPs thread here...seems like a side rant...He can start his own thread and bitch about it there in one location and we can discuss it there...

 

[formulav8]

How many posts here have stuck to the op? Mine and yours included.

 

[piesquared]

I dont think anyone would doubt that AMD can fine tune this...The issue is most can agree that AMD will not be able to fine tune it enough to live up to the prebilling of this chip....

The price point is high and the performance isnt in line. I think the FX at its current price point, cost of supporting chipset, etc is too high to move these chips against much cheaper intel setups that it doesn't outperform. I think the current line of FX chips is DOA.

trust me i want AMD to survive...I hae owned many of their chips...Thunderbird, Athlon64, and X2's and opterons 265's....

The sales numbers provided show that there is indeed a market for these chips regardless of all else. There will always be a market for them, just like there there is a market for intel's abysmal graphics. And since the FX line is aimed at the high end desktop, which is a small and dying market, the volume is inherently low. APU's cover the volume market and consumers seem to be buying them in droves. They aren't listed in the sales figures I quoted, but these are the direct competitors of intel's line.

 

[jones377]

The sales numbers provided show that there is indeed a market for these chips regardless of all else. There will always be a market for them, just like there there is a market for intel's abysmal graphics. And since the FX line is aimed at the high end desktop, which is a small and dying market, the volume is inherently low. APU's cover the volume market and consumers seem to be buying them in droves. They aren't listed in the sales figures I quoted, but these are the direct competitors of intel's line.

There is 'someone' (who I bet you loathe) that uses the same kind of argument to defend Itanium year after year

 

[Cerb]

So the Bulldozer concept kinda works if we make a comparison to Llano's stars core? Would it be safe to Bulldozer is a design that essentially is trying to reverse Pollack's rule by going narrower, but kinda wrecks that with the high clock speed targets?

No. Pollack's rule is an observation about consequences of decisions necessary to keep processors performing faster than the last ones. It's one of those observations that tends to be approximately correct, but it is not a hypothesis capable of prediction. IE, it doesn't say too much about this time around, but rather observes that what has been has been approximately a x to sqrt(x) ratio of complexity increase to performance, on average; therefore one should expect to need around 2x the complexity for 40% more performance, on average. Narrower can be more efficient, but keeping a narrower core full, making it faster than prior ones, will still need more complicated state management, which will mean more xtors, more wires, and thus more area.

They were trying to get as much as they could with what little resources they had (which, after kicking Ruiz out, was even less than normal), and if it weren't for the high power consumption and high latencies (the latter of which they really should have avoided, even if at the expense of speed, IMO), AKA, "you meddling kids," they might have gotten away with it.

The speeds did kind of wreck it though. Given that AMD has such a history of not being able to run their chips fast enough, I can't sympathize with the thought that that was all GF's fault (for Llano, I can go with that sentiment, even if AMD did make most of the 32nm decisions before the split). I also wonder, if they could get a few latencies under control, mainly RAM and caches, if BD might at least fare pretty well at moderate speeds.

Maybe AMD was afraid Stars wouldn't be able to maintain single threaded performance (...)

It already can't. It's a souped-up K8, which was a souped-up K7, and they were pushing it about as far as it could go, already. Incrementally improving on it, v. a major change like BD may or may not have been the better way to go, but some real change in direction was needed, ASAP. If you want to be positive about it, there may be a chance that it could end up far better in the next iteration (PD), not having a middle CPU (IE, Llano) sucking up R&D.