Has Xbitlabs already benched Bulldozer?

[dac7nco]

I'm certain at this point that Zambezi will be an adequate CPU for gaming and mundane tasks, and excel for everything else including the kitchen sink. I've had a feeling that the quicksync under Sandy/Ivy Bridge was an emergency response (in 2008!) to Bulldozers possible transcoding performance. I have no doubt that a pair of 3GHz socket-C32 CPUs will own a single SB-E of any variety on Handbrake and Lightwave, and for a lot less cash, plus Dual Intel NICs and ECC support.

I'm an Intel guy, but these benches we're seeing all look a bit suspect. AMD IS behind Intel in management, cash and recruitment, but do you think they would place their bets on a CPU that performs like a doubled-up Phenom-1, which would put them in the dirt forever? AMD doesn't go out of their way to hire stupid people... they do get unlucky at times with what they tell them to say.

Daimon

 

[Arg Clin]

I'm certain at this point that Zambezi will be an adequate CPU for gaming and mundane tasks, and excel for everything else including the kitchen sink. I've had a feeling that the quicksync under Sandy/Ivy Bridge was an emergency response (in 2008!) to Bulldozers possible transcoding performance. I have no doubt that a pair of 3GHz socket-C32 CPUs will own a single SB-E of any variety on Handbrake and Lightwave, and for a lot less cash, plus Dual Intel NICs and ECC support.

I'm an Intel guy, but these benches we're seeing all look a bit suspect. AMD IS behind Intel in management, cash and recruitment, but do you think they would place their bets on a CPU that performs like a doubled-up Phenom-1, which would put them in the dirt forever? AMD doesn't go out of their way to hire stupid people... they do get unlucky at times with what they tell them to say.

Daimon

I totally agree. Common sense highly suggests that the supposed leaked benches will not be representative of real world performance of BD. In fact, if BD was that poor, AMD would have done better to just scrap it and focus on a simple die shrink of Phenom II. I wouldn't make any sense to bring a CPU of the diesize of BD to the market, if it's only going to compete with the likes of Core i3. That would just not be profitable.

AMD didn't survive this long by making stupid decisions - in fact they have been forced by the sheer size and somewhat questionable business tactics of Intel over the years to be extremely brilliant just to stay alive.

 

[Vesku]

Or they would scrap it and hold off on desktop release until they could get BD+ out the door.

 

[LOL_Wut_Axel]

I'm certain at this point that Zambezi will be an adequate CPU for gaming and mundane tasks, and excel for everything else including the kitchen sink. I've had a feeling that the quicksync under Sandy/Ivy Bridge was an emergency response (in 2008!) to Bulldozers possible transcoding performance. I have no doubt that a pair of 3GHz socket-C32 CPUs will own a single SB-E of any variety on Handbrake and Lightwave, and for a lot less cash, plus Dual Intel NICs and ECC support.

I'm an Intel guy, but these benches we're seeing all look a bit suspect. AMD IS behind Intel in management, cash and recruitment, but do you think they would place their bets on a CPU that performs like a doubled-up Phenom-1, which would put them in the dirt forever? AMD doesn't go out of their way to hire stupid people... they do get unlucky at times with what they tell them to say.

Daimon

I totally agree. Common sense highly suggests that the supposed leaked benches will not be representative of real world performance of BD. In fact, if BD was that poor, AMD would have done better to just scrap it and focus on a simple die shrink of Phenom II. I wouldn't make any sense to bring a CPU of the diesize of BD to the market, if it's only going to compete with the likes of Core i3. That would just not be profitable.

AMD didn't survive this long by making stupid decisions - in fact they have been forced by the sheer size and somewhat questionable business tactics of Intel over the years to be extremely brilliant just to stay alive.

You really think that its that fast even though it needs eight cores to compete with Intel's four...? They won't put so many cores on it and wouldn't clock it so high unless they absolutely needed to.

 

[dac7nco]

What I posted was Handbrake and rendering performance, and that I expected moderate performance elsewhere. I didn't post "Buldozer is the shizzle and is going to own Sandy Bridge-E". Misread this: In highly-threaded workloads Zambezi will be fast. I use a 12-core 2.6GHz dual-C32 system at work which is an excellent workstation using two ugly Istanbul CPUs.

Daimon

 

[evident]

my gut feeling tells me that the bulldozer release is one of those movies that were so crappy that advance screenings weren't allowed to be shown to movie critics.
i seriously hope i'm wrong though because i want intel to have some serious competition for once.

 

[Deanodarlo]

You really think that its that fast even though it needs eight cores to compete with Intel's four...? They won't put so many cores on it and wouldn't clock it so high unless they absolutely needed to.

I thought one Module (with two cores, shared FPU) was the same as a single Intel hyperthreading core in terms of price - AMD always intended that to be so.

A single module is bigger than a single Intel core, but smaller than Intel dual cores.

They have sacrificed single thread for future performance in multi-threaded applications. Intel have a year and a half or so lead in fab technology - AMD had to come up with a different approach where they could compete. Single threaded performance core vs core was never going to be practical. They lost that race when the Core2 came out and now they are going after APU's, low power solutions and multi-threaded environments.

I don't think you can compare a single Intel core to a single AMD core anymore - just like you couldn't compare a P4 Ghz to AMD Ghz in the P4/Athlon era - they have gone different design routes. Like for like is now on different parameters.

One AMD module = one Intel core
8 core Bulldozer = 4 core Sandy Bridge

That's how they will be priced.

One thing we can compare them on is performance per watt. Which CPU is using power and design more efficiently for the best results in today's applications and server environments? What about tomorrows?

Anyone know the transistor count and die size of a 4 module Bulldozer vs a 4 core Sandy bridge? Interesting to also know which one costs more to produce!

AMD may have got it completely wrong like Intel did with the P4 - but they simply can't compete by producing like-for-like, so they've gone for a module split into two smaller cores vs core + hyperthreading.

 

[aigomorla]

Movieman has already stated his next build will be using a bulldozer cpu after he saw some real benches, so maybe there's merit to this article?

Movieman is a great friend of mine.

I know him in real life... infact i know a lot of the people u guys go WOW to about scores in Real Life.. :biggrin:

He wants a 48p Magoney... straight up 12 x 4 monster ... he's a core whore...

He knows he cant win in straight speed, so he wants to win in massive coreage...

 

[Arg Clin]

You really think that its that fast even though it needs eight cores to compete with Intel's four...? They won't put so many cores on it and wouldn't clock it so high unless they absolutely needed to.

One thing is a tech analysis - another is common business sense. If AMD wants to bring this product to the market, it needs to perform competitively in real world applications. If it didn't they might as well just do a die shrink of the K10.5 - much cheaper than a new architecture. Remember that the vast majority of end users don't care how they get there, they just want the most FPS in games or that it encodes their DVD rips quickly.

 

[Accord99]

One thing is a tech analysis - another is common business sense. If AMD wants to bring this product to the market, it needs to perform competitively in real world applications. If it didn't they might as well just do a die shrink of the K10.5 - much cheaper than a new architecture.

Or it's entirely possible that they thought it would be better than a die shrink of K10.5 and now that it's here and under-performing, they don't have the time or resources to do a K10.5 die shrink.

I doubt Intel thought Prescott was going to be that hot at first but when it did arrive, they weren't able to rush a Northwood shrink.

 

[Idontcare]

Or it's entirely possible that they thought it would be better than a die shrink of K10.5 and now that it's here and under-performing, they don't have the time or resources to do a K10.5 die shrink.

I doubt Intel thought Prescott was going to be that hot at first but when it did arrive, they weren't able to rush a Northwood shrink.

I thought Llano was a K10.5 shrink + IGP. Even has a 6% IPC boost over 45nm K10.5 (so would that make it a K10.75? )

 

[Accord99]

I thought Llano was a K10.5 shrink + IGP. Even has a 6% IPC boost over 45nm K10.5 (so would that make it a K10.75? )

True... To clarify, what I meant was that AMD didn't have the resources to develop a 6-core or 8-core K10.5 die shrink that would be ready for 2011 once/if they realized that BD was under-performing.

 

[Idontcare]

True... To clarify, what I meant was that AMD didn't have the resources to develop a 6-core or 8-core K10.5 die shrink that would be ready for 2011 once/if they realized that BD was under-performing.

It would have been an utter PR and marketing nightmare even if they could do it economically and technically after all the ballyhoo that has been invested into bulldozer.

Intel's marketing team would make a lot of hay if AMD stepped back from bulldozer and instead opted to just double-down on a well played-out K10 horse.

 

[Abwx]

I thought Llano was a K10.5 shrink + IGP. Even has a 6% IPC boost over 45nm K10.5 (so would that make it a K10.75? )

As already pointed, it has better IPC thats its predecessors, sometimes
by roughly more than 6%....
A die shrink of the X6 using this core would had made sense
since it would be about 200mm2..

 

[Accord99]

It would have been an utter PR and marketing nightmare even if they could do it economically and technically after all the ballyhoo that has been invested into bulldozer.

I hadn't thought of this factor but this would certainly be important for the decision-making process.

 

[sm625]

Is AMD that close to parity for wprime? If I did my math right, a 3.23 GHz llano should outscore this 3.1 GHz i5-2400. Seems odd dont it?

 

[Abwx]

Is AMD that close to parity for wprime? If I did my math right, a 3.23 GHz llano should outscore this 3.1 GHz i5-2400. Seems odd dont it?

That s a favourable case since Wprime use a lot of integer arithmetic operations and it happens that Llano core has two (amongst the three)
ALUs that have an added hardware multiplier and divider , and this
help a lot in this bench.
Indeed, i rest my case when i said that this core did
deserve some 32nm shrink..

 

[Arg Clin]

Or it's entirely possible that they thought it would be better than a die shrink of K10.5 and now that it's here and under-performing, they don't have the time or resources to do a K10.5 die shrink.

I doubt Intel thought Prescott was going to be that hot at first but when it did arrive, they weren't able to rush a Northwood shrink.

Bulldozer has been underway for how long? You don't think they've done any kind of evaluation underway? They would have been dead long ago if they had that poor R&D management. Intel isn't really comparable to AMD in this aspect. Intel just needs to put out something new every once in a while. No matter how small the performance gain is, corporate buyers will bite simply because it's newer than what they have. AMD is not in this fortunate position - they have to deliver on performance every single time, so adopting the same strategy as Intel might have on the Prescott would be suicide.

It would have been an utter PR and marketing nightmare even if they could do it economically and technically after all the ballyhoo that has been invested into bulldozer.

Intel's marketing team would make a lot of hay if AMD stepped back from bulldozer and instead opted to just double-down on a well played-out K10 horse.

And an underperforming BD wouldn't be "an utter PR and marketing nightmare"?

In fact, a shrinked K10.5 would just be renamed to an FX name, and only the tech savy buyers would even notice. Marketing would be the least of their concerns.

 

[Accord99]

Bulldozer has been underway for how long? You don't think they've done any kind of evaluation underway?

Maybe they did, maybe it look good early on. Maybe the Powerpoint slides looked really good, or it had significant supporters at AMD management who would wave away any problems found by saying it would be fixed in the future.

It wouldn't be the first time that some "revolutionary" product ended falling short of expectation even after having huge amounts of resources poured into it.

They would have been dead long ago if they had that poor R&D management. Intel isn't really comparable to AMD in this aspect.

In reality, AMD should have been dead a long time, other than for the exceptional generosity of shareholders and the UAE.

Intel just needs to put out something new every once in a while. No matter how small the performance gain is, corporate buyers will bite simply because it's newer than what they have. AMD is not in this fortunate position - they have to deliver on performance every single time, so adopting the same strategy as Intel might have on the Prescott would be suicide.

It seems to me starting from 2006, Intel has been steadily increasing its performance advantage over AMD.

 

[nagaz]

I'm certain at this point that Zambezi will be an adequate CPU for gaming and mundane tasks, and excel for everything else including the kitchen sink. I've had a feeling that the quicksync under Sandy/Ivy Bridge was an emergency response (in 2008!) to Bulldozers possible transcoding performance. I have no doubt that a pair of 3GHz socket-C32 CPUs will own a single SB-E of any variety on Handbrake and Lightwave, and for a lot less cash, plus Dual Intel NICs and ECC support.

I'm an Intel guy, but these benches we're seeing all look a bit suspect. AMD IS behind Intel in management, cash and recruitment, but do you think they would place their bets on a CPU that performs like a doubled-up Phenom-1, which would put them in the dirt forever? AMD doesn't go out of their way to hire stupid people... they do get unlucky at times with what they tell them to say.

Daimon

I don't know what to expect but i think Zambezi will be good as you said.

Don't talking about AMD becouse i'm not interested in this contest there are out there many bigger companies (and i say BIGGER) that use the service you described and is documentated.
And what you call supid people are in reality other companies that provide that service and they do it carefully and professionally. Low cost doesn't mean poor and sometimes remain to be seen what is low.
So about me you haven't the correct perception about that stupid thing.

Anyway you can choose to look out there or not.

 

[gplnpsb]

Anyone know the transistor count and die size of a 4 module Bulldozer vs a 4 core Sandy bridge? Interesting to also know which one costs more to produce!

Semiaccurate reported a Orochi die size of 315 mm2 here.

Other estimates based on cache dimensions and photos of Orochi wafers were in the range of 290 mm2. So it's probably safe to say the die is around 300 mm^2. The transistor count is unknown, but AMD stated in their ISSCC abstract that one module with its associated L2 cache is 213 million transistors. That would put the transistor count for the modules alone at 852 million. I'd estimate that 8 MB of L3 cache would take around 610 million transistors (from the 458 million transistor difference between deneb and propus for 6 MB L3 cache). That would make the Bulldozer transistor count 1.46 billion at a minimum. More transistors would have to be spent on things like hypertransport links, and memory controllers.

This is in comparison to Sandy Bridge, which has a transistor count of 995 million, including a gpu. The gpu has around 114 million transistors, putting the count for the cpu portion at 881 million. The four core sandy bridge with GT2 graphics is 216 mm2.

Now, AMD has been doing alright selling gigantic 346 mm^2 thubans for a while, but those numbers are a bit concerning. It would kind of disappointing if all those transistors in all that die space can only match four Intel cores with hyperthreading (and half the total cache).

One silver lining is that the SB-E dies are enormous as well, becuase of their quad channel memory controllers, 40 lanes of on die pci-express, and 2.5 MB of L3 cache per core. The eight core die is around 400 mm^2. Judging by Intel's MDDS database there is a 4 core die as well, but no six core die as of yet. Searching for Q-Spec Q19J, that of a LGA 2011 six core that showed up on ebay, indicates that it is based off the eight core die.

Based on sandy bridge's scalability, and a low res SB-E die shot, I'd guess that the four core die is around 270 mm^2.

 

[Deanodarlo]

Semiaccurate reported a Orochi die size of 315 mm2 here.

Other estimates based on cache dimensions and photos of Orochi wafers were in the range of 290 mm2.....

Thank you for that info - very interesting.

That's how I will judge if the CPU design is a winner, plus performance per watt and overall design efficiency for its intended purpose.

They are obviously chasing the benchmarks where hyperthreading gives a boost to Intel - if it fails to gain a decent advantage against multithreaded loads after deciding to move away from challenging single thread performance, it will be an utter design failure. This new module approach has to give appreciable gains for what it's targeted to do and the extra space/cost it requires to achieve two core modules vs a single Intel core.

Especially when Intel have a GPU and dedicated video rendering and media hardware built in as well - they have space to spare for more CPU only additions if they choose while AMD will have to cut down a Bulldozer core to make it viable.

How will AMD's Bulldozer APU's compare? Loosing CPU benchmarks by a massive margin probably, while offering better graphics. The problem is, for desktop use everyone adds a dedicated graphics card.

Thinking about it, it looks like AMD will indeed be giving up the dying desktop market and be concentrating on APU mobile devices, specific multithreaded workloads/applications/servers where they have a chance to challenge Intel at completive prices.

If you want the best desktop CPU for general use, it will probably be Intel from here on in. AMD will be filling in the gaps where Intel has weaknesses (value orientated PC market, built-in graphics/stream processing APU's mainly for mobile devices, specific multi-threaded workloads where one module beats one hyperthreading core for a similar or cheaper price).

We'll find out in a few weeks or sooner given a leak, but I feel we already know what's coming.

 

[Arg Clin]

Maybe they did, maybe it look good early on. Maybe the Powerpoint slides looked really good, or it had significant supporters at AMD management who would wave away any problems found by saying it would be fixed in the future.

It wouldn't be the first time that some "revolutionary" product ended falling short of expectation even after having huge amounts of resources poured into it.

I respectfully doubt this is the case. This is AMD - not some random happy-go-lucky techie with a good idea.

In reality, AMD should have been dead a long time, other than for the exceptional generosity of shareholders and the UAE.

Shareholders aren't generous, they are investors

It seems to me starting from 2006, Intel has been steadily increasing its performance advantage over AMD.

It seems to me that AMD has kept up pretty well in terms of performance pr $ spent in the mid end market, up until the release of SB. Remember that while performance crown gives good marketing, it's far more commercially interesting who provides the better value in the $ 200 segment.

 

[podspi]

Semiaccurate reported a Orochi die size of 315 mm2 here.

Other estimates based on cache dimensions and photos of Orochi wafers were in the range of 290 mm2. So it's probably safe to say the die is around 300 mm^2. The transistor count is unknown, but AMD stated in their ISSCC abstract that one module with its associated L2 cache is 213 million transistors. That would put the transistor count for the modules alone at 852 million. I'd estimate that 8 MB of L3 cache would take around 610 million transistors (from the 458 million transistor difference between deneb and propus for 6 MB L3 cache). That would make the Bulldozer transistor count 1.46 billion at a minimum. More transistors would have to be spent on things like hypertransport links, and memory controllers.

This is in comparison to Sandy Bridge, which has a transistor count of 995 million, including a gpu. The gpu has around 114 million transistors, putting the count for the cpu portion at 881 million. The four core sandy bridge with GT2 graphics is 216 mm2.

Now, AMD has been doing alright selling gigantic 346 mm^2 thubans for a while, but those numbers are a bit concerning. It would kind of disappointing if all those transistors in all that die space can only match four Intel cores with hyperthreading (and half the total cache).

One silver lining is that the SB-E dies are enormous as well, becuase of their quad channel memory controllers, 40 lanes of on die pci-express, and 2.5 MB of L3 cache per core. The eight core die is around 400 mm^2. Judging by Intel's MDDS database there is a 4 core die as well, but no six core die as of yet. Searching for Q-Spec Q19J, that of a LGA 2011 six core that showed up on ebay, indicates that it is based off the eight core die.

Based on sandy bridge's scalability, and a low res SB-E die shot, I'd guess that the four core die is around 270 mm^2.

I think it is important to point out that

1) According to rumors, there is only one die for client of server, so Zambezi might have a lot of extra stuff unnecessary for client CPUs (to support MP, etc)

2) From what I've heard (also rumor of course!) the layout of BD is such that there is room for a GPU/another module, and so a lot of that space on the die is empty (no transistors).


Both of those things might affect your analysis, since Intel has enough volume to have separate dies for its Xeons and i7s, and not all the die-space in Zambezi may actually be used (which from what I can tell, you assumed).



All of that being said, I too find it very disappointing (esp as a shareholder) AMD is going to be selling a 300mm2 CPU that from the looks of it (via rumor again!) has trouble competing with the 2600K.

Maybe we'll find out this Thursday?

 

[Abwx]

All of that being said, I too find it very disappointing (esp as a shareholder) AMD is going to be selling a 300mm2 CPU that from the looks of it (via rumor again!) has trouble competing with the 2600K.

This is even more surpising when taking account of GloFo s process
that has as much as 40% higher density than intel s 32nm..:biggrin: