The economics of Bulldozer...

[Arachnotronic]

So I was thinking, everyone's thinking that 4 module/8 core BD is going up against 4 core SNB, right? If Bulldozer 8 core isn't faster than SNB, we have a serious problem, and here's why:

How many transistors in a Bulldozer? 213/module * 4 modules = 852 million transistor. Then add in 8MB L3 cache. To roughly estimate this, I reference http://www.anandtech.com/show/2836 which says a Phenom II x4 has 758 mil transistor and an Athlon II x4 has 300M. So that's 758 - 300 = 458 mil for 6 mb L3 cache, so this gives us roughly 610 cache transistors. Adding we see: 852 + 610 = 1462 million transistors.

So if BD 8C is slower than SNB 4C (which has roughly 850 mil total CPU transistors), then Intel can provide the same or better performance with nearly half the transistors. This would be disastrous to AMD. This is why people claiming an 8C BD will sell for $320 is AMD's attempt to "gain marketshare" is ridiculous -- the costs of building this thing are HUGE!

Thoughts?

 

[frostedflakes]

The thing about Intel is that they expect relatively large profit margins on all the products they sell. AMD is fine with smaller profit margins. Considering their position in the CPU market, they can't afford to be picky about margins. They need to price stuff competitively to get people using their products.

This is something I've thought about, though. And you have to keep in mind that the GPU in SB takes up a fair amount of die space as well. If you get rid of that (which Intel plans to do with the SB-E chips), that frees up room for another two CPU cores that you could add without even increasing die size compared to current chips. It seems like it would be extremely trivial for Intel to sell 32nm products with 8 cores or possibly even more, as long as clock speeds were lowered enough to keep TDP in check. Whether AMD will provide enough competition for Intel to bother releasing SB chips with more cores is another issue, though.

 

[Idontcare]

The cost to actually produce cpu's is typically vastly over-rated and over-estimated.

Here's a walk-through of a back-of-the-envelope cost calc for two chips from a while back.

The absolute numbers change up every node and for the relative die-sizes but the take home message remains the same.

The business overhead is where all the expenses come in which result in the necessity of ASP's in the triple-digit range in order for the business to turn a profit.

This is why the added expense of SOI has never really been a big issue for AMD, it is a marketing talking point for Intel but if there really was as big of a cost issue with SOI as Intel makes it out to be then you can be guaranteed that AMD would have migrated away from SOI years and years ago.

AMD's problems are not the manufacturing expense associated with their processors. The expense associated with developing new process nodes combined with the challenge of getting their ASP's up into triple digits (and I don't mean $101) is what dooms their profitability.

Nvidia makes large GPU's to compete against AMD's smaller ones. Its not the end of the world for Nvidia either. To be sure it is not ideal, they'd love to lower production costs as well as increase ASP's.

Just about everyone in this industry will gladly raise production costs by 10% if it means they can produce a more competitive chip that enables them to raise the ASP by 10%.

 

[Idontcare]

This is something I've thought about, though. And you have to keep in mind that the GPU in SB takes up a fair amount of die space as well. If you get rid of that (which Intel plans to do with the SB-E chips), that frees up room for another two CPU cores that you could add without even increasing die size compared to current chips. It seems like it would be extremely trivial for Intel to sell 32nm products with 8 cores or possibly even more, as long as clock speeds were lowered enough to keep TDP in check.

Not sure how relevant they are anymore, but I made a handful of photoshopped diemaps of various hypothetical SB chips of differing core counts and layouts in this post from last fall. There might be something useful in that thread.

 

[Arkadrel]

So I was thinking, everyone's thinking that 4 module/8 core BD is going up against 4 core SNB, right? If Bulldozer 8 core isn't faster than SNB, we have a serious problem, and here's why:

How many transistors in a Bulldozer? 213/module * 4 modules = 852 million transistor. Then add in 8MB L3 cache. To roughly estimate this, I reference http://www.anandtech.com/show/2836 which says a Phenom II x4 has 758 mil transistor and an Athlon II x4 has 300M. So that's 758 - 300 = 458 mil for 6 mb L3 cache, so this gives us roughly 610 cache transistors. Adding we see: 852 + 610 = 1462 million transistors.

So if BD 8C is slower than SNB 4C (which has roughly 850 mil total CPU transistors), then Intel can provide the same or better performance with nearly half the transistors. This would be disastrous to AMD. This is why people claiming an 8C BD will sell for $320 is AMD's attempt to "gain marketshare" is ridiculous -- the costs of building this thing are HUGE!

Thoughts?

Wrong though

4 "core" bulldozer will go up against a 4 "core" (8 thread) Intel CPU. (not a 8c BD vs 4c SB)
Will it be slower in stuff that can use more than 4 threads? probably.
It ll be smaller and cheaper too, most likely.

As long as for stuff that uses 4 threads, its close enough to sandy bridge performance, thats bound to be good for amd. Atleast better than Phenom II is.

 

[Nemesis 1]

The thing about Intel is that they expect relatively large profit margins on all the products they sell. AMD is fine with smaller profit margins. Considering their position in the CPU market, they can't afford to be picky about margins. They need to price stuff competitively to get people using their products.

This is something I've thought about, though. And you have to keep in mind that the GPU in SB takes up a fair amount of die space as well. If you get rid of that (which Intel plans to do with the SB-E chips), that frees up room for another two CPU cores that you could add without even increasing die size compared to current chips. It seems like it would be extremely trivial for Intel to sell 32nm products with 8 cores or possibly even more, as long as clock speeds were lowered enough to keep TDP in check. Whether AMD will provide enough competition for Intel to bother releasing SB chips with more cores is another issue, though.

We all know that AMD is not the salvation Army as history shows. AMD 64 was faster than intel in many Apps. But not multi threaded until X2 . We all know what price AMD charged when they had the best Arch . So get off this AMD is the value company . They are the value company only when their Arch can't keep pace . Proven truth!

 

[podspi]

The cost to actually produce cpu's is typically vastly over-rated and over-estimated.

Here's a walk-through of a back-of-the-envelope cost calc for two chips from a while back.

:thumbsup: I had no idea MC for these chips was so low! I was expecting more in the ~ $50 to $80 dollar range. Of course, once you factor in other costs I am sure the AC/die is much higher :biggrin:

Wrong though

4 "core" bulldozer will go up against a 4 "core" (8 thread) Intel CPU. (not a 8c BD vs 4c SB)
Will it be slower in stuff that can use more than 4 threads? probably.
It ll be smaller and cheaper too, most likely.

As long as for stuff that uses 4 threads, its close enough to sandy bridge performance, thats bound to be good for amd. Atleast better than Phenom II is.

Personally, I think it will be something like: 4c BD vs 2c(4 thread) SB, 6c BD vs 4c SB, and 8C BD vs 4c (8 thread) SB.

SB pretty much has the same singlethread performance across the board, with only throughput separating the different CPUs.

 

[frostedflakes]

Not sure how relevant they are anymore, but I made a handful of photoshopped diemaps of various hypothetical SB chips of differing core counts and layouts in this post from last fall. There might be something useful in that thread.

Interesting stuff, don't think I had seen it before. Thanks.

We all know that AMD is not the salvation Army as history shows. AMD 64 was faster than intel in many Apps. But not multi threaded until X2 . We all know what price AMD charged when they had the best Arch . So get off this AMD is the value company . They are the value company only when their Arch can't keep pace . Proven truth!

Never claimed that AMD is a "value" company, I remember just as well as anyone here how they charged $1k for the FX-58 and other flagships many years ago. I guess a more accurate statement would have been that competitive pressure has forced AMD to be fine with smaller profit margins. They obviously price their CPUs the way they do because they must.

 

[drizek]

:thumbsup: I had no idea MC for these chips was so low! I was expecting more in the ~ $50 to $80 dollar range. Of course, once you factor in other costs I am sure the AC/die is much higher :biggrin:

Yes. In AMD's case, it is higher than what we pay for them

 

[veri745]

Yes. In AMD's case, it is higher than what we pay for them

The calculations that IDC provided don't include package costs, though, and that is a major portion of the manufacturing cost for a CPU (probably the same or more than the cost of the Si)

Variations in package cost are more dependent on performance and electrical requirements, rather than die size (i.e. a 130W high performance CPU that can pull 100A will cost more than a 35W cpu that will only be pulling 40A through the package).

 

[gevorg]

Who cares about transistors counts? The only thing that matters is price AND performance.

 

[Nemesis 1]

The calculations that IDC provided don't include package costs, though, and that is a major portion of the manufacturing cost for a CPU (probably the same or more than the cost of the Si)

Variations in package cost are more dependent on performance and electrical requirements, rather than die size (i.e. a 130W high performance CPU that can pull 100A will cost more than a 35W cpu that will only be pulling 40A through the package).

You coudn't give me for free a cpu that pulls 100Amps or 40 amps for that matter. My house sockets are only 20 amp

 

[veri745]

you coudn't give me for free a cpu that pulls 100amps or 40 amps for that matter. My house sockets are only 20 amp

115v @ 20a = 2300w

1.1v @ 100a = 110w

I'm talking about current through the package, not current through your outlet.

 

[Nemesis 1]

Who cares about transistors counts? The only thing that matters is price AND performance.

Thats the whole deal . Performance dictates price. Unless they want a price war with intel . That would be a mistake on AMDS part.

 

[AtenRa]

the costs of building this thing are HUGE!

AMD Thuban Hex-Core = 904M transistors at 346mm2 @45nm process

AMD Bulldozer (4 Module) = 1.4/1.5B transistors at 290-300mm2 @32nm process (estimates).

Now if we have the same yields between the two processes, a 4 Module BD will be cheaper because we could have more dies on the 300mm wafer.

 

[Nemesis 1]

115v @ 20a = 2300w

1.1v @ 100a = 110w

I'm talking about current through the package, not current through your outlet.

An AMP is an amp . 25 hp electic motors pull 25 amps at the motor easily .under load But I see where your coming from .

 

[Khato]

The calculations that IDC provided don't include package costs, though, and that is a major portion of the manufacturing cost for a CPU (probably the same or more than the cost of the Si)

More accurately, they don't include packaging and test. The types of defects that can be discovered before a die is packaged unfortunately are in the minority, resulting in the majority of defective die being packaged before they fail any testing.

But the point that pretty much all the expenses are research and capital expenditure is indeed quite correct. That's why Intel cares about its die sizes - smaller die and higher yields mean less equipment to purchase for the same number of products sold. It's not as much of an issue for AMD, at least until Global Foundries can't meet their production demands and they end up capacity constrained... at which point AMD simply doesn't sell as much product as they otherwise would have.

 

[drizek]

They should include a gift certificate for a broken CPU with the purchase of every black edition retail one.

So I get a 2011 stepping of retail bulldozer plus the certificate, and so when the next generation comes out in 2012, I can redeem it for one of the failures. Run it at 1.7 volts until it blows up, then go back to my 2011 stepping.

 

[JFAMD]

All of this is pretty immaterial.

Everyone is thinking that a.) somehow economics are tied to performance (but 95% of the market or more is not buying on raw performance) and that b.) somehow die size and transistor counts are crucial.

The reality is that there are 2 kinds of economics at play, manufacturer and customer.

At the manufacturer level AMD has been in business for more than 40 years and has been profitable for the last few quarters, prospects are looking better every day. Not having to chase fab technology is allowing for better profitability and better future outlooks (more stability).

At the customer level they care about performance, power consumption and price. They don't buy transistors and die sizes.

You really can't mix the two. As soon as you do you are crossing lines and making assumptions that won't hold.

 

[Ben90]

An AMP is an amp . 25 hp electic motors pull 25 amps at the motor easily .under load But I see where your coming from .

An amp is an amp in the same way a volt is a volt.

What kind of black magic makes it so your CPU can run at 1.1V while its plugged into a 120/240V outlet?

 

[drizek]

Oh shi-

You're supposed to run it at 1.1V?

No wonder it keeps sparking...

http://www.youtube.com/watch?v=TjHOUonZO6U

 

[sm625]

I think the $1000 FX-58 was more a remnant of the tech bubble than anything else. I dont think it could happen now even if BD was 30% faster in IPC.

 

[Idontcare]

I think the $1000 FX-58 was more a remnant of the tech bubble than anything else. I dont think it could happen now even if BD was 30% faster in IPC.

Oh? So that is what the 990X is then, a remnant of the tech bubble. :hmm:

 

[Topweasel]

Oh? So that is what the 990X is then, a remnant of the tech bubble. :hmm:

No. Money is out there for super machines. But its a server CPU first and foremost, if anyone is taking a risk on that product its the motherboard manufacturers building 1366 boards. People don't buy the CPU, then just starting marking the new ones as Xeons that would sell for more generally anyways.

But not so much the tech bubble, but of the GHz wars. It was at the edge where a faster processor actually meant something.

Same thing is going to start happening to cars now. As the big push for HP happened, getting the New R/T, SS, or GT meant something. But now you have family sedans hitting 250+ HP, you have the cheapest mustangs hitting 300 HP. All with a 70 mph speed limit, and laws limiting "drag racing". The new emission and fuel efficiency laws are probably the best things to happen to ICE engines. There will be teething issues, manufacturers like Chrysler will probably shelve their V8's for another decade or so, and there will be a decent drop off in performance numbers, and then they can develop their way back up.

That's what the core wars are now. But now AMD is in even less of position of chasing ASP. In this climate, decent margin mixed with good market penetration is more important then chasing that $1000 gaming CPU. Why because they can just sell it as a server CPU for more.

 

[podspi]

Oh? So that is what the 990X is then, a remnant of the tech bubble. :hmm:

Marketing. First on the race track, first in the show room.