Why did AMD release bulldozer if they can still make good CPU's?

[NostaSeronx]

Why did AMD release the underperforming Bulldozer CPU's?

It wasn't underwhelming, it was just very delayed.
* Sandtiger: The code name for a family of server chips with eight to 16 cores that will be released in 2009. Will be produced using a 45nm process and include support for DDR3 (Double Data Rate 3) memory. Due in 2009.
* Falcon: Code name for the first Fusion chip that will combine a CPU and graphics processor. Designed for laptops, Falcon will offer up to four Bulldozer cores. Due to ship in 2009.
* Eagle: Code name for an upcoming notebook chip package based on the Falcon processor. To ship in 2009.

K9 (Netburst/Tejas competiton) => K10 (Core/Core 2/Nehalem competition), take that the 32nm was just a shrink with very little optimization. Well, the design was still targeting Core/Nehalem.

2.1 GHz Bulldozer = 2.66 GHz Nehalem, so it was clearly overwhelming, but by the time it launched Intel was at Sandy Bridge. With all the bottlenecks that Nehalem suffered being fixed with addition of 256-bit ops.

As AMD went through Steamroller/Excavator, they were leaving Minimal Multithreading to Tree Multithreading. Which if continued would lead to the FPU being duplicated across the internal cores. If given a good enough shrink, via 20LPM/14XM/22FDX.

 

[scannall]

Also a very pivotal time overall in computing history. AMD64 probably saved us all from being forced into an Intel only instruction set.

BTW quote is broken for me so have to do it manually.

It was an interesting time. I understand what Intel was trying to do with Itanium, and it wasn't a bad idea overall. At least for Intel. IBM basically anointed Intel and Microsoft at the start of the whole PC thing. And that's an interesting history in and of itself. Part of that was IBM required more than one source, hence the licensing with Via, AMD, Cyrix and others. Motorola had a *bit* more performance from their CPU's at the time, but was unwilling to license anything. Hence Intel got the nod. CP/M was IBM's first choice, but couldn't reach an agreement with Gary Kildall. And when he missed a meeting for negotiations IBM said 'screw it', and contracted a garage company called Microsoft to make what was essentially a clone of CP/M.

So this pair of 'second choices' as it were, have pretty much run the industry since about 1990.

What a lot of young'uns may not understand is at the time, if IBM called you damned well better answer the phone. With your lips puckered. Even now, if IBM calls it needs to get your attention if not your focus.

 

[A///]

When his new CPU from Apple comes out expect it to blow up the industry.

What? I thought he left Apple in 2011?

 

[Topweasel]

What? I thought he left Apple in 2011?

You know I thought he left to go back to Apple to work on their next major CPU probably to be that bridge CPU were they could use an Arm like system for MacOS (which I have always felt was their eventual goal). But I miss remembered. He went to Tesla.

 

[A///]

Why did Intel make Netburst when they can make a good CPU? Why did Intel try and come out with Larrabee? Because companies are not perfect, people are not perfect.

Not that I claim to be an expert in uarch, but I'd imagine sunken costs were one factor and they truly believed they would win in the end. I remember the Intel guys thinking they'd hit 5 Ghz with Pentium 5 or what would be Prescott in '04. They thought they could easily hit 10 Ghz in a few years, but that wasn't the case. In the P4 days, the later models would generate so much heat they could effectively warm up a small room. I'd like to believe they began Core uarch around that time or a year earlier.

 

[A///]

You know I thought he left to go back to Apple to work on their next major CPU probably to be that bridge CPU were they could use an Arm like system for MacOS (which I have always felt was their eventual goal). But I miss remembered. He went to Tesla.

Oh that. I heard Apple were doing or planning something like that. It's a rumor from what I know, but funny enough it came out around the time of AMD's news leaks back about six months or more ago, in regard to an ARM processor embedded in a traditional CPU. It might be true but who knows. I can see Apple getting into bed with AMD to make something amazing and at a lower per unit price compared to Intel's offerings.

 

[Insert_Nickname]

Now they had a lot of good ideas and some of those are in Zen and are responsible for getting Zen to clock as high as it does, so Bulldozer isn't a complete failure.

Undoubtedly. Don't forget they also managed to turn Bulldozer into something semi-decent with Excavator. On the back-burner while focused on Zen no less.

Now if XV had been the starting point for the Bulldozer architecture, who knows how it would have played out...?

 

[CentroX]

Jim Keller will be back with AMD again when they design their next architecture from scratch in 2021.

 

[cytg111]

Althon's were good (or decent) and Ryzen appears to be solid so why did AMD release the underperforming Bulldozer CPU's? Was it a question of too much money sunk into the project before they realized it was a failure or was it one person at the top not willing to change the path or perhaps at the time they did not have a better solution?

Inquiring minds want to know.

I guess you need a massive crystal ball when you are laying down your bets on a new design.. Will process tech achieve the projected characteristics, will software evolve to make use of features(more cores, isa extentions), will we have perfect execution, no delays other mishaps.

 

[cytg111]

Jim Keller will be back with AMD again when they design their next architecture from scratch in 2021.

I have a hard time seing AMD going all in on a new design so soon. We might see Ryzen II by then though.

 

[tential]

I have a hard time seing AMD going all in on a new design so soon. We might see Ryzen II by then though.

Ryzen 2 is coming out shortly after Ryzen. Shouldn't take them 4 years.

 

[CentroX]

Ryzen 2 is coming out shortly after Ryzen. Shouldn't take them 4 years.

actually I think it goes like this:

2018: Ryzen+ on 14nm
2019/2020: Ryzen 2 on 7nm.

 

[Greyguy1948]

Althon's were good (or decent) and Ryzen appears to be solid so why did AMD release the underperforming Bulldozer CPU's? Was it a question of too much money sunk into the project before they realized it was a failure or was it one person at the top not willing to change the path or perhaps at the time they did not have a better solution?

Inquiring minds want to know.

Most likely because of bad simulation. The L1 cache was very small and both L2 and L3 were too slow. Double size of L1 i and L1 d like you have in Excavator is not bad but they never tested L3. The problem with Phenom was the high energy. In fact the fp is not bad in Bulldozer regarding it is 4C/8T. They have to learn much about high speed of cache. Maybe they have it in Ryzen

 

[Markfw]

Also a very pivotal time overall in computing history. AMD64 probably saved us all from being forced into an Intel only instruction set.

BTW quote is broken for me so have to do it manually.

Use reply. Its quotes only the post you "reply" on, but it works.

 

[Greyguy1948]

actually I think it goes like this:

2018: Ryzen+ on 14nm
2019/2010: Ryzen 2 on 7nm.

2019/2020 is very optimistic. AMD is very new on 14 nm. They have much work on 14 nm for both CPU and GPU.

 

[CentroX]

2019/2020 is very optimistic. AMD is very new on 14 nm. They have much work on 14 nm for both CPU and GPU.

GloFlo is starting 7nm production in 2018. Should be ready 2020.

 

[Dresdenboy]

actually I think it goes like this:

2018: Ryzen+ on 14nm
2019/2010: Ryzen 2 on 7nm.

Timing sounds plausible. But I think they either call the cores Zen+, Zen++, or Zen 2, Zen 3, not mixing this.

SKUs likely will be Ryzen 2xxx, 3xxx, etc. And the designs might further be called based on mountain related things, were in the past some Zen related history happened.

 

[CentroX]

Timing sounds plausible. But I think they either call the cores Zen+, Zen++, or Zen 2, Zen 3, not mixing this.

SKUs likely will be Ryzen 2xxx, 3xxx, etc. And the designs might further be called based on mountain related things, were in the past some Zen related history happened.

Yeah it makes sense. I fully expect Zen+ to be 2000X and above. Maybe AMD will release 1900X this autumn under the current design.

 

[Cali3350]

By the time AMD realized what a dud Bulldozer was going to be, they had too much sunk cost to not release. Not releasing would have likely put them out of business. They also believed the design was somewhat salvageable, but that never really came to fruition. I think the Dozer lines would have been different if AMD had a better fab (GloFo REALLY let them down), but in the end the design just didnt work as well as AMD hoped.

But Zen is brand new. It doesnt really share anything with the core Bulldozer design. AMD has lots of smart people working for them. They missed once, but they didnt miss twice

 

[french toast]

If they didn't release Bulldozer what would they sell? 5 year old Athlons?

They would have been better improving phenom 2, imagine how many hundreds of millions were spent on bulldozer and derivatives?
Imagine what phenom could have been had they spent half of that dev cost on phenom 3, SMT, better caches, wider core, new process etc etc.

Would have been better imo.

 

[cytg111]

GloFlo is starting 7nm production in 2018. Should be ready 2020.

Some day I like to hear the story about how they did NOT mess 14nm up... this was the biggest concern in my book.

 

[MajinCry]

They would have been better improving phenom 2, imagine how many hundreds of millions were spent on bulldozer and derivatives?
Imagine what phenom could have been had they spent half of that dev cost on phenom 3, SMT, better caches, wider core, new process etc etc.

Would have been better imo.

I could be wrong, but wasn't Phenom II a one hit wonder? In order to improve it further, the whole core would have had to been rebuilt, essentially creating a new architecture rather than a successor.

actually I think it goes like this:

2018: Ryzen+ on 14nm
2019/2020: Ryzen 2 on 7nm.

Ya. That's what the "Ryzen will be tock tock tock" thing referred to. We're going to get a few revisions of Zen. Intel's way of doing "tock tock tock" would be Sandybridge, Ivybridge, Daveybridge, Haswell.

 

[AnandThenMan]

Use reply. Its quotes only the post you "reply" on, but it works.

Heh, works thanks. Don't know why I didn't try that.

 

[StinkyPinky]

I thought 7nm is not possible on silicon? Are AMD using something else for 7nm?

 

[french toast]

I could be wrong, but wasn't Phenom II a one hit wonder? In order to improve it further, the whole core would have had to been rebuilt, essentially creating a new architecture rather than a successor

Yes perhaps, but they would have been better off doing that then the flawed idea that was bulldozer, the concept was interesting but severely flawed, as ST performance was about equal to phenom 2, the module was massive vs intels HT core, so much so they marketed it as 1 module = 2cores, when in originality CMT was supposed to be a direct rival to HT (1 module =1 intel core), just more efficient.

They should have spent all that money on a more traditional core design or redesigned phenom with SMT, hell i thought they were going to do that with Barcelona.