Why BD failed? AMD Ex-Employee speaks out!

[Frick]

OMG what a f* disaster. :shadedshu I've just seen a block diagram and explanation for the Bulldozer architecture: it's not an 8-core chip at all. The two 'cores' are siamesed into one building block and share lots of resources and lose lots of performance. This thing is actually merely a quad core processor with a souped-up version of HyperThreading or whatever the AMD equivalent is.

Here's the block diagram taken from Hexus' review:

http://img.techpowerup.org/111013/BDS.jpg

See what I mean? The only thing it seems to be good at is overclocking and that only brings it within reach of Intel's top parts at stock, it doesn't even beat them, all the while using a huge amount of power and requiring expensive cooling. This chip is a lemon like Vista was for Microsoft, or perhaps even worse.

I'll soon be upgrading my aging E8500 to an i5-2500K. No point in waiting any more.

8 integer cores = HT? I'm not that knowledgeble about this, but that is a simplification.

 

[qubit]

Seriously you've only just realised this?

It's literally the main selling point for bulldozer.

It means instead of building a whole entire core they use 25% of one and still get MOST of the performance of an extra core. ( In theory a shit ton better than hyper threading )

To put it simply, AMD may of had an underwelming release but they've just started a core race with Intel ( I imagine it's due to silicone not going to go much further IPC and clock speed wise) Piledriver will be 10 cores and the chips after that will be 20 core monsters. ( I suspect AMD are going to try and push OPEN CL big time in order for the cores to be taken advantage of easily)

They've a plan, they may not pull it off but if they do they'll become server CPU kings and if they can increase IPC a wee bit to bring it into line with competing processors then it will be a beast.



I'm aware it doesn't seem like they're pulling that off at the moment but I'm fairly certain something is wrong either with software or hardware ( rather than design) as a chip with twice the transistors of phenom x6 should be smashing it.

I've only just looked at the architecture in any detail, yes. Taken together with the benches, it's not hard to see why performance falls short. Also, the memory bandwidth is not especially high in the benches, which leads to bottlenecks. The Intel chips have higher bandwidth with less cores, so they're gonna be able to breath better.

This is definitely a 4 core chip with two threads per core, not a true 8 core chip, end of story.

I think having AMD say that the software isn't optimized for it is a bit of an excuse. Hexus said it was a dangerous argument to base a purchasing decision on, too.

I definitely don't think AMD should be marketing this as an 8 core chip.

8 integer cores = HT? I'm not that knowledgeble about this, but that is a simplification.

http://images.anandtech.com/doci/4955/BDArch.png

Yes, the integer units are separate, but most of the core is all shared resources, so it's just one core that's able to handle two threads. Just look at a block diagram for something like a Phenom or my E8500. You'll see that everything is separate, apart from the cache and glue logic.

 

[pantherx12]

I definitely don't think AMD should be marketing this as an 8 core chip.

That's what I thought when I first read about the architecure Knew it would bite them in the arse as people would expect you know, the whole performance of a core


From a definition POV though it does count as a "core" as it's actually scheduling and excuting two threads at the same time.

( hyper threading is more like smart HW based thread scheduling from my understanding, correct me if I'm wrong folks!)



What I'd like to know is, what was the BD release supposed to coincide with?

Generally AMD try and get their launch dates close a launch of something else so help boost sales.

So what was it? lol

 

[HalfAHertz]

the whole idea behind bulldozer was to decrease the size of a core from X transistors to 50-75% of X while still keeping a similar level of performance. The bad thing is that at the last moment someone decided to add all these other "useful" things like a humongous L2 cache, AES hardware decription and new instruction sets and the whole thing quickly grew out of proportions.

See once again AMD tried to do too much at once and ended up with a botch job just like with the original phenom. Be aware of your abilities and don't overstate them - that simple :/

$ntel found out long ago that a smaller but faster cache beats a bloated large cache every time yet if you look at the picture above roughly 1/3 of the die is L2 cache - it's almost as big as the module itself :/

AMD are trying to blame Glo Fo but imagine how much leaner the original phenom could have been if they had decided to keep the L3 for the next process node or if BD had "only" 512kb L2 per module and didn't require a 2 billion die...

 

[Nick89]

If the price on the 8150 drops to 200$ or less it will be my next processor.

 

[Mindweaver]

OMG what a f* disaster. :shadedshu I've just seen a block diagram and explanation for the Bulldozer architecture: it's not an 8-core chip at all. The two 'cores' are siamesed into one building block and share lots of resources and lose lots of performance. This thing is actually merely a quad core processor with a souped-up version of HyperThreading or whatever the AMD equivalent is.

Here's the block diagram taken from Hexus' review:

http://img.techpowerup.org/111013/BDS.jpg

See what I mean? The only thing it seems to be good at is overclocking and that only brings it within reach of Intel's top parts at stock, it doesn't even beat them, all the while using a huge amount of power and requiring expensive cooling. This chip is a lemon like Vista was for Microsoft, or perhaps even worse.


I'll soon be upgrading my aging E8500 to an i5-2500K. No point in waiting any more.

Hey NEWS guy.. Each Bulldozer core has 2 threads... This has been know for quite some time now... I don't see what all the fuss is over? Nobody in this thread has a BD.. Wait till the actual chip gets released, before passing judgment.. Use common sense.. if the best BD Stock chip only performs on par with a 2500k... Then they would not set the price higher than a stock 2500k. There is a lot more people that buy pc's and leave them stock then OCing them. Those people will buy a faster AMD pc.. What ground will amd lose by staying where people expect them to be? None... Does everyone in here think that the few OCer's in these forums keep AMD alive? Well no.. they do not pay there bills..

 

[Benetanegia]

Seriously you've only just realised this?

It's literally the main selling point for bulldozer.

It means instead of building a whole entire core they use 25% of one and still get MOST of the performance of an extra core. ( In theory a shit ton better than hyper threading )

It was obvious for me from the start, when I first saw the diagrams, that all those claims were complete bull. Or in other words it was taking the whole thing with 100% optimism and wishful thinking. The key of the matter is that each module has a single scheduler capable of issuing 4 instructions. The fact that the entire module can execute 8 instructions (4 per "core"), which is what AMD marketing department has been selling all the time is almost completely irrelevant. 99% of the time Zambezi is capable of only 4 instructions per module, that is 4 x 4 = 16 instructions for the entire chip. For comparison Thuban is capable of issuing and executing 3 instructions per core, so 3 x 6 = 18. Of course this only accounts for the maximum capacity and in a sense this is where BD was going to excell. If the software only requires 2 instructions, then BD can still execute 16 (2 per core) and Thuban is left with only 12. This was the logic behind the BD concept and this is what any kind of multi-threading tries to maximize the use of the resources. After all has been said and done and benchmarks have been performed, the bad thing for AMD is that Intel's SMT more than suffices for this task and has proven to be a lot more efficient.

And come on AMD's claims about 25% more area for 80% more performance was complete bull. I mean 2 billion transistors for 4 modules (8 non real cores), when Thuban is only 900 million or 750 million for Deneb. How is that a 25% increase at all? They could have made a real 8, hell probably even a 12 core Thuban-based chip on the same transistor budget and it would have been 100x times better, even if core scaling kind of sucked in comparison.

 

[Frick]

This is definitely a 4 core chip with two threads per core, not a true 8 core chip, end of story.

This has been the talk for quite some time now, so yeah.

And remember that it only fails in single threaded performance. When increasing load there it is on par with SB.

 

[pantherx12]

Could W1zz or someone else try and kept a statement from AMD asking them what's up?

Or have AMD just pulled of a GTX480 and packed to much into one die?

I'm still just shocked something with twice the transistor count can do so badly.( well not badly, I'd say performance was acceptable)

 

[naoan]

They could have made a real 8, hell probably even a 12 core Thuban-based chip on the same transistor budget and it would have been 100x times better, even if core scaling kind of sucked in comparison.

I wish they would just do this

 

[Therion_I]

There are some early suggestions that it could be a bug with BD working with Asus Crosshair V and GTX580 which most benchamarks have been run with. See this thread on Toms:

http://www.tomshardware.co.uk/forum/315775-10-asus-crosshair-giving-biased-results-bulldozer

To sum up, benchmarks run on ASRock 990FX Extreme4 with a Sapphire Radeon HD 6950 2GB Toxic Edition and 2x 4GB Corsair Vengeance DDR3-1866 seem much closer to i7 2600k.

Not sure what to believe but could the poor performance really be just a bios bug...?

 

[xenocide]

There are some early suggestions that it could be a bug with BD working with Asus Crosshair V and GTX580 which most benchamarks have been run with. See this thread on Toms:

http://www.tomshardware.co.uk/forum/315775-10-asus-crosshair-giving-biased-results-bulldozer

To sum up, benchmarks run on ASRock 990FX Extreme4 with a Sapphire Radeon HD 6950 2GB Toxic Edition and 2x 4GB Corsair Vengeance DDR3-1866 seem much closer to i7 2600k.

Not sure what to believe but could the poor performance really be just a bios bug...?

I would imagine since they are using a weaker GPU, the bottleneck is there instead of at the CPU which would result in a more "level" outcome.

 

[Tank]

despite all the issues i will still buy amd. just bought another phenom today to get my pc up and running again

will buy a 8150 later on once my finances recover after i get a job again.

regardless in high end games bd's performance is acceptable to me

 

[nt300]

If the price on the 8150 drops to 200$ or less it will be my next processor.

Me too. At $200 it's a deal and a half. Use the included AMD software to optimise it for further gain.

 

[AphexDreamer]

Yeah if Benchmarks show that BD is an improvement orver my Phenom II 965 I'll be upgrading to it as well.

Maybe I'll wait for a price drop.

Perhaps this is why the Chip has been delayed? I mean if they are so crappy you'd think they would have released before all this trash got washed up on the beach.

 

[bucketface]

This thing is actually merely a quad core processor with a souped-up version of HyperThreading or whatever the AMD equivalent is.

wait, people still didn't know this??
In Anand's 2010 analysis of the bulldozer arch, thats pretty much what he said.

I'm aware it doesn't seem like they're pulling that off at the moment but I'm fairly certain something is wrong either with software or hardware ( rather than design) as a chip with twice the transistors of phenom x6 should be smashing it.

half the bloody BD chip is cache... they really needed to shave at least half the l2 cache, at least. also the cache has really, really high latencies, something in the order of 2x or 3x intels cache.

 

[nt300]

What a bios bug? Well I know for a fact ASUS is working day & night on a new bios update to resolve some known Bulldozer issues via it's AM3+ mobo's. I wonder if this will help BD perform better....

 

[H82LUZ73]

I think we are watching a CPU manufacture turn into a graphics card one...Just my own opinion.Next year they will push the A FM2 sockets to us with built in 7000 gpu`s just to laugh at us further,AMD as a cpu company died 5 years ago after the x2 64 FX chips.Why do you think they put ATI logo out the backdoor and killed it so fast.And if the 10 core -16 core cpus desktop refresh of BD are good they will kill AM3+ mobos and make the FM2 a multi purpose to have A and Fx in it.

 

[_JP_]

There are some early suggestions that it could be a bug with BD working with Asus Crosshair V and GTX580 which most benchamarks have been run with. See this thread on Toms:

http://www.tomshardware.co.uk/forum/...ults-bulldozer

To sum up, benchmarks run on ASRock 990FX Extreme4 with a Sapphire Radeon HD 6950 2GB Toxic Edition and 2x 4GB Corsair Vengeance DDR3-1866 seem much closer to i7 2600k.

Not sure what to believe but could the poor performance really be just a bios bug...?

I know that some sites are preforming reviews with 1600MHz RAM, instead of 1866MHz, which seems to be the ideal RAM speed for BD. Also, RAM timmings always contributed to improve the performance results with AMD processors, so this one should be no different.
As for board specific issues, didn't notice it.
For once, though, I think the usual PR crap about pairing and AMD CPU with an AMD GPU actually means something, but I'll have to look further into that before having any conclusions.

 

[Nesters]

Where's seronx?

 

[heky]

I know that some sites are preforming reviews with 1600MHz RAM, instead of 1866MHz, which seems to be the ideal RAM speed for BD. Also, RAM timmings always contributed to improve the performance results with AMD processors, so this one should be no different.

Really, and the Intel SB chips were tested with 1333 ram and they still beat faildozer. Stop looking for excuses and admit to yourself the chip was over-hyped and fails to deliver.

 

[de.das.dude]

Read this on OCN. Main source: Click

amd needs a new head. amds policies and strategies sucks ass.

hint hint, i am unemployed talented and i am an INDIAN.

 

[Wrigleyvillain]

I think we are watching a CPU manufacture turn into a graphics card one...

Well, hey, the graphics cards are still quite up to par, at least, time after time at the same price point with their primary competitor, now aren't they?

 

[_JP_]

Really, and the Intel SB chips were tested with 1333 ram and they still beat faildozer. Stop looking for excuses and admit to yourself the chip was over-hyped and fails to deliver.

What a productive comment.
Uhm...if you had read any of my posts/replies, you would know I never said the chip was not over-hyped.
It's only obvious that, if bulldozer is optimized for 1866MHz RAM and has mediocre results with 1600MHz, with 1333MHz sticks it would be even worse.
Also, it's no big news that Intel's memory controllers are very good and AMD has yet to come up with something to counter that.
Bulldozer is not FAIL, it just needs some tweaks. Maybe this generation of CPU will not be able to do something remarkable, but there's always piledriver.

 

[3volvedcombat]

What you call the main source doesn't say anything more than what you posted though. And that site, Insideris.com references a thread in the Macrumors forum from 2010 concerning Apple possibly switching to AMD...
The forum member who made those comments seems to be rather bitter, here's a message from right before the first part of the snippet above:



That was April 2010. Actually, he didn't state facts, only claims, since we don't have any way to verify without AMD's permission to do so... but there must be some truth in some of them, outside of the obvious extreme bitterness about the design path and methodology taken.

Its true, because basically what he said is coming true, how is it not?

8 cores at 3.6Ghz is barley competing with the same architecture revision's of previous generations? Its a simple fact that they started doing what this ex amd guy said, about bulldozer specifically. They might keep it that way to.

There is no more to be said, bulldozer has bigger number's on paper then anything around it, but it's got low efficiency, getting low number's, and not performing like marketed. You couldnt get any more suck out of that?

Unless you hear about these FX core's failing because of ET migration within 2 years or less. He also stated that no?