AMD - Bulldozer Chitchat

[Biscuit]

Well if that article turns out to be bollocks, it goes a long way to discredit the website.

[scaryjim]

Indeed, if it's a plant by AMD then it's a massive one. And I can't imagine any website would be so dumb as to make up something like that.

But it doesn't actually shed any light on real world performance. If the top end bulldozer can get within - say - ten percent of a 990X but cost only half as much, then it still wouldn't be a world beater, but it'd be a damn fine chip. And the issue AMD seem to be having with SYSmark is that it's used by the US Government as a determining factor in large-volume contracts - which as we all know are inevitably filled by massively over-specced machines that cost twice as much as you need to spend to perform the required tasks (like, for instance, the C2D E8400 in the standard office machine as specced by my employer: massive overkill). So even if BD is a better value chip, the reliance on SYSMark to the tender process might bite them in the ass (or hoist them on their own pittard, if the claim that SYSmark is used as a procurement decider due to AMD lobbying is true), as Intel can simply throw down a slightly better performing chip, even if it's more expensive.

So, real benchmarks on real silicon. Then we'll know how good, or bad, bulldozer really is.

[Biscuit]

Indeed, the thing with me is i was hoping to buy the motherboard now, run on a temporary cheapo athlon X3/X4 (which i would then sell on/give to my dad to replace his aging 939 opty) thereby spreading the cost of my upgrade a little. I have just been waiting to have some fairly solid indication that BD is going to decent... this has pushed my eagerness to take the AMD plunge right back down.

Although its not a solid source of information, it certainly gives off a whiff of credibility given AMDs recent history.

[Terbinator]

If the top end bulldozer can get within - say - ten percent of a 990X but cost only half as much, then it still wouldn't be a world beater, but it'd be a damn fine chip.

The downisde being the 2500K is within about 10% of the 990X, and costs around £160.

Its all very exciting!

[scaryjim]

The downisde being the 2500K is within about 10% of the 990X ...

erm.... are you sure?! The i7 2600K doesn't get within 10% of a 980X...

@Biscuit: ah, yes, I see that in your position the rumours will have more resonance

[CAT-THE-FIFTH]

BSN has got it all wrong many times before and I suspect they just want more hits to their website. Even,the original Athlon 64 was delayed and no one really expected it to perform so well either.

If a Bulldozer core had the same IPC as a Phenom II core,AMD would have simply built a 32NM Phenom II and either clocked it higher or added another two cores.

Half the work has been done with Llano already so it Bulldozer was so bad they wouldn't be releasing it this year. Even,the original Phenom had a decent IPC improvement over the Athlon 64.

For example a Phenom II X6 with 50% more L3 cache probably would not be larger than Llano,ie,much smaller than a Phenom II X4. A 10% increase in clocks with more refined Turbo Core over three cores would probably mean 15% to 30% more performance overall.

How much reliable information was leaked about the HD4000,HD5000 and HD6000 series before launch?? Not much IIRC.

[Biscuit]

BSN has got it all wrong many times before and I suspect they just want more hits to their website. Even,the original Athlon 64 was delayed an no one really expected it to perform so well either.

If a Bulldozer core had the same performance as a Phenom II core,AMD would have simply built a 32NM Phenom II and either clocked it higher or added another two cores.

For example a Phenom II X6 with 50% more L3 cache probably would not be larger than Llano,ie,much smaller than a Phenom II X4. A 10% increase in clocks with more refined Turbo Core over three cores would probably mean 15% to 30% more performance overall.

Sure, but how would they explain to everyone why there is no Bulldozer? AMD are definitely committed to releasing a new architecture weather it is good or not.

[CAT-THE-FIFTH]

Sure, but how would they explain to everyone why there is no Bulldozer? AMD are definitely committed to releasing a new architecture weather it is good or not.

Re-edited my answer. I added some points.

AMD said Bulldozer would be released this year. It was only leaked roadmaps which said Q2 2011 for consumer versions. The same leaked roadmaps said Q3 for all versions of Llano.

If Cray has already publically committed itself to having a supercomputer with Bulldozer CPUs,it means it must be a decent improvement. Even the Phenom had a 20% IPC improvement over an Athlon 64.

On top of this Trinity is based on Bulldozer cores. If Bulldozer cores had the same IPC then they might as well have tweaked the Phenom II cores even further.

The only possible way that Bulldozer had similar IPC if it runs at insane standard clockspeeds,ie,well above 4GHZ. Since Bulldozer is launching in four and six core versions it would mean a Phenom II X4 and X6 are only slightly slower.

[CAT-THE-FIFTH]

Look at it this way. What if a Bulldozer quad core was running at 4GHZ stock speed. It would give it barely 8% better multi -threaded performance than a Phenom II X4 980. Even if Turbo Core ramped up to around 4.4GHZ it would give you 20% better perfomance in lightly threaded applications.

It sounds more like BSN has made up more BS based on information on B0 samples leaked on other sites. He still says B0 and B1 samples had major problems. Anandtech was the first site to mention purported B1 problems and later mentioned it was actually hitting the expected clockspeeds.

[DanceswithUnix]

Well if that article turns out to be bollocks, it goes a long way to discredit the website.

ISTR the website has had its dubious moments in the past, so I do suspect they are a good (read gullible) way to "leak" information if someone wanted to stir things up.

The thing is, I can't see what anyone would be trying to achieve by leaking this. In many ways it just doesn't add up.

[watercooled]

Sure, but how would they explain to everyone why there is no Bulldozer? AMD are definitely committed to releasing a new architecture weather it is good or not.

It was explained a while ago, it's due to there being a higher demand for Bobcat/Llano so Bulldozer production got pushed back. All the fabs were/are essentially maxed out.

[Terbinator]

erm.... are you sure?! The i7 2600K doesn't get within 10% of a 980X...

Look at the gaming results

[DanceswithUnix]

Look at it this way. What if a Bulldozer quad core was running at 4GHZ stock speed. It would give it barely 8% better multi -threaded performance than a Phenom II X4 980. Even if Turbo Core ramped up to around 4.4GHZ it would give you 20% better perfomance in lightly threaded applications.

It sounds more like BSN has made up more BS based on information on B0 samples leaked on other sites. He still says B0 and B1 samples had major problems. Anandtech was the first site to mention purported B1 problems and later withdrew the statement when they actually talked to a few people.

I got the impression that a lot of the benefit will come from the shared SSE units in the module design, so a single thread gets twice the SSE throughput of a Phenom by borrowing the SSE throughput from the other core on the module. Once you hit 4 threads then all the modules are using their SSE units and further threads have to share. I can see that working nicely in the real world, but some benchies might look very odd.

[DanceswithUnix]

It was explained a while ago, it's due to there being a higher demand for Bobcat/Llano so Bulldozer production got pushed back. All the fabs were/are essentially maxed out.

But I have my Bobcat and don't want a Llano, so I they have my permission to start making me a BD

Mind you the Zacate is a TSMC 40nm fabbed part so thats a poor excuse.

[Terbinator]

http://obrovsky.blogspot.com/

[CAT-THE-FIFTH]

I got the impression that a lot of the benefit will come from the shared SSE units in the module design, so a single thread gets twice the SSE throughput of a Phenom by borrowing the SSE throughput from the other core on the module. Once you hit 4 threads then all the modules are using their SSE units and further threads have to share. I can see that working nicely in the real world, but some benchies might look very odd.

This is why the article is BS. A module running two threads is going to have lower performance per thread than when running a single thread which JF-AMD has stated plenty of times. IIRC,adding a second module ends up with upto 70% to 80% improvement over two threads. They did this as two separate cores would be much larger.

So basically IPC should really be much better in lightly threaded applications. The BSN article is twisting things to sound like Bulldozer is a failure when probably it isn't.

http://obrovsky.blogspot.com/

Don't trust him. He has been banned from loads of forums.