AMD Zambezi news, info, fans !

[informal]

It doesn't really look that far off TBQH. If we check the difference between the 1100T and the 8150, it shows the 8150 as performing 33% faster. With the 1100T lagging slightly behind the reference i7 950, it's hard to draw a solid conclusion from that first pic anyways. We also don't know how accurate that second slide is either. The subtext indicates that the numbers are AMD's own official results (and intel's own on the 3960x), but from where and with what system configuration?

--Matt

Actually what we have now in the latest slides from AMD and what we saw from that first Scorpius slide is as different as night and day. Originally , the slide was portraying Zambezi as Thuban crushing machine,it was a massive 86% faster in legacy SSE code (C11.5) that supports 8 threads, a massive 50% faster in 3dmark 06 that is not so well threaded and a solid 15 to 20% faster in serial code such as PCmark TV and Movies benchmark. What we have now in all these previews and even "official" AMD slides that leaked is quite a bit slower Zambezi,a chip that barely can beat Thuban 6C in many workloads and that is on average just a bit faster than 1100T. It still has massive OC headroom though,but none of the advantages of new floating point unit is there any more. New FlexFP is now lagging behind one thuban core in many workloads,instead of outperforming it significantly. Something is not right. We can wait until launch for real benchmarks but probably we won't see much of a change .

[drfedja]

Old slides is probably based on CPU simulation, but in realworld scenario BD marchitecture behaves very different. Maybe some of advanced features are disabled because of serious bug, and there is performance penalty. I can't wait to see BD errata list. I can't believe in that the cache architecture is simple reason for low performance because they can simulate how much performance penalty came from smaller 16K L1D, and WT based cache policy. Hit rate of 4-way 16KB is only 1 or 2 % less than 64K 2-way. 2MB L2 has also much higher hit rate than 0.5MB L2. Maybe BD is better optimized for large data work sets.

[Oliverda]

AMD lied saying that BD is an 8 core. In fact even now is lying. It is not.

Can you tell me that a non 8 core processor how can do 6.72x speedup in a multithreaded task?

cb11.png

[Dresdenboy]

There is no performance difference between 0.9x and 1.x Agesa code at all.

There are indeed only small diffs in the disassembled AGESA code. Did you perform tests?

[informal]

Old slides is probably based on CPU simulation, but in realworld scenario BD marchitecture behaves very different. Maybe some of advanced features are disabled because of serious bug, and there is performance penalty. I can't wait to see BD errata list. I can't believe in that the cache architecture is simple reason for low performance because they can simulate how much performance penalty came from smaller 16K L1D, and WT based cache policy. Hit rate of 4-way 16KB is only 1 or 2 % less than 64K 2-way. 2MB L2 has also much higher hit rate than 0.5MB L2. Maybe BD is better optimized for large data work sets.

The slide is from December 2010. Bulldozer taped out in Q2 2010. The data is most likely NOT based on simulations but on real hardware. The clocks on the initial steppings could be low but they can extrapolate performance of higher clocked models easily. The point is that by Dec 2010,the design was well pass the finish point and they were doing the sampling to partners. The only problem they may hit by that time is low clock,low yield or leaky parts with power draw issues(or any combination of those). If anything was wrong(design bug)they would have known it and figured it in all the performance expectations and hence that scorpius slide would have it too. That is why I think there is no performance problem but something else is going on.

@ Oliverda

While the scaling is that of an 8 core chip, the single core performance is abysmal. Leaks point out to a single thread score of around 24% lower than one Thuban core at same clock. Either both fmacs are not utilized in single thread test or something else is the problem. 2x fmac should not be slower in legacy code than one Thuban core,especially since we know that fmacs are more flexible and can do any of the fp commands.

[drfedja]

Can you tell me that a non 8 core processor how can do 6.72x speedup in a multithreaded task?

cb11.png

BD modules scales very well, but there is problem with very very low per module IPC. It is ~30% lower than K10.5.

The slide is from December 2010. Bulldozer taped out in Q2 2010. The data is most likely NOT based on simulations but on real hardware. The clocks on the initial steppings could be low but they can extrapolate performance of higher clocked models easily. The point is that by Dec 2010,the design was well pass the finish point and they were doing the sampling to partners. The only problem they may hit by that time is low clock,low yield or leaky parts with power draw issues(or any combination of those). If anything was wrong(design bug)they would have known it and figured it in all the performance expectations and hence that scorpius slide would have it too. That is why I think there is no performance problem but something else is going on.

Maybe data from slides was came from earlier time than Q2 2010. Maybe R&D team was thinking that some mysterious bug can be solved before launch. Maybe they don't know what's wrong with BD....

[Dresdenboy]

chew* stated something about SuperPi, a single threaded program gaining a significant amount when set to more than one thread...we will see where this progresses upon release/NDA release in my opinion.

He wrote about running SPi on a module while allowing a 2nd thread to run on it and running SPi w/ only one core enabled on that module.

[liberato87]

Actually what we have now in the latest slides from AMD and what we saw from that first Scorpius slide is as different as night and day. Originally , the slide was portraying Zambezi as Thuban crushing machine,it was a massive 86% faster in legacy SSE code (C11.5) that supports 8 threads, a massive 50% faster in 3dmark 06 that is not so well threaded and a solid 15 to 20% faster in serial code such as PCmark TV and Movies benchmark. What we have now in all these previews and even "official" AMD slides that leaked is quite a bit slower Zambezi,a chip that barely can beat Thuban 6C in many workloads and that is on average just a bit faster than 1100T. It still has massive OC headroom though,but none of the advantages of new floating point unit is there any more. New FlexFP is now lagging behind one thuban core in many workloads,instead of outperforming it significantly. Something is not right. We can wait until launch for real benchmarks but probably we won't see much of a change .

that's exaclty what I'm saying.

In the 1st slide, below the graph we see

"estimates and projections subject to change "

but "change" cannot be 100% to 0%

[Dresdenboy]

BD modules scales very well, but there is problem with very very low per module IPC. It is ~30% lower than K10.5.

In CB it's - besides other factors - a 4 vs 6 FPU fight.

[liberato87]

In CB it's - besides other factors - a 4 vs 6 FPU fight.

I think amd engineers know that fpu are less, and the related consequences.
So why they made a slide showing zambezi with DOUBLE performance of thuban in cinebench 11.5 if it can be even worse than thuban (due to fpu and other factors)?

How did they made that slide? they were drunk? we are not talking about an error of 5-10%, but we're talking about performance that had to be doubled and instead are the same. I cannot think they make such a big fail in prediction.
or we have to think that the leaked test are fake?

the two things are like night and day, as informal said.

[drfedja]

In CB it's - besides other factors - a 4 vs 6 FPU fight.

I think that in CB code FPU is underutilised. Because of that, two integer blocks can push enough data to process with FPU and thread scaling is good. But I think, there is the problem with weak or narrow integer cores, maybe some serious L1D cache design flow.
4 BD FPU's doing job pretty well comparing to integer blocks. There is more scenarios where BD shows weak integer performance.

[flyck]

In CB it's - besides other factors - a 4 vs 6 FPU fight.

Well the other thing, and the most likely thing is that R11.5 isn't limited by fpu power at all for BD. Given the scaling on the BD leaks, adding another thread on the same module doesn't drop performance of either thread. If you look at intel HT, it doesn't reach the scaling AMD gets while the fpu is also SMT like. (running 2threads on them is hurting performance of each thread dramatically)

So performance of R11.5 isn't limited by the fpu resources but by something else on BD. Would also expleain why the leaked sisoft and other benchmarks the BD is ahead of the K8like architecture in fpu on a core and clock basis.


edit: i'm with drfedja on this.

[-Boris-]

you mean fx 8150p?
It would be a fail if the 8150p < 1100t.

anyone remember this (december 2010) : zambezi performs 50% better than 1100t.

and we can read above "render performance is based on Cinebench R11.5"

Attachment 120656

somoene lied..

Please, these slides were obvious fakes already back then! There where several things that was clearly wrong. I told people not to trust them because I knew people would start accusing AMD of lying. I know you just staded that someone lied, but know people are at it again. Look at the post below.

AMD lied saying that BD is an 8 core. In fact even now is lying. It is not.
Second, if BD is further delayed for an actual launch, on 12 octomber we should have at least a "paper launch".

It's an eightcore, it has 8 integer clusters and 8 FPUs, the only thing is that a pair of cores can borrow FPUs from each other and share some L1 cache and prefetch and scheduler.

[xdan]

It's an eightcore, it has 8 integer clusters and 8 FPUs, the only thing is that a pair of cores can borrow FPUs from each other and share some L1 cache and prefetch and scheduler.

It's an 4 core, 8 integer clusters simple and plain.
More like a HT thing.
Intel Core quad aproach( to put 2 dual cores together) was more "true" core design than BD is now.
8 cores, o yeah...

[FlanK3r]

you mean fx 8150p?
It would be a fail if the 8150p < 1100t.

anyone remember this (december 2010) : zambezi performs 50% better than 1100t.

and we can read above "render performance is based on Cinebench R11.5"

Attachment 120656

and now?
watching the slides of donaninhaber we see in cinebench that

Attachment 120657

somoene lied..

My meaning, there is some problem. Multiefectivity between cores/moudles is good, score is whorse than in original slide, so problem could be in module or in core unit-more lower performance. I dont know which exactly problem, but...hope, will be fixed in Piledriver.

[-Boris-]

Well the other thing, and the most likely thing is that R11.5 isn't limited by fpu power at all for BD. Given the scaling on the BD leaks, adding another thread on the same module doesn't drop performance of either thread. If you look at intel HT, it doesn't reach the scaling AMD gets while the fpu is also SMT like. (running 2threads on them is hurting performance of each thread dramatically)

So performance of R11.5 isn't limited by the fpu resources but by something else on BD. Would also expleain why the leaked sisoft and other benchmarks the BD is ahead of the K8like architecture in fpu on a core and clock basis.


edit: i'm with drfedja on this.

Or it is FPU limited but the processor fails to share FPU power between cores. So that all cores just got their own FPU and can't successfully borrow from the others.

[-Boris-]

It's an 4 core, 8 integer clusters simple and plain.
More like a HT thing.
Intel Core quad aproach( to put 2 dual cores together) was more "true" core design than BD is now.
8 cores, o yeah...

Now your make your own definitions, I can make the definition that CPUs are turds, and accuse Intel of lying then they're saying that their CPUs isn't turds. But that doesn't mean I'm right or that Intel actually is lying. You can only accuse AMD of lying in this matter if you make up your own definitions of what the truth is.
According to the generally accepted definition BD is eight core. You can keep your definitions.
But out of curiosity. What defines a core to you?

[informal]

One thing we know for sure now. 32nm production is not where AMD wanted it to be. Llano as a high volume product is eating a lot of capacity and it is yielding much under desired level. On top of that the production is shared between current 45nm and 32nm product lines . This is further more complicating the situation. Bulldozer is large die product and with all factors combined it's no wonder why AMD pushed it back to Q4. As for results maybe none of the results are true since AMD doesn't want to generate unnecessary hype around a product that it can't deliver in large quantities. They would devalue their current product lines and would be unable to supply enough of new products (both llano and server/desktop Bulldozer).

[xdan]

It was said before by, Chew, who is a better expert.


Read also that PDF http://www.freepatentsonline.com/20080209173.pdf See that pdf figure 3 what is a true core...A true core doesn't share L1 cache, FPU, prefetch with other core...
A core could have 2 clusters which share between them all this and work in 2 threads....
Anyway AMD 4C/8treads cpu beats INTEL 2500K quad would have sound better than...
AMD 8 core beats Intel 2500K quad...

[Oliverda]

It was said before by, Chew, who is a better expert.

Read also that PDF http://www.freepatentsonline.com/20080209173.pdf See that pdf figure 3 what is a true core...

You still haven't answered my question.

[Manicdan]

AMD lied saying that BD is an 8 core. In fact even now is lying. It is not.

the whole core argument started up again right here when you said this

fact is most of us dont care at all what AMD calls it, or what you call it, or what even i call it. most posts i say module and thread now just to keep people happy and not let them derail a thread with this marketing nonsense every day. since thread scaling gets better by having L3, does that mean the latest Athlons are no real quads? or does that mean Phenom is just somehow a better quad. and if so then maybe its possible to consider BD a octo core, but just not a very good one. we can skin this cat with more posts than this thread is, but in the end our answer will not make a single difference on the performance.

[muziqaz]

Did I just step into time machine and it send me few months back? How many times did we have that discussion about cores :/ getting really boring

[spicypixel]

This thread will be cyclical until AMD releases it.

[Smartidiot89]

4 cores with 8 threads, or 8 cores with 8 threads. Who cares, it's not like it will change the price, performance or power consumption depending on what it's called.

[liberato87]

My meaning, there is some problem. Multiefectivity between cores/moudles is good, score is whorse than in original slide, so problem could be in module or in core unit-more lower performance. I dont know which exactly problem, but...hope, will be fixed in Piledriver.

that's what I am thinking too.
there is some problem or someone lied (amd or the slides are fakes).

but.. BD was for q2, then step to "60-90 days" from 1st june.. now from q3 to q4.
I think amd knows the problems, if there are problems, and they have not found these problems yesterday..
So I hope that the problems, that are big if we have less performance than thuban, had been solved in this 6 months of delay.

If the performance are these, even if only cinebench (ok, this is only one bench but... it is deep under the amd estimated performance) is worse than thuban, IMHO amd should not release bd, even if this is price competitive, because every review will kick it because, the delays, the silence (last official statement "60-90 days from 1st june", then from q3 to q4 without ONE official word) the hype created by the "return of the fx comics" and the world record, will affect against amd.

And we are talking about the "top" 8 core cpu..
I cannot think the fx 4xxx should perform less than a phenom II 955..

But I remember that movieman talked about bulldozer like a winner.. I hope he's right.