AMD Zambezi news, info, fans !

[XRL8]

It's not hard to see that Bulldozer is no K10. It'd be nuts to think there aren't some decent measurable IPC gains.
So no, I don't believe OBR.
Or at the very least, if OBR does somehow have ES chips, they're not performing like retail chips are supposed to.

The fact that he has ES chips is not debatable ;-) .As for the Core per Core IPC ,well i dont know, remember thats before Turbo, were talking Mhz per Mhz here.And Cores in Bd arent Full cores on their own.

Ok this sounds really weird...why would AMD design a brand new cpu which doesn't perform much better than they have now. I really don't see that happening...

Even if IPC core per core would be the same,you still have 33% more of them for multithreaded work.You have new instructions for specialised code (AES SSE4.1 4.2 etc) which will give biig boost for some parts of the code.And on top of that you have big turbo for single threaded works.All that at lower manufacturing costs than thubans.

I think there is no point to jump to conclusions before we see test results. If he meant that 8 cores (4 modules) of Bulldozer perform in FPU-intense tasks 33% better than 6 Phenom cores (or one Bulldozer module is as fast as 2 Phenom cores), than it is big success for AMD.

Well ,i had hoped for more than that.But yea, all in all it wouldnt be bad chip still.

repman: relax....this "bad" sample beating sb 2600k in rendering easily. Obr showed it a few weeks ago at blog. So, Sandy Bridge performance now sounds weird too? ,-)

Taking best case scenario for BD tells us nothing about its general performance ;-)

[Apokalipse]

The fact that he has ES chips is not debatable ;-)

Yes it is.

As for the Core per Core IPC ,well i dont know, remember thats before Turbo, were talking Mhz per Mhz here.And Cores in Bd arent Full cores on their own.

If you take a look at the architecture itself, it's not hard to see that BD's IPC will not merely match K10.

As far as core scaling goes, when both BD cores in each module are being used, AMD themselves said multi-threaded performance would increase by 80% (remember 80% more than 100% is 180%)
And that's not compared to K10. That's compared to doing one thread on one BD core.

AMD also said they got "50% more performance from 33% more cores"
Of course they didn't say what type of workload. But in that context I'd assume it means something that uses >=8 threads (still not THAT specific).

If I combine their two statements:
50% more performance from 33% more cores
When both cores in each module are used, you get 80% more perormance (180% OF single threaded performance)
and assuming perfect scaling when all four modules are active, I'd conclude single threaded performance to be
[1/(1.80/2)]*(1.5/1.33)~=1.253 (are these calculations right?)
i.e. single threaded performance is about 25% better in the particular type of workload AMD tested with... but becomes 90% as much as that when both cores in each module are used (unless there's some workload where sharing the cache can result in a performance benefit)
0.9*125%=112.5%

Although yes, that's taking AMD's statements at face value.

[XRL8]

Yes it is.
If you take a look at the architecture itself, it's not hard to see that BD's IPC will not merely match K10.

As far as core scaling goes, when both BD cores in each module are being used, AMD themselves said multi-threaded performance would increase by 80% (remember 80% more than 100% is 180%)
And that's not compared to K10. That's compared to doing one thread on one BD core.

Well yeah, it is debatable in that way that if you wont check it yourself you wont know for sure.However the guy is known for having its connections and getting ES stuff, while he is disliked by most guys here, no one "in the know" dismissed tha fact that he gets samples.They are even irritated because he breaks NDA`s and still gets samples.
Guy has pictures of samples, guy has sold one , he provided many screenshots.What more do you want ?

As for the scaling, as you said, its not relative performance to phenom II but to BD itself.So it doesnt give us any IPC indication, and besides that as ive said before, new instructions and turbo should give big boosts.We should know more in the weekend.


@repman, you also have 2,3,4 core phenoms now so i dont get your point.It still would be big move upwards ,4-6-8 higher clocked parts with new instructions cheaper to build instead of 2,3,4,6 slower bigger chips.Im not saying thats the case, im saying it would still make sense to produce them instead of phenoms.

[Apokalipse]

Well yeah, it is debatable in that way that if you wont check it yourself you wont know for sure.However the guy is known for having its connections and getting ES stuff, while he is disliked by most guys here, no one "in the know" dismissed tha fact that he gets samples.They are even irritated because he breaks NDA`s and still gets samples.
Guy has pictures of samples, guy has sold one , he provided many screenshots.What more do you want ?

I'd want his statements about BD's performance to be more in line with what everything else hints.
In other words, measurably higher IPC than K10.5 (although exactly what IPC is compared to say SB is difficult to infer)

wha Gigabyte Thailand reply me "900 series mobo may start selling in August 2011"

But some Gigabyte 900 series boards are already selling....
I even have a 990FXA-UD7 in my PC right now

[freeloader]

If the results he shows this weekend are with shipping silicon and they're really crappy, what will people do? Still buy a crappy CPU or bail ship and switch camps? Just asking...

[XRL8]

If the results he shows this weekend are with shipping silicon and they're really crappy, what will people do? Still buy a crappy CPU or bail ship and switch camps? Just asking...

Well the choice would be really crappy then.
Am3+ ,subpar performance in singlethreaded and short lived platform.
1155 not cheap castrated platform with crappy overclocking options
2011 ,good solid platform that in order to own i would have to sell my kidney.

[repman]

@repman, you also have 2,3,4 core phenoms now so i dont get your point.It still would be big move upwards ,4-6-8 higher clocked parts with new instructions cheaper to build instead of 2,3,4,6 slower bigger chips.Im not saying thats the case, im saying it would still make sense to produce them instead of phenoms.

What I'm trying to say is that OBR has said that core per core BD has same performance as Phenom II (if I understood that correctly), you compared a 4 module BD (8 cores) with a 6 core Phenom II and saying it has 33% more cores (If i read it correctly) but if you would compare a 6 core BD to a 6 core Phenom II (and according to OBR) they would have the same performance, since core per core performance is the same.

(sorry if I read anything incorrectly)

I think single thread performance needs to get better ( a lot better than Phenom II) for BD to compete on the market.

[AbortRetryFail?]

What I'm trying to say is that OBR has said that core per core BD has same performance as Phenom II (if I understood that correctly), you compared a 4 module BD (8 cores) with a 6 core Phenom II and saying it has 33% more cores (If i read it correctly) but if you would compare a 6 core BD to a 6 core Phenom II (and according to OBR) they would have the same performance, since core per core performance is the same.

(sorry if I read anything incorrectly)

I think single thread performance needs to get better ( a lot better than Phenom II) for BD to compete on the market.

I agree but the first spins of this new arch will not really show where BD will be in 12 (or 24 for that matter) months.

Turbo'X' and additional SIMD compatibility will boost single-thread performance, but that may not reveal the underlying strength of the initial core/module design.

That comes with the Trinity APU 'SIMD Engine Array' in 2012, and further logic enhancements to the core/module concept.

AMD consistently 'builds' incremental performance on their base designs. The best analogy that comes to me now is they step up to the plate and hit singles -- not necessarily home runs.

It's nice to smack a home run, and I hope they do with BD, but their limited resources tend to force them to be more incremental.

[Sn0wm@n]

Yes it is.
If you take a look at the architecture itself, it's not hard to see that BD's IPC will not merely match K10.

As far as core scaling goes, when both BD cores in each module are being used, AMD themselves said multi-threaded performance would increase by 80% (remember 80% more than 100% is 180%)
And that's not compared to K10. That's compared to doing one thread on one BD core.

AMD also said they got "50% more performance from 33% more cores"
Of course they didn't say what type of workload. But in that context I'd assume it means something that uses >=8 threads (still not THAT specific).

If I combine their two statements:
50% more performance from 33% more cores
When both cores in each module are used, you get 80% more perormance (180% OF single threaded performance)
and assuming perfect scaling when all four modules are active, I'd conclude single threaded performance to be
[1/(1.80/2)]*(1.5/1.33)~=1.253 (are these calculations right?)
i.e. single threaded performance is about 25% better in the particular type of workload AMD tested with... but becomes 90% as much as that when both cores in each module are used (unless there's some workload where sharing the cache can result in a performance benefit)
0.9*125%=112.5%

Although yes, that's taking AMD's statements at face value.

wasnt the 50% more performance throughput oriented from the memory controler improvement + higher speed ..


so this renders half of your logical performance analysis wrong

[DarthShader]

What I'm trying to say is that OBR has said that core per core BD has same performance as Phenom II (if I understood that correctly), you compared a 4 module BD (8 cores) with a 6 core Phenom II and saying it has 33% more cores (If i read it correctly) but if you would compare a 6 core BD to a 6 core Phenom II (and according to OBR) they would have the same performance, since core per core performance is the same.

If a x6 BD 6 core perform the same in multithreaded apps, then BD single thread perf will be higher, due to module construction.

[repman]

Even if IPC core per core would be the same,you still have 33% more of them for multithreaded work.You have new instructions for specialised code (AES SSE4.1 4.2 etc) which will give biig boost for some parts of the code.And on top of that you have big turbo for single threaded works.All that at lower manufacturing costs than thubans.

Don't forget 4 and 6 core versions of bulldozer, which would mean that the 6 core version would perform the same as Thuban (according to OBR that is), that's why I doubt AMD would invest so much into a new chip for the same performance they have now. IPC core per core should at least be 10% more for Bulldozer, but that's just me guessing