AMD Zambezi news, info, fans !

[PatRaceTin]

many calculation he he

[Smartidiot89]

beep: of course, but what is exactly IPC? Maybe here is not talking about core to core, but about cores to cores (6 vs 8 etc)

"In reality what we are doing is driving efficiency. And don’t worry about the single threaded performance –we have already stated publicly that Bulldozer single threaded performance is expected to be higher than our current core architectures."
http://blogs.amd.com/work/2010/09/13...stions-part-3/

Seems like everyone is expecting Pentium 4 v2.0 with Bulldozer just cause it's a high-clock design.

[TESKATLIPOKA]

Manicdan Nice chart, I made my own.
BD default 3.6Ghz, Turbo1 4.2Ghz, Turbo2 3.9Ghz
Thuban default 3.3Ghz Turbo 3.6Ghz
SB 2600K default 3.4Ghz Turbo 3.8->3.5Ghz
http://www.anandtech.com/show/4083/t...-2100-tested/3

Performance Index per core at respective default frequency
Thuban 100b
BD 125b, module 225b (180% of one integer core)
SB 140b, HT 35b (25% of one core)

Chart link
http://sheet.zoho.com/public/teskatl...ance?mode=html

[drfedja]

Performance Index per core at respective default frequency
Thuban 100b
BD 125b, module 225b (180% of one integer core)
SB 140b, HT 35b (25% of one core)

Chart link
http://sheet.zoho.com/public/teskatl...ance?mode=html

It is little bit complicated.
Don't forget Amdahal's law. It tells us how threads scale up with number. Also it depends on software optimisation and paralelisation. Six cores doesn't scale 50% better than four cores. Also, eight cores / threads doesn't scale 33% better than 6 cores/threads. But your chart is useful if you compare same number of threads. For example, this is useful for comparing Sandy Bridge and Zambezi and Zambezi with Thuban up to 6 threads/cores.

Attachment 119447

For example, with 95% of paralelization, 80% of gain with modular approach and 30% of gain with hyperthreading, Bulldozer with 8 threads will scale up to 6.26x, Thuban with 6 threads will scale up to 4.8x, Westmere with 6cores / 12threads will scale up to 6.24x, Sandy B. will scale up to 4.52x. That isn't 33% for 33% more cores.

Of course, SB will be faster than Thuban, because of higher IPC.
For 40% higher IPC SandyB will have 7.91 performance points, for 30% higher IPC Westmere will have 9.27 points, and for 30% better module IPC and 3.2 GHz base clock and +400MHz TB Bulldozer will have 9.06 IPC.

[TESKATLIPOKA]

drfedja Yeah, I know It's not 100% correct because I didn't bother with Amdahal's law,
but that wasn't important for me, I just wanted to know the performance between different cpu's at the same thread count.

And here is another one, BD and Thuban have the same IPC only frequency is different. Not like I believe the IPC will be the same
http://sheet.zoho.com/public/teskatl...ce-2?mode=html

[chew*]

I'm going to save a copy of all these charts you guys are making. When NDA lifts I will show you mine

[Smartidiot89]

I'm going to save a copy of all these charts you guys are making. When NDA lifts I will show you mine

I approve of this

[informal]

I'm going to save a copy of all these charts you guys are making. When NDA lifts I will show you mine

You have to cut them some slack,they are working with virtual threads while you are working with hardware ones .

[TESKATLIPOKA]

chew* I don't mind. It's just a calculation I made on my imaginary Bulldozer.

[chew*]

You have to cut them some slack,they are working with virtual threads while you are working with hardware ones .

True but they are grabbing certain "values" out of thin air, and those values they are grabbing from thin air are the most important ones that can grossly skew the end result.

Especially since I already showed them how a native 6 core part compares to a 4 core 8 thread part with much higher IPC than the 6 core part.

Sandy with all it's IPC in a multithreaded bench 8 threads versus 6 is only 00.XX% better than thuban.......

By the magical math these guys are doing, BD = 34% better in the same compare which makes it 33.XX% better than sandy but they will sell it for same price

All I'm saying is if your going to try to guess, consider all things and try to be a little more realistic considering the info people do know versus what they don't.

What we do know is suggested price.

[jimbo75]

I get what you are suggesting chew but that argument falls a bit flat if you consider BD vs Thuban, both on price and the expected performance of a shrunk Thuban with 2 more cores. If Thuban on 32nm would have been faster, AMD have just wasted years on a broken arch.

[TESKATLIPOKA]

chew* 33% was never average performance gain. Maybe my second chart will be more to your liking but that would be a pretty weak performance also AMD always had better performance/price ratio than Intel and asking 300 dollars for a BD with the same performance as a 9-10 months old SB is not very funny.

[drfedja]

True but they are grabbing certain "values" out of thin air, and those values they are grabbing from thin air are the most important ones that can grossly skew the end result.

Especially since I already showed them how a native 6 core part compares to a 4 core 8 thread part with much higher IPC than the 6 core part.

Sandy with all it's IPC in a multithreaded bench 8 threads versus 6 is only 00.XX% better than thuban.......

By the magical math these guys are doing, BD = 34% better in the same compare which makes it 33.XX% better than sandy but they will sell it for same price

All I'm saying is if your going to try to guess, consider all things and try to be a little more realistic considering the info people do know versus what they don't.

What we do know is suggested price.

On average, BD won't be much faster than SB. Look at Westmere vs SB. There isn't so much difference on desktop. That is because most of workloads are serialised. I think that in gaming, or general usage BD isnt going any faster than SB, but in workstation performance, difference will be significiant.
Let's talk about prices: BD at launch will be priced 300$, and Intel going to cut prices of SB. Also they are going to release SB-E to compete and hold performance crown.
AMD wants with BD best price/performance ratio.

If Thuban on 32nm would have been faster, AMD have just wasted years on a broken arch.

I agree!

[FlanK3r]

drfedja Yeah, I know It's not 100% correct because I didn't bother with Amdahal's law,
but that wasn't important for me, I just wanted to know the performance between different cpu's at the same thread count.

And here is another one, BD and Thuban have the same IPC only frequency is different. Not like I believe the IPC will be the same
http://sheet.zoho.com/public/teskatl...ce-2?mode=html

this like me, I think, it is near to real results. We will see

[freeloader]

I'm going to save a copy of all these charts you guys are making. When NDA lifts I will show you mine

That will be cool. Are you under NDA to tell us when the NDA lifts?

[Evantaur]

Are you under NDA to tell us when the NDA lifts?

i've heard a rumor of a rumor that said it's 9/9/2011

[TESKATLIPOKA]

FlanK3r That wouldn't be good. In that calculation BD has the same core performance as deneb when using just one integer cluster in a module. The only advantage over Thuban was the higher default clock-speed and turbo. If thats the truth they should have just scrapped BD and used 8 Llano cores instead because they have better performance(6-7%) than Deneb core without L3 cache and wouldn't have the sharing module penalty, the size would be pretty much the same.

[nex_73]

I'm going to save a copy of all these charts you guys are making. When NDA lifts I will show you mine

What we do know is suggested price.

As the calculated price per core is +18,42%, here is my new numbers:
Zambezi vs Thuban (FX 8150 vs 1100T)
Attachment 119456

[Smartidiot89]

i've heard a rumor of a rumor that said it's 9/9/2011

Get out of here, that's my birthday you heard about

[chew*]

As the calculated price per core is +18,42%, here is my new numbers:
Zambezi vs Thuban (FX 8150 vs 1100T)
Attachment 119456

dump the thuban compare, recalculate it to sandy, then compare both to thuban.

[chew*]

I get what you are suggesting chew but that argument falls a bit flat if you consider BD vs Thuban, both on price and the expected performance of a shrunk Thuban with 2 more cores. If Thuban on 32nm would have been faster, AMD have just wasted years on a broken arch.

Your forgetting X factor.

TDP.

Doesn't matter if it would be faster if it can't fit into an acceptable TDP range.

[drfedja]

Here is my latest calculations. BD performance vs IPC vs module efficiency In that excel file you can add or remove modules/cores, turn on or turn off HT. Module efficiency depends on utilisation of shared resources. Of course, it will be lower with FP intensive than pure integer.
I've also made script for turn on, or off Thuban turbo mode. If there is less than six cores, CPU is Deneb, and there is no turbo for singlethread.
Also I've calculated that with amdahal's law, and made mathematical model for calculating performances of modern CPU's in that few benchmark app's.

[BeepBeep2]

I get what you are suggesting chew but that argument falls a bit flat if you consider BD vs Thuban, both on price and the expected performance of a shrunk Thuban with 2 more cores. If Thuban on 32nm would have been faster, AMD have just wasted years on a broken arch.

IPC IPC IPC IPC IPC IPC IPC IPC IPC IPC IPC IPC IPC IPC IPC IPC IPC IPC IPC IPC IPC IPC IPC IPC IPC IPC IPC IPC IPC IPC IPC...

I dont expect BD to do much better than Thuban in multi-threaded workloads, stars architecture is amazingly good at 8-16 thread workloads provided the program takes advantage of them. I dunno how many times I've said this but per mm2 (minus cache), Deneb/Thuban was very competitive with 45nm Core i7 performance vs core size. Sure, we could have 10 core K10.5 on desktop but that doesn't help the fact that AMD is 25-30% behind intel in IPC

[TESKATLIPOKA]

chew* Thuban had the same TDP as many Phenoms and they were on the same process just older Stepping so I don't see any reason why they can't make an 8 core Thuban with the same TDP on a smaller node.

[Manicdan]

I'm going to save a copy of all these charts you guys are making. When NDA lifts I will show you mine

Great!, id love to know how close i was.

True but they are grabbing certain "values" out of thin air, and those values they are grabbing from thin air are the most important ones that can grossly skew the end result.

well theres only a few values i worried about:
clock speed with and without various turbos
IPC with one thread per module
IPC with 2 threads per module

all 3 of those are still hidden and or rumored, and they are pretty much the ONLY things my calculations are made of.
so damn right im guessing by pulling numbers out of thin air , thats why its so fun to see how close my guesses are.