Zambezi ES performance weirdness

[informal]

Ok since it would be a longer post here,I created a blog just for this.In the blog I try to speculate(just that,speculate!) what might be the deal with all the gimped Bulldozer ES scores that are floating there.
In this link I summed up my thoughts on what's really going on with the ES that are floating in the wild(mostly China).
Everything I wrote is just my personal opinion and expectation.It may be and probably is way off.This is the disclaimer .

[FlanK3r]

I think, with Llano performance is not so bad, only CPU-Z shows bull*ts...Problem is baseclock, my new "theory" is, baseclock of Llano is 100 Mhz, not 200 MHz (Its possible?)

[informal]

I think, with Llano performance is not so bad, only CPU-Z shows bull*ts...Problem is baseclock, my new "theory" is, baseclock of Llano is 100 Mhz, not 200 MHz (Its possible?)

Of course,it fits perfectly,2x less then what is shown .

[XRL8]

Makes sense Informal.
But its easy to test.Send it to chiphell, they could do simple power draw tests with phenoms and zambezi at diff multis and check.Zambezis powerdraw for the whole system should be around -60W from standard phenom II (were talking prime95 here to fully engage cores).

Also, it would mean that it should score almost the same if set manually to lets say 1600mhz

[RussC]

The SuperPi and single thead tests are not useless. As I and others have posted, the company that wins the single thread/core and internal memory/cache performance, gets the spoils, plain and simple. Again, either company can stack cores till the cows come home and play the leap frog game of performace king for a year till the next higher core count part comes out(the same way AMD and NVDA are fighting it out). But the company with the better core/thread performance wins market share, sorry.

RussC

[jimbo75]

The SuperPi and single thead tests are not useless. As I and others have posted, the company that wins the single thread/core and internal memory/cache performance, gets the spoils, plain and simple. Again, either company can stack cores till the cows come home and play the leap frog game of performace king for a year till the next higher core count part comes out(the same way AMD and NVDA are fighting it out). But the company with the better core/thread performance wins market share, sorry.

RussC

And yet you bought a phenom II?

[nex_73]

IF and i also say IF! there is some possibilities this is the case, we can se numbers like these in the coming reviews. Heck - the new B2 & C0 rev steppings ought to be better optimized right?

[FlanK3r]

Informal:finally they using hwinfo and Mumak confirmed baseclock as 100 MHz for Llano

[Blaber]

What do you guys make of this supposedly few bench numbers of Bulldozer "no stepping data , but I think its B1"

http://obrovsky.blogspot.com/

[2senile]

Thanks for the very reasonable explanation Informal. Just a quick question tho; why does "everybody" seem to think retail Zam' will launch at such a low speed as 3.2Ghz?
From what JF posted, repeatedly, the server part was designed for 3.5+ Ghz. (Maybe I misunderstood tho plus I haven't really been following things closely)

[TESKATLIPOKA]

2senile: 3.2Ghz frequency is without a turbo witch should add another 1.2Ghz when using just 1 core

Blaber: the chip had a B1 stepping, but it seems some problems were with turbo as the author of the blog mentioned, so the numbers aren't very accurate. If the turbo didn't work or just worked partially is still unknown.

[2senile]

2senile: 3.2Ghz frequency is without a turbo witch should add another 1.2Ghz when using just 1 core

I think I understand what you are saying .... however ... I assumed JF meant that 3.5 Ghz was the optimal starting frequency due to the slightly longer pipeline. He does only talk about server where all cores should be loaded anyway & Turbo (I assume) would not factor to such a great extent. I just assumed that with 3.5 being optimal the Desktop version would launch at 3.5 or higher without Turbo involved.

Many thanks for taking time to explain.

[TESKATLIPOKA]

I don't know exactly how he meant it but turbo should work even with all cores used if the TDP isn't reached and in server it should be +500Mhz boost for 16 core Interlagos.

[radaja]

JF-AMD said they didnt cripple the ES sample?so it must be a bios code problem right?

this is from AMDzone

You guys do realize that a.) engineering samples will not represent nor correlate to actual final performance and that b.) we don't "cripple" engineering samples, right?

[TESKATLIPOKA]

Good question, but who knows the only thing known is BD doesn't work as it should.

[SpuTnicK]

IF and i also say IF! there is some possibilities this is the case, we can se numbers like these in the coming reviews. Heck - the new B2 & C0 rev steppings ought to be better optimized right?

You have already been told to provide source for those numbers, otherwise, great Excel work

[nex_73]

You have already been told to provide source for those numbers, otherwise, great Excel work

The source? It's myself - my own speculations based on informals speculations in first post... shall be fun to see how close we are when real numbers comes out...

[demonkevy666]

The source? It's myself - my own speculations based on informals speculations in first post... shall be fun to see how close we are when real numbers comes out...

what was the factor of improvement ? 1.6....

I was looking a some cpu-z with the 2.8ghz did anyone else notice it show 0.969 volts ?
seem like 1.4ghz would need more volts then that.

llano's base clock at 100mhz doesn't make sense with the ram mutli's unless someone flub and stuck in bulldozers Dividers in there lol, how ever it does fit the right performance level.

[nex_73]

what was the factor of improvement ? 1.6....

Approx 2x the leaked B0 3.2GHz with C6 off score.
What is interesting here is: as AMD have hinted about 50% more perf with 33% more cores, you can see it's pretty close to that if compared to 1100T scores

[cesariuth]

Guys what do you think about clocks stock for Bulldozer...3.2GHz or 3.5GHz??? and TC??? 4.0GHz ...4.2GHz???

What about two mode of TC???

[Aussie FX]

The SuperPi and single thead tests are not useless. As I and others have posted, the company that wins the single thread/core and internal memory/cache performance, gets the spoils, plain and simple. Again, either company can stack cores till the cows come home and play the leap frog game of performace king for a year till the next higher core count part comes out(the same way AMD and NVDA are fighting it out). But the company with the better core/thread performance wins market share, sorry.

RussC

Superpi is absolutely useless, it's a relic from nearly 30 years ago when we used to install our own co-processors.
AFAIK AMD has deprecated x87 in BD and it will be emulated in the fp pipe. Don't be surprised if 1m times are out over a minute.

Still think it's a great performance indicator?

[PatRaceTin]

hmmmmmmmmmmm

[undone]

I find the bulldozer L1 write is even lower than Deneb, I don't understand.

[mbevo]

I find the bulldozer L1 write is even lower than Deneb, I don't understand.

Probably a software bug, considering l2 and l3 writes are 0.

[informal]

I find the bulldozer L1 write is even lower than Deneb, I don't understand.

It's useless to use these results as any indicator of how Zambezi performs.What happens if retail performs much better than Thuban even if the cache benchmark numbers don't change significantly?Yep,it won't matter much. We all care how it does in real world workloads.

Btw I updates my blog with some new numbers,check it out.