Zambezi ES performance weirdness

[XRL8]

A. I thought I already addressed ES on this forum, why are we still beating that dead horse?

Because you have not said anything of substance.You said basically that ES can have vastly different clocks than the final product.We know that.That does not explain B0 steppings performance.

yes this is why i had that thought, i mean if they have a working enhanced BD core already maybe this is the real reason for the zambezi delay.

im sure amd want all the performance they can get when bringing zambezi to market?

If youre implying that theyve delayed BD in order to cram BDv2 cores in there, its just not true.Its too much work and time needed for such monumental overhaul.And showing trinity does not mean it was even functional.With a year to premiere they have some time to get there.

[Greg83]

i'm sure he did address the first thing. cause i was able to understand exactly what he said.
and regarding the second. it was addressed with his first post. obviously. you just didn't read it right. try a few more times.

[FlanK3r]

xlr: I say, dont worry....

[LowRun]

showing trinity does not mean it was even functional.With a year to premiere they have some time to get there.

They showed a Trinity laptop playing a video. Most people would agree it qualifies as functional.

[XRL8]

i'm sure he did address the first thing. cause i was able to understand exactly what he said.
and regarding the second. it was addressed with his first post. obviously. you just didn't read it right. try a few more times.

He did not specifically said why current B0 ES`s are behaving like they are.And yes, i read it one more time to be sure.No need to be fanboyish about it.
He only implied that theyre not in final clocks stage (we knew that).He also said that performance is a derivative of few key points (yes i undestood which ones).
Now.Mainboards are READY. so Its not it.
He also said that they DO NOT CRIPPLE ES`s in any way or form.So thats just what they have.And its not any intentional crippling of teh cpus.
So, its not mainboards and its not AMD evil doing to hide its performance like many have suggested. Get that ?
Now, it means B0 is broken in some way (not talking about clocks that was to be expected).
So, to at least hope that performance COULD be better we must know, at which stepping AMD currently is.

You ofcourse can read his statement, which, while true ,doesnt give you any useful bit of information and just hope for the best.Myself, i like hard info.

@LowRun

If thats the case than thats good, all i have seen was a picture of trinity APU.Will search for said video.

[Greg83]

sorry to come off as a fanboy. only thing i'm not a fan of is nvidia after my 680i and dead ram and how they dealt with it.
systems went cele, k8 754, k8 939, core 2 duo, core 2 quad and another core2quad. and a bunch of others for friends systems based on their budgets. I always used my ncix PP discount to get them best bang for their buck.
next might be this 8 core. depends on how low i can set the cpu voltage and speed. i don't have huge needs for raw clock speed. but low power usage, while maintaining enough performance for what i use is what its about to me.
But raid quality is important to me too. will have to see how far its come along on amd 990fx before i switch. only need it to perform on par with intels ich9r for raid hard drives. ssd's i will only need one sata 3.0 one and would be able to handle buying a raid card specially for that.

I saw this "All of this talk about engineering samples is really annnoying because they are not designed around performance. So trying to measure something for performance when it is designed for validation is really pointless."

they're not crippled. they were never designed for performance in the first place. only validation

maybe JF can PM me regarding if 990fx would pull around 500MB/s or greater with a 6 drive raid 0. permitted the disks can handle it. also if there is write caching like intel has used a while to get those 3000MB/s plus burst speeds.

[XRL8]

...
I saw this "All of this talk about engineering samples is really annnoying because they are not designed around performance. So trying to measure something for performance when it is designed for validation is really pointless."
they're not crippled. they were never designed for performance in the first place. only validation...

Maybe i should explain myself a little bit.
We all know that ES`s are meant for validation purposes.Thats why they most often than not, have lower clocks with higher voltage and sometimes even feature turned off.
Problem is, Current B0 samples are behaving VERY badly compared clock for clock to DENEB,plus some scores are just plain weird.
Most of the people assumed it was AMD`s strategy to blindfold us and competitors from knowing true performance.
JF said it is not the case.
So it is or was some real problem with B0 stepping cpu.
So all we know for sure ATM, is that B0 BD cpus have some kind of error, in silicon or microcode.JF has not stated that, or said that there is another stepping that fixes that, he basically havent said anything.He cant, he works at AMD and current policy is not to tell anything.
Let me be clear once again, BD B0 is currently slower Than deneb or even athlon CLOCK FOR CLOCK.That cant be neither good, neither fault of the ES designation, as it was supposed to be faster clock for clock than deneb.Look at informals sig ;-).Whole another thing is different retail clocks and turbo.

As for SB950, its the same as SB850.Carbon copy.Maybe drivers will improve something.I wasnt aware that theres something wrong with Sb850 tho.

[informal]

@XRL8

None of the tested ES are tuned for performance,just validation.Does this answers your question?

[nex_73]

1. The processor
2. The processor microcode
3. The BIOS
4. The Operating system
5. The drivers
6. The compiler code

1. Obviously only final silicon is inside AMD at this time
2. As motherboards have been shipped i hope it's somewhat ready? - sure, updates bundled with bios will come
3. See above (2)
4. I don't know how ready Win7 is at the moment, but Win8 ought to have all the support for the instruction sets? -maybe updates will come via WU?
5. As the cpu isn't launched yet, this is expected.
6. Up to developers. I hope AMD will put more resources to work tighter together with developers?

[rintamarotta]

ES's ment for validation. Not speed comparsion, comparsions are invalid at this point.
There is problem in B0 stepping, we dont know it what it is but im pretty shure AMD knows where the problem is and they need figure out how to fix it, they might have allready fixed it.
I expect B1 stepping have alot of less problems compared to B0 but i think the cpu that goes to retail is B2 stepping or then they get lucky and B1 will fix problems, they might even be allready producing B2 chips.
They might even have allready fully working chip ready.

We can only speculate if its SB950 or Zambezi B0/B1 problem and what stepping they are producing at the moment.

Why all Zambezi/BD threads get derailed?

[Greg83]

So all we know for sure ATM, is that B0 BD cpus have some kind of error, in silicon or microcode.JF has not stated that, or said that there is another stepping that fixes that, he basically havent said anything.He cant, he works at AMD and current policy is not to tell anything.
Let me be clear once again, BD B0 is currently slower Than deneb or even athlon CLOCK FOR CLOCK.That cant be neither good, neither fault of the ES designation, as it was supposed to be faster clock for clock than deneb.Look at informals sig ;-).Whole another thing is different retail clocks and turbo.

As for SB950, its the same as SB850.Carbon copy.Maybe drivers will improve something.I wasnt aware that theres something wrong with Sb850 tho.

I saw this one on page 2

"Performance is determined based on the following:

1. The processor
2. The processor microcode
3. The BIOS
4. The Operating system
5. The drivers
6. The compiler code"

He is hinting at what is going on , only way he can.
Plus there is also errata, doesn't every processor have errors that must get fixed through errata updates anyway.
Thanks for the info about sb850. I'll try to find raid performance on that one.

[XRL8]

@XRL8

None of the tested ES are tuned for performance,just validation.Does this answers your question?

Of course not.Not without knowing exactly what "tuned" means.Its just a word.
Does tuned means lower clocks ?
Disabled functionality ?
If yes, what is disabled and how.

Rintamarotta pretty much explained it to us.There is a bug.Which we heard from MANY sources for a month now.It makes sense
Only other viable explanation was AMD purposefully crippling ES, as they are NOT doing that.B0 has a problem.

Getting back to the "tuned" part.JF himself makes it seem like it means higher yields/lower clocks (which is always the case with ES).But you,me,we all know that clocks isnt only problem with B0.

As for the thread derailing.Isnt this one strictly about why and how ES performs poorly ?

[informal]

Of course not.Not without knowing exactly what "tuned" means.Its just a word.
Does tuned means lower clocks ?
Disabled functionality ?
If yes, what is disabled and how.

Rintamarotta pretty much explained it to us.There is a bug.Which we heard from MANY sources for a month now.It makes sense
Only other viable explanation was AMD purposefully crippling ES, as they are NOT doing that.B0 has a problem.

Getting back to the "tuned" part.JF himself makes it seem like it means higher yields/lower clocks (which is always the case with ES).But you,me,we all know that clocks isnt only problem with B0.

As for the thread derailing.Isnt this one strictly about why and how ES performs poorly ?

I already made 2 blog posts about these buggy ES.EVerything I had to say about it ,I did it there.You can go and read what I think may be happening with these samples.

[JF-AMD]

I would bother to respond if people would actually read what I say. Unfotunately it doesn't seem to match their reality. Perhaps if you go back to my statement about engineering samples not designed for performance you'd understand that you are beating a horse that is no longer alive.

[Lightman]

I would bother to respond if people would actually read what I say. Unfotunately it doesn't seem to match their reality. Perhaps if you go back to my statement about engineering samples not designed for performance you'd understand that you are beating a horse that is no longer alive.

JF when will first commercial servers based around Orochi ship to customers? Will it be at the server BD launch or you're holding cards (CPU's) close to your chest before then?

I have a little server to upgrade and wondered if it's worth holding a month or two for it

[RussC]

please remind me how much money intel makes a year compared to amd ....


and btw if you believe intel's product are so great ... why dont you buy their product instead ... or their stock

Im not a shill for either company. I call them like I see them. All the equipment I have I mostly got for free by doing tech support for friends and family doing upgrades and fixes.

Sad fact is, AMD is having more issues gaining Market share due to other factors. Like Software Company requirements that say "Intel blah blah or higher" on the requiremnet specs. I never see AMD bosted on the top of the Specs sheets. You sometimes see "or Equivalent" which basically means the competitors option (AMD). And companies like mine where we blindly used Intel regarless of the perfomrnces at any price point. Even when AMD had the clearer Choice all around, They couldn't cut the market share due to these factors.

Yup, and the fact you can look at INTC like IBM, no one ever got fired for buying INTC.

Fixed that for you.

I think both AMD and Intel are working towards eager execution. In essence, instead of trying to predict which path to take when you come to a branch, take both. IMO, this is the next leap both companies need to take in single threaded performance. Bulldozer would give AMD a significant lead on developing this, IMO.

For one thing, AMD didn't make much of a dent in Intel market share even when they did have the faster chip in both single and multi threaded apps. So there is indeed more going on than simply who has the faster chip for any given application.

Of course my scenario doesn't ring true, because it hasn't been the case in recent memory. Before conroe AMD had the stronger performance in single and multi-threaded performance. And since conroe AMD hasn't had the faster chip for single or multi-threaded performance. And in that time they lost the small market share gains they did make. But if they do have the better multi-threaded performance I expect that they will regain ground in the segments I mentioned, or at least the server segment. Single thread performance doesn't mean much of anything in the server world. What matters is total throughput and power consumption while doing it. Why do you think the market was salivating at the thought of clustered ARM chips for server applications?

Right, there are a number of reasons AMD didn't take more maketshare away, two biggest reasons they didn't have the lead long enough and INTC's criminal activity(which they still in my mind haven't really been punished for). Now during that time, AMD had its best stock performance ever getting up over $40/share.

Well see about the performace of BD and see if it can make a difference. INTC still holds all the cards, to me and what they do is more important to who gets more marketshare than what AMD does. With that said, If BD is an out of the park performace leader, well see how that plays out and for how long, but we have a really good idea its not that.

Few things.

@RussC

The thing is, in this day, pure performance doesnt really matter to the 80% of population.
"good enough" is the mantra for todays computing.
And even NOT good enough atom made a killing.
And now AMD makes a killing with zacate.
Simple example, if i offer my girlfriend a new laptop with sandybridge, which will be too slow to play sims 3 without choking, shes not going to be happy.If Llano will do that, shes going to opt for that.
And it doesnt matter that some tasks that she rarely does are going to be slower.Because gaming NEEDS some kind of performance, but with encoding a movie clip she an wait a minute instead of 30secs.
Llano will do good too.Although it needs cpu part to be revamped, but its coming next year.
Only problems may be with desktop/server part as we dont have BD numbers yet.

Well, you skirted the issue, if BD cores were alot better than INTCs, the performance in the lower end would be there also and would be "good enough". Also, in that market segment, image has something to do with it also. When your cpu's are the fastest, the image of that even in the 80% segment may buy a AMD device because of all the hype. Plus, intagrators and OEMs will put alot more of your equipment into there products because of that also.

RussC

[XRL8]

I would bother to respond if people would actually read what I say. Unfotunately it doesn't seem to match their reality. Perhaps if you go back to my statement about engineering samples not designed for performance you'd understand that you are beating a horse that is no longer alive.

I have hard time believing that AMD "designs" ES parts specially for validating (i would think that youre guys just take everything that works without it needing to clock like retail).
But if youre telling that.Im gonna ask a simple question.
Did AMD "design" ES samples to perform vastly different than retail parts in something else than clockspeeds/turbo alone ?If yes, thats crippling.If no, B0 has a bug.

@Informal, yes i read your blog, it makes more sense than what JF is saying.However there are only two ways out of it.Crippling or bug.Seems like people here dont like when someone points out the obvious.

@JF, without saying what AMD did to ES in this special "design" we wont know nothing and youre not willing to say so thats impass.

It would be good to know if there are newer samples, but i doubt JF is going to tell us.

Well, you skirted the issue, if BD cores were alot better than INTCs, the performance in the lower end would be there also and would be "good enough". Also, in that market segment, image has something to do with it also. When your cpu's are the fastest, the image of that even in the 80% segment may buy a AMD device because of all the hype. Plus, intagrators and OEMs will put alot more of your equipment into there products because of that also.

RussC

Performance in the lower-end IS there for AMD, with zacate and Llano now.So i dont get what are you saying, nobody buys zacate to do serious work.As for Image, yes it has something to do with sales, however depends where and with whom.Most people dont know who Intel nor AMD is.They know Sony or apple tho.

[Manicdan]

my understanding is that B0 was being created as cheaply as possible, just so it can be recognized and have ENOUGH functions working to test them. things like half performing or half clock speeds does affect the bin they would have gone too for a retail part, but affects nothing when it comes to making a bios. simply put, they built the B0 chips to all be bottom bin, so they can have the highest yields.

-i am not an expert, so this isnt even an educated guess

[Greg83]

Seems to me.
ES only are used to get the cpu there. To program for it, optimize for it and prepare for the actual processor.
They were engineered only for people to design and program for
The bios' , the fixing of errata , the microcode updates, everything effects peformance
basically when windows sees it as something other then a Pentium 3.

All these contributed to the performance.

[SkullCracka]

JF works with servers and is on vacation. llano had some low numbers on samples we seen and now are showing great numbers. Not sure why you and some others seem to be taking all this so personaly and making it out to be some kind of grassy knoll conspiracy.

Chill ... breath... put down the mouse and walk away.

[JPQY]

Has this link already seen here?

http://www.chiphell.com/thread-210890-1-1.html

JP.

[dave_graham]

XRL8,

ES chips are designed for validation purposes, not for clockspeed, IPC, etc. They are designed so that ODMs/OEMs can integrate a mechanically and electrically compatible design. This is used to suss out HT bus signaling, socket fitment and tolerance, thermals (in some cases), as well as power delivery and pin/pad designs. Also, some work is done for BIOS development around AEGESA code integration, etc. They are NOT designed for performance. later steppings (e.g. B1/B2) can be used for validating CLK/BLK, etc. but at that point refinements to process are very much set in stone.

JF is correct.

dave

[XRL8]

To all the people.

There is a reason B0 performs badly.And its not intentional by AMD.
Design of the architecture and simulations were made before any silicon.
After that both Intel and AMD put out revisions, in which they fix bugs and increase clockspeed.
They do not design separate chip dies for validating purposes.Point of validating is having fully functional hardware, because u cant validate properly if you have a cpu that doesnt work like it should.
You can of course do it partially.
Pretty much every ES chip has the same IPC as final revision, or at least very near.
Thoroughbred A and B had the same performance, A was hot and clocked lower.They fixed that with adding another metal layer.But IPC was the same.
First Hammer ES`s at 800mhz had pretty much the same performance PER CLOCK as the one that eventually hit retail.
List is endless.Phenom B1 on the other hand, had to be crippled because it had a bug.

We all agree that B0 performs poorly (catastrophically).
JF says its a design choice by AMD.And its not about clocks.
I say Clocks are a ES problem always.However poor performance IPC wise is always due to some bug.

What im getting at, there is some technical pure and simple reason for weird performance figures.
And its not some esoteric marketing wording.
Performance comes from somewhere.It comes from IPC and Clocks.I AM DISCARDING CLOCKS in that regard!

And let me make myself clear.IM NOT TALKING THAT BD WILL BE A FAILURE ,or that it performs poorly.
However using logic im pretty confident B0 stepping has more problems than just clocks.

[dave_graham]

erff....XRL8, I've got a pair of ES X5650s with unlocked QPI (scalable from 5.8 -> 6.4 --> ?) and they don't work on anything but a binned Intel board. It was designed as a QPI validation part and consequently, even though it "works" it was never intended to do anything but help ODMs vett QPI clocking.

From an AMD perspective, B0 is never released to market as a production chip. B1/B2/B3 usually are.

dave

[LightSpeed]

To all the people.

There is a reason B0 performs badly.And its not intentional by AMD.
Design of the architecture and simulations were made before any silicon.
After that both Intel and AMD put out revisions, in which they fix bugs and increase clockspeed.
They do not design separate chip dies for validating purposes.Point of validating is having fully functional hardware, because u cant validate properly if you have a cpu that doesnt work like it should.
You can of course do it partially.
Pretty much every ES chip has the same IPC as final revision, or at least very near.
Thoroughbred A and B had the same performance, A was hot and clocked lower.They fixed that with adding another metal layer.But IPC was the same.
First Hammer ES`s at 800mhz had pretty much the same performance PER CLOCK as the one that eventually hit retail.
List is endless.Phenom B1 on the other hand, had to be crippled because it had a bug.

We all agree that B0 performs poorly (catastrophically).
JF says its a design choice by AMD.And its not about clocks.
I say Clocks are a ES problem always.However poor performance IPC wise is always due to some bug.

What im getting at, there is some technical pure and simple reason for weird performance figures.
And its not some esoteric marketing wording.
Performance comes from somewhere.It comes from IPC and Clocks.I AM DISCARDING CLOCKS in that regard!

And let me make myself clear.IM NOT TALKING THAT BD WILL BE A FAILURE ,or that it performs poorly.
However using logic im pretty confident B0 stepping has more problems than just clocks.

Man, give it a rest already. We CANNOT know what is wrong on B0, other than the fact that its an ES and not tuned for performance but rather validation. If that explanation isnt enough for you, I dont know what is. More details can come from AMD themselves, but that wont happen so youre out of luck there.

Lets just wait for B1/B2 and hope for the best