AMD presents "The Bulldozer Blog"

[terrace215]

If my app is restricted to 2 threads, or any other number below the max, your CPU will underperform compared to the competition.

Yes, it should be particularly bad (relative to SB) at any number of active threads up to 50% of the cores in the AMD part, especially in integer work. And when the green fans see the gaming results, we'll see the five stages of grief.

[Manicdan]

does anyone know how well BD will clock on normal desktop voltages?

[terrace215]

does anyone know how well BD will clock on normal desktop voltages?

That will depend on the degree to which GloFo can fix their 32nm process... so it has to be too soon to tell, at this point.

[JF-AMD]

Your goal is to sell CPUs, fooling the customers if necessary.

Fooling someone will allow me to sell a processor. Telling the truth is what helps you sell millions of processors.

There is way too much on the line for me to be "fooling" people.

[Manicdan]

That will depend on the degree to which GloFo can fix their 32nm process... so it has to be too soon to tell, at this point.

so you can just stop with the "will be" attitude about it until then,
thanks

[JF-AMD]

Yes, it should be particularly bad (relative to SB) at any number of active threads up to 50% of the cores in the AMD part, especially in integer work. And when the green fans see the gaming results, we'll see the five stages of grief.

Actually, I can't get into specifics, but I would be willing to bet that my processor actually performs better at half load than anyone here would ever expect.

It will be very interesting to see a 16-thread SB with 8 active threads vs. a 16-thread interlagos with 8 active threads.

[informal]

Yes, it should be particularly bad (relative to SB) at any number of active threads up to 50% of the cores in the AMD part, especially in integer work. And when the green fans see the gaming results, we'll see the five stages of grief.

Let me get this straight: you haven't seen gaming results for SB,let alone for BD desktop part(or any aspect of BD performance for that matter);you claim that in integer work, with 50% utilization ,BD will be "particularly bad" without providing any source or at least a reason for this line of thinking. All this makes me think that you just failed big time.

First of all,we haven't seen gaming results for SB.It may be great at gaming for all we know.It may be the same as i7 we have today or slightly better.We have few results where Turbo is not working well which show performance uplift up to 10% compared to Lynnfield and a geek bench result from a mobile SB with obviously working Turbo mode which shows ~20% uplift(which is very good).No gaming results yet.

Second of all,we haven't seen any performance results of any Bulldozer variation.Bulldozer will have noticeably improved integer units relative to Stars cores,which will be 4 issue and each core will be capable of 2 loads and 1 store(matching SB core).It will have more advanced Turbo mode too,meaning very aggressive up clocking when CPU is underutilized (the scenario you previously described as "particularly bad"). It will have much more potent fpu unit than Stars cores that can be split in two and will support AVX. It will have more L3 cache and it will have a new shared L2 level cache and probably a trace cache. There is also a hint of a so called "accelerate mode" in which BD's front end can dispatch 2x more instructions to the module's 2 cores under certain circumstances(dresdenboy's blog).There will probably be a lot more prefetching and speculative execution going on in the design too.Now we don't know how much of an improvement all this will bring to gaming(which is mainly a GPU thing) but I think it will bring very noticeable gains in traditional and parallel workloads.
Now it may be the case that when underutilized one BD MPU will trail one SB MPU but to claim this to be the case now ,you would have to have both chips in your hands today and running actual tests on them.Which I'm sure you are not.

[nn_step]

Actually, I can't get into specifics, but I would be willing to bet that my processor actually performs better at half load than anyone here would ever expect.

It will be very interesting to see a 16-thread SB with 8 active threads vs. a 16-thread interlagos with 8 active threads.

can you comment on if there has been any improvement in latency on bit-manipulation instructions: Leading Zero Count (LZCNT) and Population Count (POPCNT)?

[MrMojoZ]

We'll talk again when BD is released. If I remember about it, which is going to be difficult.

So you don't care at all about BD, you are just trolling? Just want to clear this up. If I went into the Sandy Bridge threads and spewed the same negative crap as you, you'd think that was fine? I don't understand your perspective here, please educate me.

[terrace215]

you claim that in integer work, with 50% utilization ,BD will be "particularly bad" without providing any source or at least a reason for this line of thinking.

I thought it was fairly obvious. Up to 50% activity, you have a full SB core with no HT load against 1/2 the module. Add in that Intel turbo >= AMD turbo (probably >> IMO), and there you go. For FP, at least in 256-wide work the module had to share to begin with, so the relative performance increase of half a module is greater when you stop loading the other half.

There was a thread back in the day on amdzone, where people started asking about single-threaded performance. JF's answer was first that it doesn't matter to server customers, and then, when pressed, well, it will be better than current k10.5 single threaded performance. Reading between the lines, it won't be great. Clearly not in SB's league. But what many people miss is that single-threaded performance directly relates to 50% loaded performance (more so integer), too, due to BD's module design, and SB's HT design.

So there, you have 2 indicators.

Like I said, a point of relative weakness. It's relative strength will be stuff that can truly continuously use all the cores.

[terrace215]

Actually, I can't get into specifics, but I would be willing to bet that my processor actually performs better at half load than anyone here would ever expect.

Are you kidding? Maybe better than *I* would expect, but certainly worse than the biggest fans here are expecting.

Fooling someone will allow me to sell a processor. Telling the truth is what helps you sell millions of processors.

There is way too much on the line for me to be "fooling" people.

Oh, come now. The marketing gentleman doth protest too much.

Sure, sure, if you told outright lies, that would poison the market, though you might be able to fool millions before too much feedback was out there.

But lets not pretend that your goal is to present the unvarnished, objective truth of the tradeoffs between your product and the competition. Your employer would (rightfully) not be be pleased.

So it's a gray area called "marketing". That's fine. You're an advocate for your products. You seek to emphasize their strengths, and minimize their shortcomings.

[informal]

I thought it was fairly obvious. Up to 50% activity, you have a full SB core with no HT load against 1/2 the module. Add in that Intel turbo >= AMD turbo (probably >> IMO), and there you go. For FP, at least in 256-wide work the module had to share to begin with, so the relative performance increase of half a module is greater when you stop loading the other half.

So you skip everything else I have written (because you know I have a point) and make a claim ,again,without knowing few important things like : a) intel's Turbo being much more aggressive than AMD's version ? b)a performance of "half the module" (you do realize that "half the module" is actulay a full fledged 4 issue improved integer core with an access to 256 FMAC unit,shared L2 cache+shared L3 cache and a new and improved Turbo Core ability?).
The perf. "penalty" of fully loading up a module is ~10% so single active core will be actually performing better even without the Turbo,let alone with it.

[Manicdan]

Are you kidding? Maybe better than *I* would expect, but certainly worse than the biggest fans here are expecting.



Oh, come now. The marketing gentleman doth protest too much.

Sure, sure, if you told outright lies, that would poison the market, though you might be able to fool millions before too much feedback was out there.

But lets not pretend that your goal is to present the unvarnished, objective truth of the tradeoffs between your product and the competition. Your employer would (rightfully) not be be pleased.

So it's a gray area called "marketing". That's fine. You're an advocate for your products. You seek to emphasize their strengths, and minimize their shortcomings.

your company tries to compare turbo to the invention of email and wireless routers. intel thrives on the idiocracy of mankind

[crazydiamond]

ing typical

cant open 1 amd thread without same trolls talking same nonsense

[LesGrossman]

Fooling someone will allow me to sell a processor. Telling the truth is what helps you sell millions of processors.

There is way too much on the line for me to be "fooling" people.

I know you are a server guy, but in the blog you were so straightforward with bulldozer being compatible with the current platform in the server world, can you dig and share some info about bulldozer desktop version running on AM3 boards?

[madcho]

Yeah, marketing always overhypes things. I said Magny Cours was going to be 50-60% higher than istanbul. I was really wrong.

80-120% was more like it.

What people forget about any predicitions is that it does no good to overhype your products. If I said it was going to be 300% faster and it wasn't a.) I blow all credibility and b.) the product is viewed as a failure.

I totaly agree, K10 was a bit overhype ... And i understand the way to send conservative numbers.

Also, don't forget that we have a boost technology. The 50% number is based on a fully utilized processor. If you were only running a single thread you would be looking at a very different uplift.

That's awesome

I hope AMD is gonna deliver some new FX chips. So long we havn't seen them.

[terrace215]

So you skip everything else I have written

When BD's advocate downplays the importance of its single-threaded performance, I take the hint.

So either you disagree with this point, or you don't see that relative single-threaded performance is a nearly perfect indicator of relative 50%-of-fully-threaded performance. It is only past 50% that the big design differences (modules vs HT) start to kick in. And until 50% we don't really need to worry that power or memory bandwidth is going to be all that constraining, and besides, they're on fairly even ground with those factors.

I'll assume you do understand the latter, so I guess you are arguing that BD is going to be very close to SB in single-threaded performance.

I don't buy it.

[Manicdan]

When AMD's Server Marketing dude discusses server workloads, I take the hint.

fixed

[JF-AMD]

can you comment on if there has been any improvement in latency on bit-manipulation instructions: Leading Zero Count (LZCNT) and Population Count (POPCNT)?

no, I can't

[informal]

When BD's advocate downplays the importance of its single-threaded performance, I take the hint.

So either you disagree with this point, or you don't see that relative single-threaded performance is a nearly perfect indicator of relative 50%-of-fully-threaded performance. It is only past 50% that the big design differences (modules vs HT) start to kick in.

I'll assume you do understand the latter, so I guess you are arguing that BD is going to be very close to SB in single-threaded performance.

I don't buy it.

Actually he replied to you:

Actually, I can't get into specifics, but I would be willing to bet that my processor actually performs better at half load than anyone here would ever expect.

It will be very interesting to see a 16-thread SB with 8 active threads vs. a 16-thread interlagos with 8 active threads.

[terrace215]

Actually he replied to you:

So, what's your prediction for single-threaded performance of BD vs SB?

[-Sweeper_]

So, what's your prediction for single-threaded performance of BD vs SB?

look who you are asking..

[madcho]

there is Xtrem funny people here ^^

[Manicdan]

we need to have a paypal setup gambling section of the forum. let the fanboys take some real risk with what they say. (this applies to all variations of fanboys)

[STaRGaZeR]

Fooling someone will allow me to sell a processor. Telling the truth is what helps you sell millions of processors.

There is way too much on the line for me to be "fooling" people.

A good product sells itself, a bad one doesn't. Marketing is there to help the good products to sell even more and to help the bad ones to sell something. Marketing teams almost always tell the truth. The problem is not what they tell you, it's what they don't tell you. You, with emphasys in you, are fooling your potential customers just as much as any other marketing guy.

I'll say this one time only: let's have a clear and sincere conversation. Sincere doesn't mean sharing results. Don't try to fool me because it's not going to work. Don't take this as an offense, again I'm sure you know what I'm refering to, so please.

Actually, I can't get into specifics, but I would be willing to bet that my processor actually performs better at half load than anyone here would ever expect.

It will be very interesting to see a 16-thread SB with 8 active threads vs. a 16-thread interlagos with 8 active threads.

I'll take that bet. I can guarantee you right now that your processor won't be able to match my expectations.

And yes, it'll be very interesting. I'm looking forward to it.