AMD's Magny Cours Architecture revealed

[abinstein]

What I am attempting to say is that I do not want to have EVER AGAIN have the experience of my development box being faster than production. Personally I think the production box should ALWAYS be faster than development or test. But I have had the opposite. That is a BAD situation. It ruins any guess I might have at how long production runs will take.

This makes perfect sense. Destroying the performance correlation between the development and production servers makes any guesstimate during the development phase a lot harder. We had a similar situation when we borrow a Nehalem based server for performance comparison. The performance scaling from 1 thread to 4 wasn't good, but we don't know whether its due to the lock/barrier implementation on the borrowed (remotely accessed) machine, or due to the fact that Turbo Mode was engaged when only 1 thread is running. That's when I found that Intel had a very ambiguous description of the Turbo Mode. I guess we will never know.

On the other hand, I really don't mind either Intel or AMD implements extra features into their processors, as long as I can easily disable them (best without reboot).

[JF-AMD]

JF-AMD, I don't wanna to see TurboBoost feature in Opterons period.

You will be able to disable it. Maybe it is just me, but I think more people are interested in the "reverse turbo" (i.e. powernow) than turbo.

[AussieFX]

I fully accept that AMD apparently has zero interest in competing int he high end workstation market. But I'll keep asking just the same.

I'm not so sure about that. I think that that AMD knows full well where it is positioned in the marketplace and their strategy reflects that. Moreso now with new managememnt.

I just can't see anything changing until Bulldozer and if it turns out that it's still behind then we will see more of the same.

Different machines can be good at different tasks and there is nothing wrong with that. A Buggatti Veyron will not beat an Enzo Ferrari around a racetrack, yet it is the faster car.

[Polonium210]

Since people seem to accept that BoXX knows what they are doing in the market, and has credibility then I guess I can cite BoXX and their product line as an authority on the subject at hand.


And very few AMD offerings outside the APEX series. Which as we know is a "workstation" setup, meaning it has a lot of extras not required by the RenderBoXX setup.

So, which extras of the APEX series are "intolerable"? What was your company's turnover for the last financial year?

[eaima]

Turbo mode alike would have a great success on the consumer side (i.e. Phenom III or IV) but I think it would mean unwanted transistor on the opteron system.

JF-AMD wrote:

Maybe it is just me, but I think more people are interested in the "reverse turbo" (i.e. powernow) than turbo.

I am more interested in energy saving feature too. Seeing a single thread running on multiple core will definitely be more powerful than instantly boosting an heavy single thread, even more when we will get 12 and 16 core on a chip.

About the future platform, don't it would make a great workstation for stream computing if we have graphic card connected via an hyper transport3 expansion slot? http://www.hypertransport.org/default.cfm?page=HTXOverview That would be great while waiting for fusion.

And having the fiorano platform supporting both DDR2 and DDR3 would be great too as consumer with particular need and deeper pocket may switch to DDR3 right away.

[abinstein]

Assume we care only about per-processor raw performance where the application scales well to multiple number of cores. Then it's apparent that AMD, even if lacking behind somewhat, is "pushing" Intel harder than ever in recent years.

The release of Istanbul is just a quarter behind that of Nehalem (actual availability for servers). The release of Magny Cours will probably be just one or two quarters away from Nehalem EX. All these "chasing up" if you like before the core microarchitecture update of Bulldozer.

Nehalem is touted as the biggest update in Intel's history. Well so far, at least on server space, it seems AMD's Shanghai and Istanbul compete just fine with Intel's biggest update in history. If we take into account price and power as important metrics, then AMD is winning.

This is of course only for the server markets. For enthusiast clients, and for certain workstations market where single-threaded performance is the king, Intel has a decisive lead and AMD is about 1-2 years behind. But this market is very small. For 99% of consumers/clients, things like power consumption, upgradability, and cost are more important.

[enumae-k]

...The particular example is that if one out of 4 cores are active, then 2-step clock rate increase; if two out of 4 cores are active, then 1-step clock rate increase; both for up to 1 minute....

I will post this again. Could you link to the paper you have specifically stating the "up to 1 minute" time limit?

[abinstein]

...The particular example is that if one out of 4 cores are active, then 2-step clock rate increase; if two out of 4 cores are active, then 1-step clock rate increase; both for up to 1 minute....

I will post this again. Could you link to the paper you have specifically stating the "up to 1 minute" time limit?

Can you? If you can, then do it yourself. If you cannot, then don't expect anyone else to do it for you. (You do understand what white paper I was talking about, don't you?)

In case that you still can't, I'll tell you that the 1 minute is not described in the paper (which makes it ambiguous in the first place), but measured/observed from actual performance. It is as simple as running my code on a Nehalem system for different durations. If it's less then 1 minute, the single-threaded program is accelerated. The longer it's over 1 minute the less the acceleration becomes. Apparently you are agreeing with me by action that Intel's description of Turbo mode is very ambiguous, such that without dirtying one's hands even you don't understand it's performance implications.

BTW, I want to echo what JF & keith & a few others said about Intel's Turbo mode, that while it's falsely marketed as "free" performance boost, it really is counter-productive and even harmful in a server environment. Take a look at the following excerpt from Intel's Turbo Boost white paper:

IntelTurboBoost.jpg (59.57 KiB) Viewed 1280 times

So in order to accelerate single-threaded performance, ALL CORES are dissipating more power. Does that make sense? Probably only for benchmarking.

[enumae-k]

...In case that you still can't, I'll tell you that the 1 minute is not described in the paper (which makes it ambiguous in the first place), but measured/observed from actual performance....

The point is I was looking for clarification in regards to your comment...

In the Turbo Mode white paper (at least the version that I read), Intel described only by example how Turbo Mode will be engaged. The particular example is that if one out of 4 cores are active, then 2-step clock rate increase; if two out of 4 cores are active, then 1-step clock rate increase; both for up to 1 minute. Such description is IMHO too ambiguous to be useful....

So your comment is in fact misleading people because you didn't read it, nor was it described by Intel in the Turbo white paper. Is that correct?

[JF-AMD]

So in order to accelerate single-threaded performance, ALL CORES are dissipating more power. Does that make sense? Probably only for benchmarking.

Actually, they turn it off for many benchmarks. The reason is that if you are actually getting good threading and all 4 cores are active, you are probably unlikely to get much of a boost. And running the cores hotter could potentially drive the processor to throttle down for thermal purposes.

Some benchmarks/applications take the clock speed and essentially "lock" it as a variable, so even if the actual clock rate goes up, it assumes that the "locked" rate is the rate, so you don't get to see the performance improvement.

This is just my guess, but I am thinking that some highly threaded applications may prefer a fixed clock rate.

[abinstein]

In the Turbo Mode white paper (at least the version that I read), Intel described only by example how Turbo Mode will be engaged. The particular example is that if one out of 4 cores are active, then 2-step clock rate increase; if two out of 4 cores are active, then 1-step clock rate increase; both for up to 1 minute. Such description is IMHO too ambiguous to be useful....

So your comment is in fact misleading people because you didn't read it, nor was it described by Intel in the Turbo white paper. Is that correct?

So you mean I was misleading people by telling the actual implication of Turbo mode, those that Intel should have but failed to state in the white paper? I don't think so....

Thank you for the nitpicking, though. Lets go back to the subject. Turbo mode increases thermal pressure, and in my experience a 1-minute engagement could cause CPU to throttle itself thermally. Sadly I do not have physical access to the server and cannot do what most enthusiasts sites did, to give the Nehalem baby a supreme cooling for maximum Turbo mode duration.

[dm7000s]

In the Turbo Mode white paper (at least the version that I read), Intel described only by example how Turbo Mode will be engaged. The particular example is that if one out of 4 cores are active, then 2-step clock rate increase; if two out of 4 cores are active, then 1-step clock rate increase; both for up to 1 minute. Such description is IMHO too ambiguous to be useful....

So your comment is in fact misleading people because you didn't read it, nor was it described by Intel in the Turbo white paper. Is that correct?

So you mean I was misleading people by telling the actual implication of Turbo mode, those that Intel should have but failed to state in the white paper? I don't think so....

Thank you for the nitpicking, though. Lets go back to the subject. Turbo mode increases thermal pressure, and in my experience a 1-minute engagement could cause CPU to throttle itself thermally. Sadly I do not have physical access to the server and cannot do what most enthusiasts sites did, to give the Nehalem baby a supreme cooling for maximum Turbo mode duration.

Totally irrelevant BS. I'm curious too. Where is the "up to 1 minute" stated?

[abinstein]

Whether Intel states it or not is irrelevant. It happens and it's a fact: engaging in Turbo mode after 1 minute (or probably shorter) can cause the processor to thermally throttle itself.

Actually, the problem is precisely that it was not explicitly stated by Intel.

This is BS!

[MU_Engineer]

So in order to accelerate single-threaded performance, ALL CORES are dissipating more power. Does that make sense?

Yes, it does as Intel said the Nehalem has two PLLs- one for the cores and one for the uncore. Thus, the cores all have to run at the same clock speed. Although, Intel could have gone the way of the Phenoms and had per-core PLLs and specifically clocked each core differently with Turbo Boost.

An unloaded core running at a higher clock speed does consume more power and run hotter than that unloaded core running at a lower clock speed, but the difference is not all that much compared to the thermal dissipation of a loaded core at a higher vs. lower clock speed. This is how Intel can actually increase a core's clock speed Turbo Boost and not exceed the TDP and why Turbo Boost only kicks in fully when one or two of the cores are executing instructions.

[enumae-k]

lol...

Whether Intel states it or not is irrelevant...

That was my point, so it is very relevant. You claimed to have read an Intel white paper in which Intel describes a scenario that allows Turbo Boost to only work for up to one minute. You mislead and were called out.

[abinstein]

You need to learn to quote correctly:

Whether Intel states it or not is irrelevant. It happens and it's a fact: engaging in Turbo mode after 1 minute (or probably shorter) can cause the processor to thermally throttle itself.

Cutting a sentence out of its context is not a good way of discussion. Unfortunately it is the way that you've always preferred.

Nehalem can thermal throttle itself after engaging in Turbo mode for merely 1 minute. Intel's white paper said something very ambiguous about this without giving any precise duration. OTOH, I told you the actual measured duration after which thermal throttle can happen.

Which one is misleading?

If anything, you only called out your own mental blockage when it comes to critics of Intel.

And in case you (or dm7000s) still lack the mental capacity to hold information long enough for productive discussion, I quote a previous paragraph I wrote below for your reference:

I'll tell you that the 1 minute is not described in the paper (which makes it ambiguous in the first place), but measured/observed from actual performance. It is as simple as running my code on a Nehalem system for different durations. If it's less then 1 minute, the single-threaded program is accelerated. The longer it's over 1 minute the less the acceleration becomes. Apparently you are agreeing with me by action that Intel's description of Turbo mode is very ambiguous, such that without dirtying one's hands even you don't understand it's performance implications.

[enumae-k]

I know you will not listen, but I really hope you understand that I was calling you out for having made the statement of:

In the Turbo Mode white paper (at least the version that I read), Intel described only by example how Turbo Mode will be engaged. The particular example is that if one out of 4 cores are active, then 2-step clock rate increase; if two out of 4 cores are active, then 1-step clock rate increase; both for up to 1 minute. Such description is IMHO too ambiguous to be useful....

Clearly now that you have explained, there isn't any such paper in which Intel described Turbo Boost to only work for "up to 1 minute".

[abinstein]

I know you will not listen, but I really hope you understand that I was calling you out for having made the statement of:

In the Turbo Mode white paper (at least the version that I read), Intel described only by example how Turbo Mode will be engaged. The particular example is that if one out of 4 cores are active, then 2-step clock rate increase; if two out of 4 cores are active, then 1-step clock rate increase; both for up to 1 minute. Such description is IMHO too ambiguous to be useful....

Clearly now that you have explained, there isn't any such paper in which Intel described Turbo Boost to only work for "up to 1 minute".

I know you won't/can't read, but let me quote myself for you again, which I did just two posts ago:

I'll tell you that the 1 minute is not described in the paper (which makes it ambiguous in the first place), but measured/observed from actual performance. It is as simple as running my code on a Nehalem system for different durations.

So what are you b itching about anyway?

Since you're so much interested in statement accuracy, please also tell me which of the following, stated right beside the 1-minute measurement, is true or not:

• Intel only described Turbo mode ambiguously by example
• In the example, CPU overclocks 2 steps when one core is active, 1 step when two cores are active
• CPU may downclock when Turbo mode caused increased temperature or power/current.

Simple "yes" or "no" will do. I look forward to your answers.

The 1-minute duration can be verified by anyone with access to a Nehalem server without supreme cooling. Not that I recommend anyone to spend a dime on those benchmarketing suckers, and it's Intel's fault that they didn't fully disclose the implication of Turbo mode. Clearly Intel is telling Turbo mode as yet another "blue crystal" to the fools.

So is this irrelevant to you, as long as Intel's paper does not say anything about the "1 minute," even if it is what will happen?

EDIT: Actually I can kind of understand your b itching. Being a big Intel pumper you must be very frustrated to learn the inadequacy of the Turbo mode implementation. That's why you fall into the fallacy to argue things away from the main point -- inadequacy of the Turbo mode. This is like a defense lawyer criticizing the love affair of the prosecutor on a homicide case. While I can understand it, I think it's pretty low, even for your level.

[enumae-k]

You lied about reading an Intel white paper in which Intel describes a time limit on turbo. Move on.

[keithlm]

You lied about reading an Intel white paper in which Intel describes a time limit on turbo. Move on.

POP QUIZ: Was the main point or idea in his post incorrect?

Please post something that shows that the main point or idea is not correct or desist. You aren't going to succeed at distraction so you can stop trying.

[enumae-k]

He can acknowledge a mistake or he will continue to ignore it. It makes no difference anymore.

[abinstein]

enumae-k, I don't know how to converse with you productively, because apparently you're more fond of personal attacks than the materials under discussion. Lying is an act of purposeful deception; it is not an improper wording or an erroneous expression. Yet you feel the urge to accuse others of lying when they (I) expressed opinions that you do not want to accept/agree. As I said, it is not a way of productive discussion, and I refuse to further partake in such quarrel.

On the other hand, let me ask you again about the truthfulness of the following statements, just to show that we, as well as all readers of this thread, are on the same page with respect to the implication of Intel's Turbo mode. As before, simple "yes" or "no" will do:

• Intel only described Turbo mode ambiguously by example
• In the example, CPU overclocks 2 steps when one core is active, 1 step when two cores are active
• CPU may downclock when Turbo mode caused increased temperature or power/current.

Now while you're thinking about them, let me ask you two more questions, again with simple "yes" or "no" answer:

• Is it either technically or morally acceptable for a company to describe a potentially harmful feature ambiguously to its customers?
• Can you guarantee that Turbo mode will last more than 1 minute under all proper cooling condition?

You can honestly answer these questions, or keep ignoring them and continue your personal dogfight. It makes no difference to me.

HINT: the answers in each list above are the same, but in both lists are different.

[enumae-k]

enumae-k, I don't know how to converse with you productively, because apparently you're more fond of personal attacks than the materials under discussion....

Like this:

EDIT: Actually I can kind of understand your b itching. Being a big Intel pumper you must be very frustrated to learn the inadequacy of the Turbo mode implementation. That's why you fall into the fallacy to argue things away from the main point -- inadequacy of the Turbo mode. This is like a defense lawyer criticizing the love affair of the prosecutor on a homicide case. While I can understand it, I think it's pretty low, even for your level.

I simply I called you out for your BS claim, and you know it.

In regards to your other questions, it doesn't matter how or if I answer them, you have already made up your mind.

[The_Ghost]

I know you will not listen, but I really hope you understand that I was calling you out for having made the statement of:

In the Turbo Mode white paper (at least the version that I read), Intel described only by example how Turbo Mode will be engaged. The particular example is that if one out of 4 cores are active, then 2-step clock rate increase; if two out of 4 cores are active, then 1-step clock rate increase; both for up to 1 minute. Such description is IMHO too ambiguous to be useful....

Clearly now that you have explained, there isn't any such paper in which Intel described Turbo Boost to only work for "up to 1 minute".

you are right, intel does not say that TB will last up to a minute

As an independent and complimentary feature, Intel® Hyper-Threading Technology (Intel® HT Technology) along with Intel Turbo Boost Technology increases performance of both multi-threaded and single threaded workloads. Intel Turbo Boost Technology is activated when the Operating System (OS) requests the highest processor performance state (P0).

The maximum frequency of Intel® Turbo Boost Technology is dependent on the number of active cores. The amount of time the processor spends in the Intel Turbo Boost Technology state depends on the workload and operating environment, providing the performance you need, when and where you need it.

Any of the following can set the upper limit of Intel Turbo Boost Technology on a given workload:

Number of active cores
Estimated current consumption
Estimated power consumption
Processor temperature
When the processor is operating below these limits and the user's workload demands additional performance, the processor frequency will dynamically increase by 133 MHz on short and regular intervals until the upper limit is met or the maximum possible upside for the number of active cores is reached. Conversely, when any of the limits are reached or exceeded, the processor frequency will automatically decrease by 133 MHz until the processor is again operating within its limits.
http://www.intel.com/technology/turboboost/

the problem with intel cpu's, is that they will not run at the real rated speed all the time, so they come out with this so called marketing trick and actually lower the mhz of the cpu, which was done on the P4 when it got too hot, so is this a new trick, or is intel back up too it's old tricks, and just put a new name on it?

[dm7000s]

OK, all clear now. I just wanted to be sure that I didn't misinterpret abinsteins post, which led me to believe that Intel attached some sort of guaranteed effective period to turboboost (which might be misleading).

1. They don't attach any time timetables to it at all. Good!
2. Abinstein showed that -indeed- its use may be (in his particular case) very short-lived.

On to the next issue.