AMD's Magny Cours Architecture revealed

[abinstein]

One has to wonder how does a thread about AMD's Magny Cours architecture morphed into one advertising/defending Intel's Turbo mode without some intellish fanboism. Well, one can only wonder.

Gustafson's Law is a special case of Amdahl's Law. Not the other way around.

1. There is no such thing as "Gustafson's Law." It's not a law, but merely an observation, or understanding if you will, that was made/obtained by competent computer scientists for at least 15 years: Amdahl's law does not limit throughput oriented, data parallel workloads. That simple.

2. Google again will serve you well with respect to generalizing Amdahl's law:
http://arxiv.org/pdf/cs/0209029
But this has nothing to do with the discussion above.

[abinstein]

1.All the core already running at full load, the turbo mode engage to give a peak of power;
2.The cpu temperature rise promptly up to it's set limit as trowing more electromagnetic noise and heat to the mainboard and every circuit all around;
3.Then all the cooling system have to get out all that heat from the cpu, which also touched the mainboard as well;
4.After some time, the cpu have again reach is normal temperature, the turbo mode is on again;

But every thime this cycle complete, don't the mainboard will be hotter and hotter, leading to lost computing cycle due to ECC error and system unstability.

This is a very good point. So although CPU may have thermal detector to throttle its own clock rate, but the attempts to exceed TDP will affect the whole system as the extra heat must be dissipated to the environment. I believe this is why Turbo mode is turned off in many/most IT and datacenter environments.

[kaa]

One has to wonder how does a thread about AMD's Magny Cours architecture morphed into one advertising/defending Intel's Turbo mode without some intellish fanboism. Well, one can only wonder.

...and that innuendo is The Straw.

FOAD Abinstein.

If clearing up misunderstanding and misconception is "advertising/defending" anytime it's about something you don't like, then this site's future is in danger.

[Smartidiot89]

One has to wonder how does a thread about AMD's Magny Cours architecture morphed into one advertising/defending Intel's Turbo mode without some intellish fanboism. Well, one can only wonder.

Well Turbo Mode is a good function, pretty much sums it all up and AMD should use it in Bulldozer.

[abinstein]

One has to wonder how does a thread about AMD's Magny Cours architecture morphed into one advertising/defending Intel's Turbo mode without some intellish fanboism. Well, one can only wonder.

Well Turbo Mode is a good function, pretty much sums it all up and AMD should use it in Bulldozer.

You are right. The concept of increasing the single-core clock rate while all/most other cores are idle is a good one. However, I don't feel the need to defend the particular implementation of Intel's Turbo Mode.

One big problem of Intel's Turbo Mode is its ambiguity. 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. For example, Intel rated their processor TDP using a set of artificial benchmarks that stress the cores under rated clock rate. Will Turbo Mode guarantee that this TDP is never exceeded? I bet not, otherwise why would Intel limit the period of overclocking to 1 minute only in its example?

So while I agree with you that the concept of accelerating single-core performance is good, I do not think the current Turbo Mode implementation is clean, and definitely not a clear one.

A clean implementation of such idea should take TDP restriction and software controllability into account. The latter is favorable to server/worktation environment so a system reboot is not needed to enable/disable the feature. In addition, it might also be preferable to implement better thread/process affinity to cores, so the running threads/processes won't jump across different cores excessively.

[Smartidiot89]

One has to wonder how does a thread about AMD's Magny Cours architecture morphed into one advertising/defending Intel's Turbo mode without some intellish fanboism. Well, one can only wonder.

Well Turbo Mode is a good function, pretty much sums it all up and AMD should use it in Bulldozer.

You are right. The concept of increasing the single-core clock rate while all/most other cores are idle is a good one. However, I don't feel the need to defend the particular implementation of Intel's Turbo Mode.

One big problem of Intel's Turbo Mode is its ambiguity. 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. For example, Intel rated their processor TDP using a set of artificial benchmarks that stress the cores under rated clock rate. Will Turbo Mode guarantee that this TDP is never exceeded? I bet not, otherwise why would Intel limit the period of overclocking to 1 minute only in its example?

So while I agree with you that the concept of accelerating single-core performance is good, I do not think the current Turbo Mode implementation is clean, and definitely not a clear one.

A clean implementation of such idea should take TDP restriction and software controllability into account. The latter is favorable to server/worktation environment so a system reboot is not needed to enable/disable the feature. In addition, it might also be preferable to implement better thread/process affinity to cores, so the running threads/processes won't jump across different cores excessively.

I fully agree that Intels implementation isn't a good one and is mainly only worth it for benchmarking purposes, and also the fact that Nehalem isn't exactly cool.

AMD's Overdrive has an excellent function in my opinion, while I don't use it myself (thanks to my motherboard BIOS refusing to comply to my voltage settings with AOD or k10stat) I really like the way its executed as I can get the best of both worlds: energy efficient and a overclocked state when it actually is needed/preferred.

But I do believe that in some shape or form, all future CPUs and even GPUs will have a "Turbo-Mode".

[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....

Could you link to the paper you have specifically stating the "up to 1 minute" time limit?

[BaronMatrix]

Baron,

10 years ago I was willing to pay more than 5x as much as standard for "special" Xeons when I needed them badly enough.

I'm darn sure those Xeons did not cost Intel 5x as much to make.

There is a sound business reason for AMD to adopt a similar product strategy for the "performance at all costs" crowd with folks like CB in it.

There are customers with deep pockets who will pay whatever they can justify to run their business critical apps at as high performance as they possibly can.
For many of them, a 1% increase in performance is the difference between beating their competitors or their competitors beating them.
IOW, their company survival depends on that 1%.

Of course, even better is for AMD to find out what critical features are needed in their dies to keep these customers happy and then find a way to economically incorporate them into their mainstream products.

That is not happening as long as the reaction to the CB's of the world is
"You are only 1% of the market, We don't care enough about you to try and understand your problem."

I don't think AMD officially said that to him. I think it was The Ghost. Besides, every test I see has Nehalem and Istanbul basically neck and neck on real wrld workloads like SPECjbb and LINPACK so again where are the numbers that say Nehalem is so much more powerful it's worth the extra money for HIS WORKLOADS - they can only be FP anyway, whether it's 3rd order differentials or NURBS or specular highlighting.

[maduroutmb]

What if the die was intelligent enough to see such conditions and increase the CR of the active transistors so that we could get out of this "low IPC bubble" ASAP? Under ideal conditions, we'd like to be able to goose the CR enough so that the total work done per unit time by the die remains constant.

If Turbo Boost were that fine-grained, then yes it would be awesome. At the core level, it's a lot less exciting.

[AussieFX]

That is absolutely correct. It's also why I'm not screaming crazy like certain people here are over the subject. I'd MUCH prefer AMD, but if I can't get it, I'll deal.

HOWEVER, As I stated in another post, Intel advertises on TV with ads saying the have the fastest processors on the planet. As we all know, perception is everything. We also know for a fact, that 99.9% of the buying public is not going to buy dual Xeon, Quad core 3.2ghz systems to search for porn and MP3's, and read blogs. They are more likely to go to Best Buy to get the 399.00 HP Pavillion with 19" flat screen monitor. A purchase that AMD would be a MUCH better buy for him to get, but because he saw the Intel ads, it's more likely he'll buy Intel. It's simple advertising.

AMD on the other hand has little customer recognition beyond work of mouth.

Now that 1% of extreme power users sales isn't really going to hurt AMD sales, but then again, the perception does hurt lower end sales. For a brief time, you largely went unchallenged recommending AMD for the highest end systems. Including BoXX which I pushed a lot of systems and sales for them. To try to make the same recommendations now will get you scoffed at. So at least indirectly, I say that not having the fastest around IS hurting them in overall sales to some degree. If people had the perception that AMD had the fastest processors around, I absolutely guarantee that would trickle down into sales on the low end.

Clearly with their numbers not looking that good, you can't exactly say AMD is doing just fine, there is plenty of room for improvement.

The problem AMD faces is not that they currently don't have the fastest processor, it's that they don't have a marketing budget anywhere near intels.
I have never, ever seen an AMD commercial, the closest thing to advertising I have seen from AMD is the logo on a Ferrari F1 car and it's not easy to see at 200mph.( I'm not including the internet here.)
Yet every day on television and in the newspapers I see intel advertising. This is currently an insurmountable force to deal with.
A lot of people seem to forget that when AMD were walking over intel with K8 they weren't exactly making huge profits. IIRC most of the time it was under $100 million and when you compare that to the billions intel was making it is insignificant. The performance crown helps but not having it isn't the end of the world as some are making out. (Not you btw)
Let's face it AMD have no chance to catch and pass intel until Bulldozer is released, I see no shame in that and so they should just put their heads down and do the best with what they have. In most scenarios AMD are plenty fast enough and in the few scenarios like yours where they aren't, then the answer is simple you should buy intel. But I don't want to hear about an intel cpu in your house.

There is always room for improvement whether you're fastest or not.

[kaa]

What if the die was intelligent enough to see such conditions and increase the CR of the active transistors so that we could get out of this "low IPC bubble" ASAP? Under ideal conditions, we'd like to be able to goose the CR enough so that the total work done per unit time by the die remains constant.

If Turbo Boost were that fine-grained, then yes it would be awesome. At the core level, it's a lot less exciting.

Of course. As I've said elsewhere, TB in and of itself is not that impressive

Just as P-M was a foreshadowing of C2, TB is a foreshadow of much more interesting stuff.

As I've said, everyone, not just Intel, is going to be putting increasingly more sophisticated implementations of dynamically changing CR into their dies.

I say again that I'm predicting that such a dynamic sprint mode will be in Bulldozer.

[keithlm]

ions of dynamically changing CR into their dies.

I say again that I'm predicting that such a dynamic sprint mode will be in Bulldozer.

WHY would they bother?

FIRST:

Actually I see TurboBoost as being "kind of" dishonest. If the chip can run faster and has passed testing and validation at the higher speed, then why isn't it just clocked higher? It is easy to see that a much more eloquent and cleaner design would be to go ahead clock it at the fastest it can be tested and validated at and then put in a mechanism to downclock cores independently if they are not needed. OH WAIT WE ALREADY HAVE THAT.

Thus: With the availability of dynamic underclocking the only reason for TurboBoost to exist is as a marketing gimmick. It might fool some people, and cause others to ignore the truth, but in reality it isn't that useful.


SECOND:

Something even better than TurboBoost would be to allow the system to overclock differently for various specific applications. The user should be able to specify how much to overclock and set other parameters such as voltage. OH WAIT WE ALREADY HAVE THAT. But of course since it's not in microcode or in the bios then many people don't consider that "allowed" for benchmarking. And these same people will adamantly insist that TB be allowed for benchmarking because "real system performance" is more important than clock per clock comparisons. But that same argument can used to defend TB also applies to using tools such as AOD. (Personally I think BOTH should not be used for comparative benchmarking. Comparative means we want to KNOW how they compare. Dynamic features only confuse the issue.)

THIRD:

And one of the most relevant points for SERVERS: Most experienced administrators do NOT want added complexity in their systems. Period. Especially when it only adds another point of failure. And as others have mentioned: The small amount of "the best performance at all costs" people don't make up a large enough demographic to spend time and money on.

[The_Ghost]

SME (Subject Matter Experts) and "power users" like CB can help us identify exactly what changes are needed to address the problem and increase performance by enough that AMD can start winning, rather than continuing to lose, the purchase dollars of the IT departments in CB's field.

...and =that= should be something we all want.

the problem is that this represents a small market share, amd's target is the larger market for a larger share of it

the funny thing that i remember was that at a time when amd was ahead of intel in performance, your thoughts and ideas was just the opposite, that intel was not worried or needed the top performance cpu, because that market share was so small, it is funny now to see how you have made a reverse decision on this matter

i have no problem with cb and the company that he works for needing the maximum performance cpu, the facts are, that is not where intel and amd make their money at

[The_Ghost]

That's a silly comment. Particularly out of a self-professed CPU designer.

if AMD wants to outsell Intel in CB's work environment, AMD has to execute that critical path more quickly than Intel can. It's as simple as that.

do you feel good about your self when you keep on throwing these personal attacks around?

the money is not that great for amd where cb's work enviroment calls for

[The_Ghost]

That is absolutely correct. It's also why I'm not screaming crazy like certain people here are over the subject. I'd MUCH prefer AMD, but if I can't get it, I'll deal.

HOWEVER, As I stated in another post, Intel advertises on TV with ads saying the have the fastest processors on the planet. As we all know, perception is everything. We also know for a fact, that 99.9% of the buying public is not going to buy dual Xeon, Quad core 3.2ghz systems to search for porn and MP3's, and read blogs. They are more likely to go to Best Buy to get the 399.00 HP Pavillion with 19" flat screen monitor. A purchase that AMD would be a MUCH better buy for him to get, but because he saw the Intel ads, it's more likely he'll buy Intel. It's simple advertising.

AMD on the other hand has little customer recognition beyond work of mouth.

Now that 1% of extreme power users sales isn't really going to hurt AMD sales, but then again, the perception does hurt lower end sales. For a brief time, you largely went unchallenged recommending AMD for the highest end systems. Including BoXX which I pushed a lot of systems and sales for them. To try to make the same recommendations now will get you scoffed at. So at least indirectly, I say that not having the fastest around IS hurting them in overall sales to some degree. If people had the perception that AMD had the fastest processors around, I absolutely guarantee that would trickle down into sales on the low end.

Clearly with their numbers not looking that good, you can't exactly say AMD is doing just fine, there is plenty of room for improvement.

this is what i mean, i have no problem with your posts, and needs

you have clearly stated your needs, and your company has to satisfy their needs

[The_Ghost]

...and that innuendo is The Straw.

FOAD Abinstein.

If clearing up misunderstanding and misconception is "advertising/defending" anytime it's about something you don't like, then this site's future is in danger.

yea, i think that you have gone over the top again, the straw that broke the camels back

[JF-AMD]

Here's where all of this technology is coming from.

Processor manufacturers have to deal with specifying TDPs for their processors. In order to get the max power (TDP for AMD, max power for Intel) you have to do a "worst case scenario." Essentially, how much power do you draw if every transistor is active and firing. The reality is, even in HPC, you never see all of the transistors firing. So, when you say that AMD has a TDP of 105W, you are probably more likely, in most average workloads, to see ~50-60W of total power. That is why we came up with ACP - it's a more realistic way to look at power.

Well, now people are starting to say, if there is such a gap between typical and max power, can you use that margin to squeeze a few more megahertz out of the processor. The challenge here is that while it sounds appealing (and we will be implementing a similar technology some time in the future), there are some downsides to it. This is essentially opening the processor up wider and increasing the power draw (power and clock speed are closely tied together.)

Imagine if MPG ratings for cars were all formulated at 90mph. If you were typically driving 55-60MPH you'd see much better mileage than indicated (i.e. you would not see the 105W TDP that your processors are rated at).

As a driver would you say"I like the better mileage, this is saving me more money than I expected", or would you say, wow, what great mileage, I'm going to crank the car up to 85 to get more performance."

Most people will take the gas savings, and a few lead footed folks will look for more performance. Based on where we see customers today, my guess is that most will want better power efficiency. Turbo is being marketed like it is "extra performance for free." It really is not that. If they marketed the feature as "when your processor is not consuming enough power, it kicks in to eat more," then nobody would use it.

In reality, another 1-200MHz is not going to make a tremendous difference, especially if it is fluctuating.

[AzmountAryl]

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

[Cat-Barf]

One has to wonder how does a thread about AMD's Magny Cours architecture morphed into one advertising/defending Intel's Turbo mode without some intellish fanboism. Well, one can only wonder.

I was wondering how long it would take you to revert to form and start whining fanboy......

if it's any consolation, you took a couple pages longer than I expected.

[Cat-Barf]

The problem AMD faces is not that they currently don't have the fastest processor, it's that they don't have a marketing budget anywhere near intels.
I have never, ever seen an AMD commercial, the closest thing to advertising I have seen from AMD is the logo on a Ferrari F1 car and it's not easy to see at 200mph.( I'm not including the internet here.)

AMD HAS had commercials in the past. And true, they'll never have Intel's budget. Their advertising is basically little more than word of mouth.

In a community like ours (graphics) I have seen little talk or mention of AMD in over 2 years. Professionally or from people recommending others what to buy. It's ALWAYS buy the i7 series or whatever. When AMD did have the top dog, I heard lots of AMD recommendations.

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.

[Cat-Barf]

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.

So imagine my (not really) surprise when I went to look at BoXX's "RenderBoXX" offerings to see how they were being equipped, and the ONLY options available were Intel based Nehalems. Something that BTW, in the space department, 10 motherboards take up the space of a single 4U rackspace.

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. And of course it was render farms that I've been getting flamed for.

Unlike some forum fanboys, BoXX quite well understands the market and their product line reflects that. BoXX understands it's customer base and supplies the maximum performance for demanding applications. Their product line reflects the needs of the market. A few years ago their inventory was very AMD system heavy, not so much now. Who knows, it might be again some day.

If enough people keep asking for it, maybe AMD will provide it.

[keithlm]

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

BINGO. NAIL ON THE HEAD. HE SHOOTS HE SCORES. YOU PASS GO AND COLLECT $200 Million DOLLARS.

I also do NOT want to see this garbage EVER on my production servers.

But then nowadays I'm no longer an admin... I'm just a "normal" power user. (With 8 years as a UNIX admin. 3 as a DBA for Informix. 1 for Oracle. 10+ as a developer.)

By "accident" one time I noticed some kernel parameters were not correct on a Sun Starfire box for the database. We developers were getting grief because of it. But hey I had been an admin AND a DBA. Long story short: I had previously been a high paid consultant flying around the country and installing these major database installations. I got tired of travel and got a non-travel lower paying job. So I was a "normal employee". When I noticed problems on the UNIX box I had to write a 30 page whitepaper explaining WHY the box was configured wrong. It took six months then they hired a $300.00 an hour consultant for 3 weeks who came in, read my whitepaper, and basically said to me: "You had a few things that you suggested that wouldn't matter. The rest of the things you suggested are very important and I'm basically going to repeat exactly what you told them." [six months before] Duh. Final result was a 25%-30% increase in performance for the jobs we ran. And I am talking about some jobs that took a week to run. Guess what? People REALLY notice that kind of change. "Oh it used to take 8 days.. and we now got it done in just under 6? Huzzah!"

Moral of the story: The average current system administrators can't handle (understand?) the hardware they have now. I don't want NEW questionable features added to the system that they can screw up. And make NO MISTAKE; these marketing toys that Intel is now hyping (oh that IS a bad accidental pun) are AWFUL. Either the performance "gain" because of them will be minimal OR they will hurt the final outcome in the long run. For the Turbo-Boot feature it is very simple: systems that run at a high amount of CPU use do not need more "clutter" in the hardware that won't gain them anything. (IN my brain I think: "Great an added 'feature' that some idiot is going to make me spend 3 months testing".... only to find out it has minimal effect.)

For systems that aren't that busy I have two things to say:

First: Turbo-Boost would actually work. Things WOULD be faster. But no Huzzah for this. Read on McDuff.
Second: ON these systems it won't be important whether it exists or not and in fact I do not WANT it working.

REASON: The main production systems are paramount. And even if Turbo MIGHT be good on my development machine: I do NOT want it on my prodution box if it doesn't actually have benefit. But in that case I also do not want it on my development or test boxes making me incorrectly judge how long jobs are going to take to run.

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.

Okay. Sorry for the novel. I'm having Bitburger Pills. (I used to live in Bitburg many decades ago.)

NOTE: Sad part is that most of this will just fly over the heads of people without the same experiences. Personally I always wonder: How many people who have years of experience advocate these new "features". And do NOT get me wrong: Even on SPARC and PPC there are many experienced admin who will NOT enable the SMT (hyperthreading). Ever. Not in this lifetime. Oh sure... younger admins will come in and scream "We should turn it on along with Turbo-Boost". Luckily more experience also generally means the authority to say: "No. Deal with it. Go pout in the corner." (I have basically had that happen to me: On the receiving end: years ago when I was young and thought things like performance was more important than user comfort/convenience. But that's another story. After years I started understanding. Now I can communicate the same things only often now it is ONLINE in forums instead of just person to person like it was back in the day. Lucky for these younger guys. Get slammed in a forum instead of in real life. Much easier.)

[MU_Engineer]

If Turbo Boost were that fine-grained, then yes it would be awesome. At the core level, it's a lot less exciting.

Of course. As I've said elsewhere, TB in and of itself is not that impressive[/quote]

I assume when you say "fine-grained," you are talking about being able to clock certain components of the core like FPUs at higher-than-the-rest-of-the-core frequencies? That would be a bit difficult to pull off as you would have to have a PLL for each region of the core and separate clock domains. Also, I don't know if it would really pay off much because when you have one part of the core working (such as the FPU), you are executing instructions and probably are hitting other parts of the core as well, such as I/O, ALUs, and caches and probably want to have them running fast as well. I think that having the whole core running at one speed and then clock-gating off unused portions of logic to save power is a better and easier-to-implement idea. Intel is currently using that approach in the Nehalem.

Just as P-M was a foreshadowing of C2, TB is a foreshadow of much more interesting stuff.

The Core was much more a foreshadowing of the Core 2 than the Pentium M was. The Pentium M was basically a Pentium III with more cache, SSE2, and the P4's FSB. The Core was basically the same microarchitecture as the PIII and Pentium M, but at least its macroarchitecture was much more similar to the Core 2 with a monolithic 65 nm dual-core die with a unified L2 cache, which were not present in the PIII/P-M but present in the Core 2.

As I've said, everyone, not just Intel, is going to be putting increasingly more sophisticated implementations of dynamically changing CR into their dies.

I say again that I'm predicting that such a dynamic sprint mode will be in Bulldozer.

I don't know. It may show up, or AMD might just split the difference and clock their big-die CPUs to have a guaranteed top speed somewhere between the minimum rated full-load speed of Intel's chips and the maximum rated single-thread clock speed. This would result in AMD having slightly lower performance in single-threaded tasks but higher performance in multithreaded tasks and far more predictable behavior all around (because all of the clock speeds are guaranteed.) My notebook has the precursor to Nehalem's Turbo Boost, called Intel Dynamic Acceleration, but it is fairly random in its action, short in duration, and tends to not kick on very much when the CPU has been running a while and has warmed up the heatsink. I'd personally rather have a CPU that I know always can run at 3.0 GHz on all cores rather than have some workloads run at 2.8 GHz, some at 3.0 GHz, and some at 3.2 GHz, depending on what the power-state controller, thermal diode, and ammeter are thinking at the moment.

[MU_Engineer]

Here's where all of this technology is coming from.

Processor manufacturers have to deal with specifying TDPs for their processors. In order to get the max power (TDP for AMD, max power for Intel) you have to do a "worst case scenario." Essentially, how much power do you draw if every transistor is active and firing. The reality is, even in HPC, you never see all of the transistors firing. So, when you say that AMD has a TDP of 105W, you are probably more likely, in most average workloads, to see ~50-60W of total power. That is why we came up with ACP - it's a more realistic way to look at power.

Well, now people are starting to say, if there is such a gap between typical and max power, can you use that margin to squeeze a few more megahertz out of the processor. The challenge here is that while it sounds appealing (and we will be implementing a similar technology some time in the future), there are some downsides to it. This is essentially opening the processor up wider and increasing the power draw (power and clock speed are closely tied together.)

Imagine if MPG ratings for cars were all formulated at 90mph. If you were typically driving 55-60MPH you'd see much better mileage than indicated (i.e. you would not see the 105W TDP that your processors are rated at).

As a driver would you say"I like the better mileage, this is saving me more money than I expected", or would you say, wow, what great mileage, I'm going to crank the car up to 85 to get more performance."

Most people will take the gas savings, and a few lead footed folks will look for more performance. Based on where we see customers today, my guess is that most will want better power efficiency. Turbo is being marketed like it is "extra performance for free." It really is not that. If they marketed the feature as "when your processor is not consuming enough power, it kicks in to eat more," then nobody would use it.

In reality, another 1-200MHz is not going to make a tremendous difference, especially if it is fluctuating.

Okay, so here's what I am gathering from this discussion; correct me if I am wrong.

1. Both Intel and AMD CPUs are able to stably run at speeds well above what's shipping, but the TDPs get out of hand if you validate them to run at those speeds and voltages (basically, running them at full load on all cores for a long period of time.)

2. Intel and AMD CPUs almost always run under the rated TDPs at their nominal rated speeds under most real-world workloads.

3. Intel allows some of their CPUs (E5520 and higher) to clock themselves up to the point where they actually reach their rated TDP by using Turbo Boost.

4. AMD openly admits their CPUs run under the TDP values and introduces another consumption figure, ACP, that better reflects what the CPU actually is using for power.

5. An AMD CPU's ACP and an Intel CPU with Turbo Boost's TDP are relatively indicative of how much wall power each CPU draws. The Intel CPU will have a slower nominal rated speed and fully-loaded speed but may have a higher Turbo Boost speed for running a few threads at low CPU utilization.

6. So an AMD CPU with an X watt ACP that is fully loaded on all cores will probably run faster than an Intel CPU with an X watt TDP when fully loaded on all cores but the AMD CPU will be slower when only running one or two threads.

Does that sound about right? If so, I would imagine that Turbo Boost really won't matter too much on a server as servers tend to run a lot of threads and the ability to accelerate a single thread on a 8+-core server isn't that important. It might be great on some certain desktop apps, but not very useful in a server.

[scientia]

Amdahl's Law is perfectly valid. What is NOT valid is this conclusion

Maybe you should switch to Gustafson's Law. For example:

Amdahl's Law approximately suggests:

“ Suppose a car is traveling between two cities 60 miles apart, and has already spent one hour traveling half the distance at 30 mph. No matter how fast you drive the last half, it is impossible to achieve 90 mph average before reaching the second city. Since it has already taken you 1 hour and you only have a distance of 60 miles total; going infinitely fast you would only achieve 60 mph. ”

Gustafson's Law approximately states:

“ Suppose a car has already been traveling for some time at less than 90mph. Given enough time and distance to travel, the car's average speed can always eventually reach 90mph, no matter how long or how slowly it has already traveled. For example, if the car spent one hour at 30 mph, it could achieve this by driving at 120 mph for two additional hours, or at 150 mph for an hour, and so on.


This does make Gustafson's Law a bit less pessimistic than Amdahl's Law however there are of course problems that are not parallelizable.