AMD chief say BD will offer only 35% not 50% more performance than previous gen

[Terzo]

Nope.

If the first increase is 11%, then you have to add 13% onto the result of the 11% increase, so it's actually larger. Then you do the multiplication and then add the 11%, multiply, add the 15%.

You can actually just multiply through, where a percentage increase is just 1.X (e.g. 50% increase is 1.5) and the base is 1. Probably not the greatest explanation, so example is below.

11%, 13%, and 11% increase...
1.11*1.13*1.11=1.3923, or a 39.23% increase.

Just throwing this out there in case you didn't know, though I'm assuming you described it that way to make for a simpler explanation.

 

[intangir]

:hmm:

Aren't you supposed to add the numbers? 11+13+11=35...

Heh, no, not with percentages. With percentages, the base 100% changes with every addition, so a % is worth more with each increase.

It's the same concept behind compound interest. That's how 6% a month for 12 months actually doubles your investment (201.2% to be exact). 6*12 = 72 is most definitely not 101.2! I hope my taxes didn't pay for your education.

 

[RussianSensation]

he uses an average, not one test result.

We have been over this 100x.

Phenom II is only as fast per clock as a 1st generation Core 2 Quad (even then it often isn't).

I also provided an average from Computerbase.de. It's 45-50%.

Let's do this another way:

Core 2 Q (65nm) ~ Phenom II
Core 2 Q (45nm Penryn) ~ 5% faster than C2Q
Core i7 ~ 15-20% faster than Core 2Q
Core i7 (2nd) ~ 15-20% faster than i7

Cumulative advantage taking mid-points: 1.05 * 1.175 * 1.175 = 1.45x

Also ipc numbers are useless since they change by the frequency the design is running at... This is especially the case for K8 like architecture which is bottlenecked by caches and others.

In deriving IPC, we compare processors at the same/similar clock speeds. No one sits there and unfairly compares a 3.3ghz Phenom to a 4.7ghz SB.

At stock speeds, A64's IPC advantage was mostly offset by Pentium 4's high frequency advantage. But once overclocking was taken into account on a percentage basis, it was all over for the Pentium 4.

OTOH, Phenom I/II processors did not have more overclocking headroom than Intel Core 2Q/i5/i7 chips. Inferior IPC/watt is precisely why Phenom I/II was a disappointment.

If BD is able to clock to 5.5-6.5ghz, than SB's IPC advantage would be negated by a massive clock speed advantage of BD. Otherwise, BD will need to bring at least a 15% improvement in IPC or hope that in the next 1-2 years most programs benefit from 8 threads.

I am doubtful that BD will be able to overclock much better than SB given the amount of trouble AMD already has just getting these chips to 4.0ghz.....Since SB is also an amazing overclocker, IPC/watt is literally the most important metric of comparison for modern processors (unless you need more cores for rendering, etc.). Not only that, but most programs are only coded to use 2-4 cores. In games, an i3-2120 (2C/4T) is competitive with an i5-760 (4C/4T).

Bulldozer still needs to beat a $220 2500k @ 4.6ghz, nevermind a 2600k with HT. In other words, over the next 2 years, the majority of users are going to be better off buying the fastest IPC 4-core architecture rather than a mediocre IPC 8-core CPU. By the time we'll need 8-16 core CPUs, we'll be on Haswell/Bulldozer V.3 (or whatever).

 

[Spikesoldier]

Well, looks like the only thing we can look forward to is waiting for the BD refresh if this proves to be true. Q1 2012 I recall?

Looks like a GPU only upgrade for me this fall instead of a complete overhaul.

 

[OCGuy]

Not looking good for AMD.

 

[Makaveli]

Looks like I made the right decision in jumping ship when Nehalem came out.

Everything from AMD has been a disappointment after socket 939 and it looks like the ship is still slowly sinking.

 

[thilanliyan]

In deriving IPC, we compare processors at the same/similar clock speeds. No one sits there and unfairly compares a 3.3ghz Phenom to a 4.7ghz SB.

I think what he was saying is that IPC even on the same processor can change depending on its frequency due to other bottlenecks. Correct me if I'm wrong and I really don't know whether that statement is correct, but it's just what I inferred he was saying.

 

[bryanW1995]

:hmm:

Aren't you supposed to add the numbers? 11+13+11=35...

think about it. Say that the baseline is 100%, +11% would be 1 x 1.11 = 1.11. +13% over the 1.11 would be 1.11 x 1.13 = 1.2543 of the original. Then 1.2543 x 1.11 again is = 1.392273 of the original.

Not looking good for AMD.

I believe that you are understating the situation. Not looking good would be something like a GI's perspective of the battle of Corregidor. This is more like a Japanese cabinet meeting on Aug 10, 1945.

4 years ago when barcelona came out I remember Gary Key mentioning that BD had better be good b/c it was AMD's best hope. What are they going to say this time? Is AMD's best hope now getting bought out or turning into another SCO?

 

[intangir]

I think what he was saying is that IPC even on the same processor can change depending on its frequency due to other bottlenecks. Correct me if I'm wrong and I really don't know whether that statement is correct, but it's just what I inferred he was saying.

Your interpretation and the original statement is correct. IPC is not strictly a constant for a particular microarchitecture on a particular task; you'll get a higher ratio measuring at 2.0GHz than at 3.0GHz, due to i/o speeds not scaling with core clock frequency.

Think about it. Any task where doubling clock speed does not double benchmark scores is an instance where IPC is not constant wrt clock speed. Well, assuming that the benchmark scores in question actually are on a linear scale, and twice the score means twice the performance.

 

[SHAQ]

35% is probably in special circumstances as well and will be less on the average. They always highlight the extremes to make it sound better and mislead people so they will buy it.

 

[Idontcare]

35% is probably in special circumstances as well and will be less on the average. They always highlight the extremes to make it sound better and mislead people so they will buy it.

That's the key point to many of us who simply avoid descending into these threads and arguments.

AMD has said "up to" over and over again.

That means one special case, a corner case, that they sought to find, legitimizes the statement.

We all remember AMD's "40% better than clovertown" statement...

This is why I feel like some of the marketing presence in the forums has been to everyone's detriment, the marketing guys choose their words wisely, judiciously, so as to ensure they have technical grounds to fall back on when everyone's expectations - set by their posts and marketing verbiage - falls short of reality when the product actually hits newegg.

The fact that the specific terminology was "50% more throughput", and expressly not stated as "50% more performance", was the red flag for me that marketing was doing what marketing makes a paycheck to do...as is the presence of the persistent caveat "up to" in all the continuing marketing hype.

 

[LOL_Wut_Axel]

WHAT? I got to go back and reread your llano post. What the hell. Be consistant. Your all over the place.

What are you talking about? Deneb and Thuban do not have integrated (whether powerful or not) GPUs in either the same die or the package. Llano is very different from both; it's a different market.

Nope.

If the first increase is 11%, then you have to add 13% onto the result of the 11% increase, so it's actually larger. Then you do the multiplication and then add the 11%, multiply, add the 15%.

If it's 39% it doesn't really change anything. The overall statement remains the same. Thanks for pointing it out, though.

 

[OCGuy]

I called out the admitted AMD "employee" on this forum before. Personally I think he should have to have the exact same disclaimer that Rollo had in the GPU forum.

 

[RussianSensation]

Your interpretation and the original statement is correct. IPC is not strictly a constant for a particular microarchitecture on a particular task; you'll get a higher ratio measuring at 2.0GHz than at 3.0GHz, due to i/o speeds not scaling with core clock frequency.

Think about it. Any task where doubling clock speed does not double benchmark scores is an instance where IPC is not constant wrt clock speed.

Why would you assume i/o bottlenecks exist? You should pair a fast processor with an SSD. The IPC of the processor shouldn't change by going from 2 to 4ghz if the architecture is well designed.

Umm....in theory there may be some drop off in IPC eventually for the processor as other parts of the design like IMC or cache or System Agent may become bottlenecks. But we won't be able to reach those speeds on air cooling anyway with either 1st or 2nd generation Core i series. Intel designs its chips to last 2-3 years since they understand they'll be increasing frequencies on the same design with each Tick. As such the possibility of IPC being bottlenecked with higher frequencies is practically nil. For instance, Core i7 920 2.66ghz scales without problems if you compare its single threaded performance to the Core i7 980X 4.4ghz.


========================================================================
Look at the benches I posted from Xbitlabs of Q6600 vs. Phenom II 940. You can see that they are very close in IPC at 3.6-3.8ghz range.

Now look at Anandtech's review of Sandy Bridge. Go through the pages and you'll see that a 2500k 3.3ghz is about 2x faster than Q6600 2.4ghz. Of course 2500k gets 1 Turbo bin when using 4 cores.

Let's continue. The only way for 2500k 3.4ghz (with 4 cores pegged) to be nearly 2x faster than a Q6600 2.4ghz (4 cores pegged) is if Sandy Bridge is 1.5x in IPC compared to C2Q (which is exactly what I have been saying regarding Phenom II).

Math = 2500k @ 3.4ghz x 1.5 IPC / 2.4ghz Q6600 = 2.1x. Q.E.D.


==========================================================================
Now, regarding your theory that IPC changes with frequency. Do you have any evidence to support the view that this is so with either Core i7 or 2nd generation Sandy Bridge designs?

Here is my support that your theory doesn't hold true:

SB enjoys excellent scaling in performance with higher clockspeeds.

Games
Starcraft II (one of the best CPU benches for gaming)

2500k 3.3g (min 3.4g Turbo) = 51 fps
2500k 4.7g = 69 fps (+35&#37
Frequency increase = +38% (4.7ghz / 3.4ghz) <-- almost linear scaling

Rendering

2500k 3.4g (Turbo 4 cores pegged) = 5.47
2500k 4.9g = 7.71 (+41%)
Frequency Increase = +44% (4.9 / 3.4ghz) <-- almost linear scaling

While you can argue that in games you'll be GPU limited most of the time, the IPC advantage of SB isn't going anywhere, even at 4.9ghz overclocks. If anything, from the Xbitlabs link right above, it's the Phenom II 1100T that barely scales with higher clock speed.

Therefore, it looks to me like we have known data that SB clocks high and its IPC remains intact due to a robust architectural design. As a result, I do not see how you can dismiss IPC in this comparison.

 

[dennilfloss]

At this point I'll be very surprised if this doesn't turn out to be Phenom launch v2.0. Not looking good at all. Intel will soon have a gun to our collective heads, if they don't already.

Not my head. I'd be quite happy with a 35% improvement over my nonoverclocked PH II 955 if the price is reasonable.

 

[bryanW1995]

I called out the admitted AMD "employee" on this forum before. Personally I think he should have to have the exact same disclaimer that Rollo had in the GPU forum.

Why, is this AMD employee consistently trolling everyone under the sun who disagrees with him, and even trying to get longtime members in good standing banned? If so, then I wholeheartedly agree!

Oh, and he needs to be banned, then allowed back b/c the AMD marketing dept asked Anand nicely.

 

[Abwx]

Now look at Anandtech's review of Sandy Bridge. Go through the pages and you'll see that a 2500k 3.3ghz is about 2x faster than Q6600 2.4ghz. Of course 2500k gets 1 Turbo bin when using 4 cores.

Let's continue. The only way for 2500k 3.4ghz (with 4 cores pegged) to be nearly 2x faster than a Q6600 2.4ghz (4 cores pegged) is if Sandy Bridge is 1.5x in IPC compared to C2Q (which is exactly what I have been saying regarding Phenom II).

Math = 2500k @ 3.4ghz x 1.5 IPC / 2.4ghz Q6600 = 2.1x. Q.E.D.

.

Really, you should definitly give up any basic maths comparison.

If a 2.4g Q6600 is two times slower than a I2500K , then
the IPC improvement is 2 x (2.4 / 3.4 ) = 1.41

That is, 41%, assuming the 2500K has not a turbo enabled,
otherwise the difference would be less...

 

[Riek]

I think what he was saying is that IPC even on the same processor can change depending on its frequency due to other bottlenecks. Correct me if I'm wrong and I really don't know whether that statement is correct, but it's just what I inferred he was saying.

Correct. That is indeed what i mean.

Why would you assume i/o bottlenecks exist? You should pair a fast processor with an SSD. The IPC of the processor shouldn't change by going from 2 to 4ghz if the architecture is well designed.

Umm....in theory there may be some drop off in IPC eventually for the processor as other parts of the design like IMC or cache or System Agent may become bottlenecks. But we won't be able to reach those speeds on air cooling anyway with either 1st or 2nd generation Core i series. Intel designs its chips to last 2-3 years since they understand they'll be increasing frequencies on the same design with each Tick. As such the possibility of IPC being bottlenecked with higher frequencies is practically nil. For instance, Core i7 920 2.66ghz scales without problems if you compare its single threaded performance to the Core i7 980X 4.4ghz.

Look at the benches I posted from Xbitlabs of Q6600 vs. Phenom II 940. You can see that they are very close in IPC at 3.6-3.8ghz range.

Now look at Anandtech's review of Sandy Bridge. Go through the pages and you'll see that a 2500k 3.3ghz is about 2x faster than Q6600 2.4ghz. Of course 2500k gets 1 Turbo bin when using 4 cores.

Let's continue. The only way for 2500k 3.4ghz (with 4 cores pegged) to be nearly 2x faster than a Q6600 2.4ghz (4 cores pegged) is if Sandy Bridge is 1.5x in IPC compared to C2Q (which is exactly what I have been saying regarding Phenom II).

Math = 2500k @ 3.4ghz x 1.5 IPC / 2.4ghz Q6600 = 2.1x. Q.E.D.

Now, regarding your theory that IPC changes with frequency. Sandy Bridge enjoys excellent scaling in performance with higher clockspeeds. Look at Starcraft II (one of the best CPU benches for gaming).

2500k 3.3g (min 3.4g Turbo) = 51 fps
2500k 4.7g = 69 fps (+35&#37
Frequency increase = +38% (4.7g vs. 3.4g)

While you can argue that in games you'll be GPU limited most of the time, the IPC advantage of SB isn't going anywhere, even at 4.7ghz overclocks. If anything, from the link right above, it's the Phenom II 1100T that barely scales with higher clock speed. Therefore, we have known data that SB clocks high and its IPC remains intact due to a robust architectural design.

Just because you can find 1 or more tasks where ipc doesn't differ alot doesn't mean it isn't there.

While you increase your frequency of the cpu you're ram stays the same. ergo: the latency to your ram doesnt change alot and neither does the bandwidth. This means that more cycles are wasted when you have to wait on this.

In case of phenom they have a fix frequency for their level 3cache. Meaning that while the cpu frequency scales, relative to the cpu speed the cache becomes slower. (this is the same for the latency towards the memory)


If you look at most reviews a 10% clockspeed increase will give on average a 7% performance increase.

Just took some results at 3GHz and 3,7GHz. This is a 23% clockspeed difference and the performance increase is on average (took 6non gaming tests) 18%. So ipc dropped. (for gaming the result was more in line with 10% on average)

 

[sangyup81]

In case of phenom they have a fix frequency for their level 3cache. Meaning that while the cpu frequency scales, relative to the cpu speed the cache becomes slower. (this is the same for the latency towards the memory)

This is why we phenom overclockers insist on overclocking the CPU-NB. It has the pleasant side effect of increasing the speed of the L3. At least that's the theory behind it.

 

[Idontcare]

Why, is this AMD employee consistently trolling everyone under the sun who disagrees with him, and even trying to get longtime members in good standing banned? If so, then I wholeheartedly agree!

Oh, and he needs to be banned, then allowed back b/c the AMD marketing dept asked Anand nicely.

Anand allowed rollo back not because NV marketing asked him but because there was, and still is, a standing open-invitation to all banned members to petition the moderator establishment in the Mod Disc forum to request leniency and forgiveness to be allowed back in.

Perma-ban simply makes little sense in terms of the decades that can pass in one's life while being banned. Its not really meant to be a life sentence, just a ban until you get your act together which for some people really can take a lifetime.

For Rollo the issue is that he was allowed back too soon, it is probationary after all, and thus he was booted from the forums once again.

Whether we want to admit it or not, people with passion are what we like to see inaction.

 

[JFAMD]

That's the key point to many of us who simply avoid descending into these threads and arguments.

AMD has said "up to" over and over again.

That means one special case, a corner case, that they sought to find, legitimizes the statement.

We all remember AMD's "40% better than clovertown" statement...

This is why I feel like some of the marketing presence in the forums has been to everyone's detriment, the marketing guys choose their words wisely, judiciously, so as to ensure they have technical grounds to fall back on when everyone's expectations - set by their posts and marketing verbiage - falls short of reality when the product actually hits newegg.

The fact that the specific terminology was "50% more throughput", and expressly not stated as "50% more performance", was the red flag for me that marketing was doing what marketing makes a paycheck to do...as is the presence of the persistent caveat "up to" in all the continuing marketing hype.

I have addressed the 40% claim several times. In memory stream performance we were ~40% faster. In 4P performance, we were 40%+ faster. Where that got wrapped around the axle was people taking statements out of context and applying them to client parts, it was a server statement.

Server performance = throughput; client performance = speed. Different criteria that cannot be correlated because the workloads and environments are different.

 

[RussianSensation]

Really, you should definitly give up any basic maths comparison.

Did you even bother checking the links I provided and noted that I was estimating? You want exact math?

Cinebench R10: Single threaded

2500k = 5,860
Q6600 = 2,778

2500k is 2.11x faster. I simply rounded the figure to ~ 2x for simplicity sake and for approximation in my previous post.

3.4 ghz * (1+X) = 2.4ghz * (5,860 &#247; 2,778)

X = 48.9&#37;

Still don't believe me?

Computerbase.de did an extensive test. All the data is right in front of you.

2500k @ 2.8ghz = 171%
Q6600 @ 2.4ghz = 100%

Adjusting SB's clock speeds to 2.4ghz:

2.8 ghz 1.71
------- = -----
2.4 ghz Y

Y = 46.6% relative to a Q6600 2.4ghz

SB is 45-50% faster in IPC than C2Q/Phenom II.

snip

What does your discussion of RAM bandwidth/latency and their dependency in a Phenom architecture have anything to do with SB's IPC advantage over Phenom II? You are trying to negative the important of IPC in how it relates to Phenom II's bottlenecks. Yet, SB is not really affected by either latency or memory bandwidth bottlenecks in consumer applications. There are plenty of tests that show this on the Internet.

Therefore, SB's 40-50% IPC advantage over Phenom II remains even in overclocked states. As such, Bulldozer has to be compared to a 2500k processor running at 4.5ghz+ for us overclockers. If Bulldozer has identical IPC to Phenom II, what do you think is going to happen?

My personal view is that there are 4 main ways for Bulldozer to be competitive (And of course some combination of these factors):

1) Reliance on heavily multi-threaded benchmarks to show a marked improvement over SB.
--> Throwing more cores sounds like AMD's niche strategy to gain market share by targeting heavy multi-taskers. I admit this will be beneficial in the workstation space, server space and professional space. Alternatively, it's a great marketing strategy for the average Joe like the Mhz myth was. Still, most consumer programs don't benefit from more than 4 threads. Intel also has Quick Sync, which takes care of the market's video transcoding needs.

2) Much higher overclocking headroom than SB
--> Unlikely since Intel's CPUs overclock to 4.6ghz+
--> FX-8130 needed 1.5V on 32nm to get a 4.63ghz overclock in this thread. That's a very high voltage for 24/7 operation on 32nm. Doesn't look to me like BD is going to be hitting 5.5-6.0ghz any time soon. Obviously, in stock form, if BD ships with 4.0ghz clocks, then it will be competitive with a 3.3ghz SB. But ultimately, us enthusiasts do want to look at top-end performance in overclocked states.

3) Much higher IPC than Phenom II
--> IMO, the most important variable which will ultimately decide how good BD's performance will be. Ironically, the variable for which we have the least amount of information from AMD. Hmm....

4) Price

Remember when AMD owned Intel in performance?

--> If we are to believe that AMD is going to give us an 8 core processor for $300 that mops the floor with SB, would AMD be giving their highest SKU for $300? AMD is not a charity.

 

[LOL_Wut_Axel]

Did you even bother checking the links I provided and noted that I was estimating? You want exact math?

Cinebench R10: Single threaded

2500k = 5,860
Q6600 = 2,778

2500k is 2.11x faster. I simply rounded the figure to ~ 2x for simplicity sake and for approximation in my previous post.

3.4 ghz * (1+X) 2.11
--------------- = ------
2.4ghz 1

X = 48.9%

Don't believe me? Computerbase.de did an extensive test. All the data is right in front of you.

2500k @ 2.8ghz = 171%
Q6600 @ 2.4ghz = 100%

Adjusting SB's clock speeds to 2.4ghz:

2.8 ghz 1.71
------- = -----
2.4 ghz Y

Y = 46.6% relative to a Q6600 2.4ghz

You can sit here all day and argue maths, but SB is 45-50% faster in IPC than C2Q.

No. You're inflating the numbers by focusing on one or two benchmarks. You need to use overall performance in all the CPU benchmarks. Sandy Bridge has 30% higher IPC than Core 2 45nm and 40% higher IPC than Core 2 65nm and K10.5.

Again, a compilation of most CPU benchmarks:

 

[Idontcare]

Again, a compilation of most CPU benchmarks:

That is an excellent graph, summarizes the strengths and weaknesses of the architectures quite nicely.

Would be nice to see where Llano falls in this.

 

[RussianSensation]

No. You're inflating the numbers by focusing on one or two benchmarks.

I showed Cinebench for a quick Single threaded IPC comparison.

Then I showed a 46.6&#37; average from Computerbase, which averaged 22 benchmarks:

3DMark 11
Cinebench 11.5
SiSoft Sandra (CPU + AES)
SiSoft Sandra (Mem + AVX)
WinRAR 4.0 (integrated)
Autodesk 3ds Max 2011
MainConcept H.264/AVC Pro
dBpoweramp R14
TrueCrypt 7.0a
SPECjvm2008
Paint.NET
WinRAR 4.0 (real pack)
x264 HD Benchmark 3:19
PCMark Vantage
Jurassic Park - Gothic 4
ArmA 2: Operation Arrowhead
Call of Duty: Modern Warfare
Civilization V
F1 2010
Medal of Honor
Resident Evil 5
Two Worlds

That is an excellent graph, summarizes the strengths and weaknesses of the architectures quite nicely.

The benchmark provided only focuses on timed tests and tells us nothing about gaming performance or non-timed tests.