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

[intangir]

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.

Um, I'm not assuming. I'm telling you what theory predicts, what the data shows, and what decades of experience with computing has taught me.

And by I/O, I mean any communication that the CPU does, including with RAM. An SSD can be considered just an added layer of the memory hierarchy.

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.

It doesn't just happen in theory! The bottlenecks are observable at clock speeds achievable now. As mentioned, overclocking a Phenom II does not necessarily increase L3 cache speed, and this has demonstrable damping effects on performance gains. And while this problem doesn't occur with Sandy Bridge (L3 cache runs at core clock frequency for that architecture), the BCLK and QPI speeds still do not necessarily increase.

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, using the exact same Xbitlabs data you used:

Performance per clock is not quite proportionate to IPC, as IPC is normalized per number of active cores, but the trend behavior for each individual processor should be intact.

You will notice that every line is downwardly sloped. If IPC were constant versus clock, the lines would be flat. If IPC went up with increasing clock, they'd slope up. And if there were no correlation, they'd look more scattershot. But no, all the lines slope DOWN. What do you think that means?

SB enjoys excellent scaling in performance with higher clockspeeds.

I am not denying that, and Sandy Bridge is indeed an amazing CPU, but it is not immune to the laws of computation. "Excellent" scaling is not perfect scaling. And to assume IPC to be constant with clock speed is to assume that clock scaling is perfect.

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.

I am hardly dismissing IPC. I am simply saying that IPC does vary with clocks, and sometimes quite widely, so this effect must be taken into account. Look at the i5-760 numbers for Cinebench. Perf/cycle drops from 1.4 to 1.2 going from 2.8 to 4.1 GHz. That's 15%! Obviously this has to be considered if we are talking about 5-10% differences in performance.

Scaling of performance with clock speed is actually another property of processor microarchitectures, and can vary enough between CPUs so that what the performance rankings are at 3 GHz are not necessarily what they are at 4 GHz. As an example, observe the Cinebench numbers where the lines of the i7-950 and the X6 1100T cross!

 

[LOL_Wut_Axel]

I showed Cinebench for a quick Single threaded IPC comparison.

Then I showed a 46.6% average from Computerbase, which averaged 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
Games (640x480)
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
Games (1680x1050)
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

Not being rude, but that's not a good way to do architecture comparisons, which is what we're doing. You use all the single-threaded applications you can to compare CPU architectures, not multi-threaded ones. For example, the fact that some Sandy Bridge CPUs support HyperThreading skews the results completely since it's not part of the IPC. And, again, games is one of the worst ways to do architecture comparisons. There's the GPU factor to take into account, which also skews the results.

The table I posted above is taken from the time it took all the CPU architectures listed to complete the tests, at the same clock speed, and all with one core and one thread enabled. This is the most accurate way to represent architectural differences.

 

[RussianSensation]

Not being rude, but that's not a good way to do architecture comparisons, which is what we're doing.

You use all the single-threaded applications you can to compare CPU architectures, not multi-threaded ones.

So we can't compare a 2500k 3.3ghz 4C/4T CPU to a Phenom II X4 @ 3.3ghz 4C/4T? That's news to me. Also, I am suddenly being biased by using a wide variety of benchmarks as opposed to limiting my findings to only timed tests?

Alright, let's just wait until September 19th. I am not going to try to prove anymore that SB is significantly faster than 35% in IPC than Phenom II is.

If your theory proves true, around September 19th AMD will deliver a 4.2ghz Turbo 8-core FX-8150. Adjusting for a 35% IPC disadvantage (using your insistent estimate) and assuming BD at least maintains identical IPC to Phenom II, we'll have an equivalent of a 3.1ghz 8-core SB processor....for $300.

 

[LOL_Wut_Axel]

So now using a wide variety of benchmarks as opposed to limiting your findings to only timed tests is not a good way?

Alright, let's wait until September 19th. I am not going to try to prove anymore that SB is significantly faster than 35% in IPC than Phenom II is.

If your theory proves true, on September 19th AMD will deliver a 4.2ghz Turbo 8-core FX-8150. Adjusting for a 35% IPC disadvantage (using your insistent estimate) and assuming BD at least maintains identical IPC to Phenom II, we'll have an equivalent of a 3.1ghz 8-core SB processor....for $300. :whiste:

You're putting words into my mouth. Sandy Bridge has HyperThreading. That gives it a ~20% advantage in multi-threaded applications. I'm not using the turbo to compare since it's dependant on other external factors. Using base clock speeds only and assuming the same IPC as Nehalem, the FX-8150 would be just slightly faster than a Core i7 990X overall. It would have around 11% lower single-threaded performance, and 25-30% higher multi-threaded performance than the i7 2600. Your 8-core SB figure you pulled out of nowhere. And as has been mentioned previously, Sandy Bridge has 39% higher IPC than K10.5.

Also, if you're comparing architectures directly you need to use one thread for both at the same clock speed. That's as accurate as you can get to comparing them. Again, you do not use multi-threaded benchmarks to directly compare architectures. It introduces too many factors into the equation.

Another thing some people seem to forget is that having higher IPC does not equal having the faster CPU.

 

[RussianSensation]

I'm telling you what theory predicts, what the data shows, and what decades of experience with computing has taught me.

It doesn't just happen in theory! The bottlenecks are observable at clock speeds achievable now.

Of course there are always other bottlenecks that exist which ensure that you don't get perfectly linear IPC scaling - Hard drive, GPU bottlenecking, perfect multi-threading in the code of the applications, Windows 7 allocation of threads, etc. etc. But you picking 1 or 2 benchmarks does not at all disprove that SB doesn't have almost linear performance scaling with an increase in clock speeds (assuming no other bottlenecks exist).

I can pick a benchmark that shows you that Core i7 has 0 scaling with frequency:

OR

I can pick a benchmark and shows a 93% scaling with a 100% increase in clock speed on the same processor:

If you want to discuss how a GPU is the main bottleneck at 2560x1600 8AA, or that an mechanical drive is the main bottleneck when doing a writing and reading application simultaneously, I won't disgree. It still stands that SB has a 45-50% IPC advantage over Phenom II, and that advantage doesn't suddenly evaporate at 4.7ghz. As such, BD will be competing against that advantage.

 

[LOL_Wut_Axel]

So we can't compare a 2500k 3.3ghz 4C/4T CPU to a Phenom II X4 @ 3.3ghz 4C/4T? That's news to me. Also, I am suddenly being biased by using a wide variety of benchmarks as opposed to limiting my findings to only timed tests?

You could, but it's not the most ideal situation. If you want to compare IPC you use single-threaded applications or make the application use a single thread.

Also, no one claimed you were being biased. I just claimed you were doing incorrect comparisons given what we're discussing.

 

[RussianSensation]

Another thing some people seem to forget is that having higher IPC does not equal having the faster CPU.

A 2C/HT Core i3-2120 3.3ghz has no problems beating a full fledged 4C Phenom II 955 3.2ghz in every single gaming benchmark in this review. How did it do that? Most games don't use 4 threads. Core i3 has a massive IPC advantage at the same clocks. If BD doesn't improve on Phenom II's IPC, it's game over against a 4.5ghz 2500k in the same benchmarks. You still stand firm that IPC doesn't matter?

 

[LOL_Wut_Axel]

Of course there are always other bottlenecks that exist which ensure that you don't get perfectly linear IPC scaling - Hard drive, GPU bottlenecking, perfect multi-threading in the code of the applications, Windows 7 allocation of threads, etc. etc. But you picking 1 or 2 benchmarks does not at all disprove that SB doesn't have almost linear performance scaling with an increase in clock speeds (assuming no other bottlenecks exist).

I can pick a benchmark and shows you that Core i7 has 0 scaling with frequency:

OR

I can pick a benchmark and shows a 93% scaling with a 100% increase in clock speed on the same processor:

It still stands that SB has a 45-50% IPC advantage over Phenom II, and that advantage remains intact at 2.0ghz, 3.0ghz, 4.0ghz or 5.0ghz.

Proof? This says otherwise, and is much more accurate for comparing architectures:

And for the umpteenth time, using games to compare architectures is less than ideal. One, most new games support some form of multi-threading; two, they introduce the GPU into the equation. Both things can skew the results. If you want to compare CPU architectures, use CPU-only benchmarks.

Also, what I picked is not one or two benchmarks. This table is based on the results for 18 benchmarks, 15 of which were completely real-world and focused on the CPU only.

 

[LOL_Wut_Axel]

A 2C/HT Core i3-2120 3.3ghz has no problems beating a full fledged 4C Phenom II 955 3.2ghz in every single gaming benchmark in this review. How did it do that? Most games don't use 4 threads. Core i3 has a massive IPC advantage at the same clocks.

You do know gaming isn't the only thing people do with their computers, right? There's other benchmarks where the Phenom II X4 955 beats it.

 

[happy medium]

I'm guessing what all these numbers mean is that theres really no way a BD's gonna beat a 2500k @ 4.5 in anything really, even overclocked. Is this safe to assume?

Can us q9xx owners hit the buy button on a 1155 system now without worry that some how BD 's gonna pull off a miracle?

 

[intangir]

Of course there are always other bottlenecks that exist which ensure that you don't get perfectly linear IPC scaling - Hard drive, GPU bottlenecking, perfect multi-threading in the code of the applications, Windows 7 allocation of threads, etc. etc.

We are in violent agreement here.

But you picking 1 or 2 benchmarks does not at all disprove that SB doesn't have almost linear performance scaling with an increase in clock speeds (assuming no other bottlenecks exist).

I did not pick the benchmarks; you did. My point is that bottlenecks always exist. Always. Except in synthetic benchmarks that do not reflect a real-life workload. And I wish to highlight that you said "almost" linear and not linear!

I can pick a benchmark that shows you that Core i7 has 0 scaling with frequency:

OR

I can pick a benchmark and shows a 93% scaling with a 100% increase in clock speed on the same processor:

If you're trying to disprove my assertion that perfect performance/clock scaling is impossible, you would do better by, y'know, actually posting benchmarks that demonstrate perfect scaling. Both of these benchmarks are actually reinforcing my point that IPC goes down as clock speed goes up.

 

[happy medium]

In civ 5 27 to 52 is pretty darn close to 100%, well within the margin of error.

 

[LOL_Wut_Axel]

I'm guessing what all these numbers mean is that theres really no way a BD's gonna beat a 2500k @ 4.5 in anything really, even overclocked. Is this safe to assume?

We don't know. We have no idea of what the IPC is as of now.

 

[jvroig]

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 ÷ 2,778)

X = 48.9%

Still don't believe me?

Correct me if I'm wrong, but the single-thread turbo of 2500K is 3.7, yes?

Using that as the real figure from your calculation, the result would be 36.83%, not 48.9.

2500K = 5,860 pts / 3.7GHz = 1583.783 pts/GHz
Q6600 = 2,778 pts / 2.4GHz = 1,157.5 pts/GHz

1583.783 / 1,157.5 = 1.368

So a Q6600 needs a 36.8% IPC boost to reach the IPC of the 2500K. That's a bit far from your 48.9%

As for whether the benchmark is valid or representative of real-world / gaming / whatever performance, no comment, I don't really care. Just really wanted to point out the mathematical error there that I perceived since you are using it as some sort of proof for your argument with LOL_Wut_Axel.

 

[RussianSensation]

I'm guessing what all these numbers mean is that theres really no way a BD's gonna beat a 2500k @ 4.5 in anything really, even overclocked. Is this safe to assume?

Can us q9xx owners hit the buy button without worry that some how BD 's gonna pull a miracael?

AMD is positioning an 8-core CPU against a 2600k (4C/HT). HT adds about a 20% boost. It most likely means that each BD core isn't as fast as a single SB core. BD should beat the 2500k in multi-threaded apps (8C vs. 4C). However, whether or not it will beat SB in most scenarios is uncertain.

What do you think? After the failure that Phenom I/II were, AMD is going to deliver a nearly 4.0ghz 8-core processor with similar IPC to SB and similar power consumption and similar overclocking headroom -- all that for $300? All that with a fraction of Intel's R&D?

So basically, in 2 months from now, we'll all be able to buy an equivalent to an 8-core 2500k @ 4.0ghz, despite 5 years of non-competitiveness, no single AMD benchmark leak, no official explanation for 3x or so they delayed BD? Oh I know, I know exactly where I heard this before (the perfect alignment of all the stars in the universe with no disadvantages) -- like that non-sense claim that HD6970 was going to be the next "R300" and beat GTX580 by 40-50% in DX11. ^_^

Either way, loyal AMD fans have done a great job since January in promoting the 2nd coming of Jesus (I mean BD), along the way dismissing architectural advancements in Core i (1st) gen and now Core i (2nd gen) over the last 3 years (and still failing to acknowledge that Phenom II was only as fast as Core 2 Quad released by Intel in 2007). Funny enough, some of them complained about the power consumption disadvantage of Fermi vs. Cypress, yet conveniently dismissed the horrendous power consumption of overclocked Phenom II X4 and X6 processors in their own systems.

During this time, they present claims that CPUs don't matter for games, file compression, photoshop, sysmark (general office performance) because all of these benchmarks are either heavily compiled to 'favor' Intel OR that you'll be limited by some other components anyway (i/o, memory speed, hard drive, gpu, etc. etc.).

At the same time, they await BD as if suddenly its performance advantage over SB will matter. They also claim that Phenom II was good enough against C2Q and Core i7 all this time, but suddenly upgrading to a BD is definitely worth it.

If we use their same logic, then it would make 0 difference if one purchased a SB processor today or waited for BD since you'll still be limited by your mechanical drive for office tasks OR you'll still be limited by your GPU in games. Ironically, they haven't done a great job of actually promoting why BD will be a game changer.

 

[RussianSensation]

If you're trying to disprove my assertion that perfect performance/clock scaling is impossible, you would do better by, y'know, actually posting benchmarks that demonstrate perfect scaling. Both of these benchmarks are actually reinforcing my point that IPC goes down as clock speed goes up.

I don't recall ever stating that there is a perfectly linear scaling with increased frequency. I said it was 'excellent' or 'nearly linear' for SB and that it will still be there at higher clock speeds against Phenom II. As such, I believe that SB's IPC advantage will also matter against BD, which is the main argument at hand in this thread.

Correct me if I'm wrong, but the single-thread turbo of 2500K is 3.7, yes?
So a Q6600 needs a 36.8% IPC boost to reach the IPC of the 2500K. That's a bit far from your 48.9%

:thumbsup: 2500k does Turbo to 3.7ghz when only 1 core is used. However, take a look at what happens in the Multi-threaded Cinebench R10:

Q6600 2.4 = 9,681
2500k 3.3 (3.4) = 20,381 (the same 2.11x increase). <-- 49% difference then.

So I don't think in that single-threaded bench the 2500k was running at 3.7ghz.

Here are more benchmarks:

Source
2500k is 54% faster than the X4 970 at similar clock speeds.

2500k is 55% faster than the X4 970.

2500k is 68% faster than the X4 970.

2500k is 44% faster than the X4 970.

2500k is 38% faster than the X4 970.

There are scenarios such as WinZip where 2500k is only 22% faster, for instance.

However, based on a multitude of benchmarks I provided from Anandtech, Hardware Secrets, Xbitlabs, Computerbase, SB's advantage in IPC over PhII is closer to 45-50%, not 35%.

 

[Idontcare]

Q6600 2.4 = 9,681
2500k 3.3 (3.4) = 20,381 (the same 2.11x increase). <-- 47% difference then.

I'm not following the conversation all that closely here, so forgive me if I am about to state something foolish, but didn't you forget to clock-normalize those performance numbers there?

IPC = Instructions/clock

 

[AtenRa]

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.

Have a look @ this Llano A8-3850 vs Phenom II vs Athlon II comparison

You will find that Llano has the same or better performance with Phenom II

 

[jvroig]

So I don't think in that single-threaded bench the 2500k was running at 3.7ghz.

No reason it wouldn't be. We can't choose what facts to believe and not to believe depending on what suits us better. Either we believe in facts for credibility, or not.

Like I said, I don't really care about the benchmarks. Just corrected the math because you were offering it as proof to suit your argument (an argument which may be correct or incorrect, but about which I don't really care to involve myself), and from a cursory glance it looked wrong to me. That is all. I'm not interested in your argument with LOL_Wut_Axel, just the math in that particular "proof".

 

[RussianSensation]

I'm not following the conversation all that closely here, so forgive me if I am about to state something foolish, but didn't you forget to clock-normalize those performance numbers there?

IPC = Instructions/clock

3.4ghz 4C was 2.11x faster than a 2.4ghz 4C CPU.

Then,

3.4ghz * (1 + IPC advantage) = 2.11x faster * 2.4ghz

IPC advantage = 48.9&#37;. So I did normalize it. Not sure why I put 47% rather than 49% (time to sleep).

No reason it wouldn't be. We can't choose what facts to believe and not to believe depending on what suits us better.

Considering there is an identical 2.11x performance advantage that 2500k enjoys in both Single and Multi-threaded bench, how can you have it working at 3.7ghz in 1 case and only at 3.4ghz in another?

 

[Idontcare]

3.4ghz 4C was 2.11x faster than a 2.4ghz 4C CPU.

Then,

3.4ghz * (1 + IPC advantage) = 2.11x faster * 2.4ghz

IPC advantage = 48.9%. So I did normalize it. Not sure why I put 47% rather than 49% (time to sleep).

Ah, explained perfectly, thanks for indulging my absentmindedness.

That is a rather impressive IPC gain in such a short timespan, a mere 4-5 yrs.

And here I thought everyone was trying to convince me that single-threaded IPC scaling was dead.

 

[Joseph F]

I can't wait for BD and SB-E to come out so I don't have to hear all of this arguing and speculation anymore.

 

[Nemesis 1]

Um, I'm not assuming. I'm telling you what theory predicts, what the data shows, and what decades of experience with computing has taught me.

And by I/O, I mean any communication that the CPU does, including with RAM. An SSD can be considered just an added layer of the memory hierarchy.



It doesn't just happen in theory! The bottlenecks are observable at clock speeds achievable now. As mentioned, overclocking a Phenom II does not necessarily increase L3 cache speed, and this has demonstrable damping effects on performance gains. And while this problem doesn't occur with Sandy Bridge (L3 cache runs at core clock frequency for that architecture), the BCLK and QPI speeds still do not necessarily increase.



Here, using the exact same Xbitlabs data you used:

Performance per clock is not quite proportionate to IPC, as IPC is normalized per number of active cores, but the trend behavior for each individual processor should be intact.

You will notice that every line is downwardly sloped. If IPC were constant versus clock, the lines would be flat. If IPC went up with increasing clock, they'd slope up. And if there were no correlation, they'd look more scattershot. But no, all the lines slope DOWN. What do you think that means?



I am not denying that, and Sandy Bridge is indeed an amazing CPU, but it is not immune to the laws of computation. "Excellent" scaling is not perfect scaling. And to assume IPC to be constant with clock speed is to assume that clock scaling is perfect.



I am hardly dismissing IPC. I am simply saying that IPC does vary with clocks, and sometimes quite widely, so this effect must be taken into account. Look at the i5-760 numbers for Cinebench. Perf/cycle drops from 1.4 to 1.2 going from 2.8 to 4.1 GHz. That's 15%! Obviously this has to be considered if we are talking about 5-10% differences in performance.

Scaling of performance with clock speed is actually another property of processor microarchitectures, and can vary enough between CPUs so that what the performance rankings are at 3 GHz are not necessarily what they are at 4 GHz. As an example, observe the Cinebench numbers where the lines of the i7-950 and the X6 1100T cross!

In order for i7-950 comparred to X6 1100T to have any meaning a 4 core AMD processor has to be used . According to your own ealier words. IPC vs core count you referranced to . I found your explanation to be closer to what in music we call allegro.

 

[jvroig]

Considering there is an identical 2.11x performance advantage that 2500k enjoys in both Single and Multi-threaded bench, how can you have it working at 3.7ghz in 1 case and only at 3.4ghz in another?

Good question, and there are several factors in multi-threaded performance than just IPC (that is measured from a single-thread bench, that is) . One of the other could simply have better scaling than the other due to their architectural differences (the other arch is simply better multi-threaded scaling-wise), or perhaps it is just that particular software that prefers one arch over another (happens very often, even in games)

Whatever the case, we have to stick to the facts, and not force them to fit a notion. In this case, doing so means skipping over factors that could affect the multi-threaded performance of the arch, or maybe in just that one particular software.

In this particular case, since we are talking about it, it seems to me that the Q6600 has poorer multi-threaded performance scaling than SB (at least in Cinebench), so despite SB going down about a measly 300MHz, it manages to maintain the same performance advantage.



EDIT: To add, since we have touched upon turbo and clock-to-clock comparisons, I would like to express my frustration in this same endeavor that you are trying to accomplish: IPC comparisons of processors ever since "Turbo" became widespread.

I remember spending more time than usual (and more frustrating than usual) looking for clock-to-clock comparisons, because that is what I need for my usage since overclocking comes into play with Turbo off. I know why review sites bench with turbo on, because that is the "out-of-the-box" performance, but as an enthusiast who will only run the chip at its highest stable clock, I don't care about the "out-of-the-box" performance, and for me to know what I need to know, I need clock-to-clock comparisons.

With all the dynamic and hard-to-predict "turbo" going on, it's not as easy as before when processors only came with one speed setting, unless a reviewer specifically allots a section for IPC comparisons, and it doesn't happen often at all.

 

[AtenRa]

Im sorry to say but Q6600 was not a quad core design but 2 dual core dies in the same PGA (Pin Grid Array).

So by comparing the two processors in benchmarks and say that SB 2500K has ~50&#37; more IPC is wrong.
In order to see the IPC of the two designs we have to measure ONLY a SINGLE core of each processor (edit: at the same frequency).

That ~50% between Q6600 and i5 2500K is performance and not IPC.