Sneak peak at FX-8150 with SuperPi & wPrime

[Ferzerp]

Or history repeats itself and they are counting on crazy clock speed scaling to make it competitive?

 

[ThatsABigOne]

I still wonder why they even bother to post SuperPi results. It uses the archaic X87 instruction set and is not good at all for disseminating single-threaded performance. It's bad for modern CPUs, even more for AMD's since it gives Intel's architecture a huge advantage.

Just to point out how bad SuperPi is, the old CPU in my laptop (Core 2 Duo P7450, now a T9600) overclocked to 2.7GHz calculates 1M in 19.465s, beating the FX-8150 by more than a whole second while being clocked 900MHz lower.

And yes, it was using SuperPi mod 1.5.

Your T9600 should be 2.8ghz stock not 2.7ghz. :whiste::biggrin:
I will run the test when I will get back to windows. :awe:

 

[ThatsABigOne]

Sorry to double post. Too lazy to edit. How do you guys tell if this is fakie?

 

[BlueBlazer]

SuperPi is in no way, shape, or form an accurate comparison for Intel vs. AMD. Intel has decided even in their new CPUs to support this old and outdated instruction set, while AMD has not. That's why Intel's CPUs get much higher performance.

And saying that it favors Intel is again incorrect. You can check this as it clearly shows difference between Netburst, K8 and Conroe. Even though x87 is present in all CPUs (both AMD and Intel) for legacy software, Intel discourages it and endorses replacement with faster SSE and SSE2 (which will inflate FP scores/performance). Since it doesn't use any special instructions, SuperPi can run on all CPUs, even on ancient Pentium Pro, to compare IPC and instruction execution efficiency on the same code/binary.

Ok we have a test system directly from AMD and it behaves like any other leaks. So facts are:
-zambezi 8C @ 3.6Ghz is slower than K10 in super pi <= not big of a deal so we can move on
-zambezi 8C @ 3.6Ghz is as fast as llano QC 3650 @ 2.6Ghz in wprime. This one is hard tp skip. This is a proof that none of the systems out there,even those that AMD show off to the press are representative of final performance. There is absolutely no chance that FX8150 is going to be slower than 2.6 QC K10.

The only logical conclusion is that all performance leaks shoot be disregarded until launch day.

Considering that this is a demo/test system from AMD, shouldn't it have the latest stepping and near shipping/final product? And there's still officially no word from AMD as to final release dates (whether its shipping nor launch). All we can find are rumors and speculations based on bits and pieces of information. :hmm:

Or history repeats itself and they are counting on crazy clock speed scaling to make it competitive?

As already said earlier, its all about "overclocking" and price. That's one conclusion I came about after going through all the information so far.

Sorry to double post. Too lazy to edit. How do you guys tell if this is fakie?

Unfortunately it is not fake.

 

[LOL_Wut_Axel]

Your T9600 should be 2.8ghz stock not 2.7ghz. :whiste::biggrin:
I will run the test when I will get back to windows. :awe:

My P7450 was overclocked to 2.7GHz when I tested it some months ago. The current CPU in my laptop is a T9600, overclocked to 3.3GHz.

 

[LOL_Wut_Axel]

]And saying that it favors Intel is again incorrect. You can check this as it clearly shows difference between Netburst, K8 and Conroe. Even though x87 is present in all CPUs (both AMD and Intel) for legacy software, Intel discourages it and endorses replacement with faster SSE and SSE2 (which will inflate FP scores/performance). Since it doesn't use any special instructions, SuperPi can run on all CPUs, even on ancient Pentium Pro, to compare IPC and instruction execution efficiency on the same code/binary. [/B]

AMD stopped supporting legacy X87 instructions on their CPUs some years ago, so yes, by default it'll place Intel CPUs much higher since they continue to have support for it. K10.5 has 8% lower IPC than Penryn. If we were to believe SuperPi, that'd be a much higher number.

Just because Intel encourages using newer instruction sets like SSE2 doesn't mean they stopped supporting the old X87 instruction set, which is exactly why there's a big difference in it comparing AMD to Intel. IIRC, ever since K10 AMD hasn't supported X87 natively.

 

[Ferzerp]

Let's see, my aged i7-930 does it in 10.182s

 

[Dresdenboy]

AMD stopped supporting legacy X87 instructions on their CPUs some years ago, so yes, by default it'll place Intel CPUs much higher since they continue to have support for it. K10.5 has 8% lower IPC than Penryn. If we were to believe SuperPi, that'd be a much higher number.

Just because Intel encourages using newer instruction sets like SSE2 doesn't mean they stopped supporting the old X87 instruction set, which is exactly why there's a big difference in it comparing AMD to Intel. IIRC, ever since K10 AMD hasn't supported X87 natively.

Sorry, that's plain wrong.

 

[BlueBlazer]

AMD stopped supporting legacy X87 instructions on their CPUs some years ago, so yes, by default it'll place Intel CPUs much higher since they continue to have support for it. K10.5 has 8&#37; lower IPC than Penryn. If we were to believe SuperPi, that'd be a much higher number.

All x86 CPUs have to support (or include) legacy x87 instructions for legacy programs (in other words, backwards compatibility for older programs). If AMD had removed x87 then SuperPi will not even run. Just like some programs and games that requires SSE2 will not run on my old Ahtlon XP but will run on my old Pentium4.

Just because Intel encourages using newer instruction sets like SSE2 doesn't mean they stopped supporting the old X87 instruction set, which is exactly why there's a big difference in it comparing AMD to Intel. IIRC, ever since K10 AMD hasn't supported X87 natively.

Did you know starting from Pentium4 the x87 has become redundant as Intel no longer "supports" it? That is no further developments on x87. That's why Pentium4's x87 performance was poorer than Pentium-III and AMD's K7. In its place Intel recommends SSE2, thus any program that uses them will have better performance while older/legacy programs using x87 will have lower performance. That's why in the early days, the benchmarks vary wildly (those that use SSE2 are likely to favor Pentium4, while those using SSE and x87 are likely to favor Athlons and Pentium-IIIs). :hmm:

 

[james1701]

the how hard would it be to just recode SPi to use sse2?

 

[BlueBlazer]

the how hard would it be to just recode SPi to use sse2?

Its not hard to do that, but it will "screw up" all previous results (previously bad results suddenly become better by leaps and bounds when re-run with a new SuperPi program that uses SSE2).

The SuperPi program has remained unchanged for many years, pretty much a good measurement of how modern processors handle legacy stuff (running old code/binaries). Besides IPC, its also sensitive to cache and memory performance.

 

[Sp12]

the how hard would it be to just recode SPi to use sse2?

I think I could do it.

Issue is that it's been the same program for years, and changing it would no longer make it a good measure of legacy X87 performance.

 

[LOL_Wut_Axel]

*snip*

Yes, but AMD hasn't optimized for it for years now. I guess I used the word "support" incorrectly.

X87 is outdated junk anyway, so who cares about the results in a program that uses it?

 

[RussianSensation]

X87 is outdated junk anyway, so who cares about the results in a program that uses it?

How come SuperPi shows very accurate IPC increases from C2D/Q to Core i5/i7 (1st gen) to Core i5/i7 (sandy bridge generation) then?

How come A64 lost to C2D in SuperPi and also ended up losing in the real world?

How come SuperPi shows Phenom I/II losing to modern Intel processors, and this also shows up as similarly inferior IPC in real world apps?

How come when AMD was faster in SuperPi during Athlon 64/X2/FX era vs. Pentium 4, no one complained that SuperPi was "useless"? :whiste:

I am not saying that SuperPi should be used to predict performance in rendering, video encoding, or decryption tasks (which are obviously multi-threaded), but thus far in the last 10 years, it has proven to be a prett accurate predictor of single-threaded/IPC performance.

If BD is the first processor to go against this trend (i.e., horrible SuperPi performance but insane real world performance due to its peculiar module design), then we can stop using SuperPi to compare AMD to Intel CPUs. The reason people still use it to gauge single-threaded/IPC performance is because it has been an accurate predictor of both up to this point!

 

[BlueBlazer]

Yes, but AMD hasn't optimized for it for years now. I guess I used the word "support" incorrectly.

Both AMD and Intel stopped optimizing or stopped development of x87 for many years already in favor of SSE, SSE2, SSE3, SSE4 and AVX.

X87 is outdated junk anyway, so who cares about the results in a program that uses it?

This program is good for measuring CPU performance running on essentially the same code/binary (no special optimizations and no special speed-up instrcutions). In this manner, legacy x86/x87 performance can be compared across all processors (from ancient to latest). For example, comparing Athlon XP with Pentium4 without using SSE2.

 

[ThatsABigOne]

My P7450 was overclocked to 2.7GHz when I tested it some months ago. The current CPU in my laptop is a T9600, overclocked to 3.3GHz.

Danger: Thread derailment ahead.

Got it. Mine is at 3.1ghz. I would go to 3.16ghz, but anything higher will result in a crash. Oh, I am undervolted.

What laptop is it?
Mine is aging Asus G51vx with GTX260m.

 

[LOL_Wut_Axel]

How come SuperPi shows very accurate IPC increases from C2D/Q to Core i5/i7 (1st gen) to Core i5/i7 (sandy bridge generation) then?

Perhaps because Intel always optimized better for it? A quick comparison: a Phenom II X4 955 completes SuperPi 32MB 5% slower than a Core 2 Duo E6700 (Conroe). We know that Conroe has 3% higher IPC than K10.5, so if the benchmark were accurate the performance difference would be in that ballpark. Yet here we see the E6700 running at a clock speed 20% lower manages a 5% better result, which means it's either biased in favor of Intel architectures or Intel optimized better for X87.

Both AMD and Intel stopped optimizing or stopped development of x87 for many years already in favor of SSE, SSE2, SSE3, SSE4 and AVX.

This program is good for measuring CPU performance running on essentially the same code/binary (no special optimizations and no special speed-up instrcutions). In this manner, legacy x86/x87 performance can be compared across all processors (from ancient to latest). For example, comparing Athlon XP with Pentium4 without using SSE2.

See above.

 

[LOL_Wut_Axel]

Danger: Thread derailment ahead.

Got it. Mine is at 3.1ghz. I would go to 3.16ghz, but anything higher will result in a crash. Oh, I am undervolted.

What laptop is it?
Mine is aging Asus G51vx with GTX260m.

ASUS G60VX, GTX 260M as well. Battery life kinda sucks (2-2:30 hours on average) and the weight is a bit high at 7 pounds, so I'm looking to replace it with a Sager NP5165.

Mine is stable at 3.35GHz max. At 3.4GHz it always BSODs. I think what limits it is the Memory, so I just decided to leave it at 3.3GHz. Stock voltage. Temps get a bit high on full load (78C), but nothing alarming.

Sorry for thread derailment, lol.

 

[BlueBlazer]

Perhaps because Intel always optimized better for it? A quick comparison: a Phenom II X4 955 completes SuperPi 32MB 5&#37; slower than a Core 2 Duo E6700 (Conroe). We know that Conroe has 3% higher IPC than K10.5, so if the benchmark were accurate the performance difference would be in that ballpark. Yet here we see the E6700 running at a clock speed 20% lower manages a 5% better result, which means it's either biased in favor of Intel architectures or Intel optimized better for X87.

Again, as mentioned before its not optimized. Do you know that it favored Athlon XPs over Pentium4? If it had used SSE2, then Pentium4 would have been walking all over the Athlon XP. Yes, that's how far back that SuperPi has been in existence.

See above.

It says "all time based values added". Its not based on only one program results. In other words, some of those programs are using SSE, SSE2, SSE3, etc.

 

[LOL_Wut_Axel]

Again, as mentioned before its not optimized. Do you know that it favored Athlon XPs over Pentium4? If it had used SSE2, then Pentium4 would have been walking all over the Athlon XP. Yes, that's how far back that SuperPi has been in existence.

It says "all time based values added". Its not based on only one program results. In other words, some of those programs are using SSE, SSE2, SSE3, etc.

Right... so can you actually explain why Conroe, which has 3% higher IPC than K10.5, performs better while having 20% lower clock speeds in SuperPi?

Also, read the whole article on Tom's. They use newer instruction sets because they perform well on both architectures.

 

[BlueBlazer]

Right... so can you actually explain why Conroe, which has 3% higher IPC than K10.5, performs better while having 20% lower clock speeds in SuperPi?

You are taking into account the average speedup. However performance varies from program to program. Differences can range from tiny to very large. Check this benchmarks, you will notice in some instances differences can be as much as 47% (such as the Excel benchmark) clock-to-clock. Now can you explain how a program written before even Conroe comes into existance, is somehow optimized for Conroe? Time travel?

Also, read the whole article on Tom's. They use newer instruction sets because they perform well on both architectures.

The new instructions leveled the playing field, and only thing left is how well the CPU executes them. The point is that these are all accumlated values, and in some instances in some programs the difference is small (as above benchmark link shows).

 

[LOL_Wut_Axel]

You are taking into account the average speedup. However performance varies from program to program. Differences can range from tiny to very large. Check this benchmarks, you will notice in some instances differences can be as much as 47% (such as the Excel benchmark) clock-to-clock. Now can you explain how a program written before even Conroe comes into existance, is somehow optimized for Conroe? Time travel?

The new instructions leveled the playing field, and only thing left is how well the CPU executes them. The point is that these are all accumlated values, and in some instances in some programs the difference is small (as above benchmark link shows).

Yes, because as its been known for a long time now, Excel runs better on Intel CPUs. You can find one anomaly, but it doesn't change anything. SuperPi runs much better on Conroe than on K10.5, just like Excel does. That much couldn't be clearer.

Because of that, there's not much point in comparing AMD and Intel using it.

 

[BlueBlazer]

Because of that, there's not much point in comparing AMD and Intel using it.

Why not? That code/binary is not optimized for either AMD or Intel architectures. Its a measure of how well modern CPUs (from both AMD and Intel) handle legacy x86/x87 code. That's a common factor. Its also sensitive to memory (as your graph shows). This is just one of the "de-facto" benchmarks around (commonly used by many to gauge performance changes), and there are others such as wPrime, Cinebench, POVray, 3dmark, etc. Whenever you see an engineering sample tested, you will likely to see these programs in action (besides SuperPi, lately Cinebench R11.5 gets the spotlight).

 

[T_Yamamoto]

54 Seconds on a Pentium D 820 @2.8 GHZ LMAO

 

[LOL_Wut_Axel]

Why not? That code/binary is not optimized for either AMD or Intel architectures. Its a measure of how well modern CPUs (from both AMD and Intel) handle legacy x86/x87 code. That's a common factor. Its also sensitive to memory (as your graph shows). This is just one of the "de-facto" benchmarks around (commonly used by many to gauge performance changes), and there are others such as wPrime, Cinebench, POVray, 3dmark, etc. Whenever you see an engineering sample tested, you will likely to see these programs in action (besides SuperPi, lately Cinebench R11.5 gets the spotlight).

I just explained this. For the same reason why you shouldn't compare K10.5 and Sandy Bridge using DiRT 3. There's applications that are made to take advantage of some CPUs over the others, so they shouldn't be the main ones we look at. Instead, we should discard them for the most part as an overall outlook on performance.

The benchmark was made to take advantage of K10.5 and Nehalem yet not of Sandy Bridge. Should we use it to base performance, therefore? No.

wPrime isn't very good, either. It puts the Atom D510 as having similar performance to a Core 2 Duo SU7300.