How Much Will I Gain in Gaming When Upgrading from a Phenom II X4 965 to an X6 1090T?

[NTMBK]

Sounds nice and dandy, except they do not use those instructions because those are not suited for gaming but for things like encryption and encoding. Games still use SSE3 and its variants.

So... no, that's not the answer.

SSE4 includes single instruction dot-product (_mm_dp_ps) and FMA3 has single instruction fused-multiply add (_mm_fmadd_ps). Both of these are incredibly useful in 3D geometry workloads, which is a big part of the majority of games.

 

[ThePeasant]

No, it's not. Did you even look at CineBench?

What does cinebench (or itunes for that matter) have to do with any game? In fact what does any benchmark other than the actual games themselves have to do with the games?

In terms of absolute IPC K10.5 is faster. The only way Bulldozer is faster is with new instructions like AVX and AES, and games DO NOT use those. Also, the 960T and 1100T have Turbo as well.

Tests were also ran on Windows 8 and all this scheduling hoopla did for performance was bring it up by less than 5%. Again, there is zero reason for the FX-4100 to be faster than the 1090/1100T in games.

Just out of curiosity, how do you know the kinds of instructions used in games?

Anyway, the bolded parts are where we strongly disagree. I don't know how else to explain that absolute statements about IPC like that, especially regarding radically different architectures, cannot be so easily made.

 

[LOL_Wut_Axel]

SSE4 includes single instruction dot-product (_mm_dp_ps) and FMA3 has single instruction fused-multiply add (_mm_fmadd_ps). Both of these are incredibly useful in 3D geometry workloads, which is a big part of the majority of games.

If games were using FMA3 AMD would be pointing to them left and right.

 

[LOL_Wut_Axel]

What does cinebench (or itunes for that matter) have to do with any game? In fact what does any benchmark other than the actual games themselves have to do with the games?



Just out of curiosity, how do you know the kinds of instructions used in games?

Anyway, the bolded parts are where we strongly disagree. I don't know how else to explain that absolute statements about IPC like that, especially regarding radically different architectures, cannot be so easily made.

Hmm... maybe the fact that games still rely heavily on single-threaded performance? And the fact that the Modular approach means you lose even more ST performance if you're running games that are more multi-threaded? It's been calculated that the loss in performance from CMT is anywhere from 10-20%. Let's even assume for a minute that the FX-4100 has higher ST performance than the 1100T (which obviously it does not). Now let's assume that the average game will use all four threads but no more, and that it loses 15% of its performance because of CMT. How does it end up 20-30% faster in games than a real quad or six-core?

And no, those new instructions won't give it that big a boost. If you look, you can see the 1st gen Core processors doing quite well and they don't have the new instructions either.

Nah, that can't be it.

 

[GaiaHunter]

Hmm... maybe the fact that games still rely heavily on single-threaded performance? And the fact that the Modular approach means you lose even more ST performance if you're running games that are more multi-threaded? It's been calculated that the loss in performance from CMT is anywhere from 10-20%. Let's even assume for a minute that the FX-4100 has higher ST performance than the 1100T (which obviously it does not). Now let's assume that the average game will use all four threads but no more, and that it loses 15% of its performance because of CMT. How does it end up 20-30% faster in games than a real quad or six-core?

And no, those new instructions won't give it that big a boost. If you look, you can see the 1st gen Core processors doing quite well and they don't have the new instructions either.

Nah, that can't be it.

If the games heavily depend on single thread performance then the conclusion would be BD and PD have higher single thread performance (or enough to make up with higher frequency).

Since that only seems to be happening in more recent games, one is led to believe it is simpler (require less work) to optimize for BD/PD architecture.

 

[2is]

Hmm... maybe the fact that games still rely heavily on single-threaded performance? And the fact that the Modular approach means you lose even more ST performance if you're running games that are more multi-threaded? It's been calculated that the loss in performance from CMT is anywhere from 10-20%. Let's even assume for a minute that the FX-4100 has higher ST performance than the 1100T (which obviously it does not). Now let's assume that the average game will use all four threads but no more, and that it loses 15% of its performance because of CMT. How does it end up 20-30% faster in games than a real quad or six-core?

And no, those new instructions won't give it that big a boost. If you look, you can see the 1st gen Core processors doing quite well and they don't have the new instructions either.

Nah, that can't be it.

Nothing you said here addresses anything in the post you quoted. Plenty of games rely on both single and multi threaded performance and none of what you said backs up your earlier claim of games not using any new instructions.

I suggest you look up the phrases "red herring" and "non-sequitur"

 

[ThePeasant]

Well it seems you aren't grasping the fundamental concepts of my argument. Let me say it one last time:

There are numerous flaws in what your argument:-

1. You are speaking in absolutes; you use a set of benchmarks to represent all program existing and yet written but IPC and "CMT tax" are stochastic.
2. You are ignoring benchmarks which contradict what you assert (fallaciously) simply because they contradict (confirmation bias), thus completing the circle.

To simplify that you are saying because these benchmarks show one architecture being superior, then these other benchmarks which show situations where that isn't true must be intrinsically false (intentionally manipulated).

That being said I am curious (not suspicious) where all these gains are coming from.

 

[tweakboy]

sad real 6 cores loses to a fake 8 core.

Now that was a great post....... lol

 

[tweakboy]

Wait til Ivy E comes out,,"6 to 12 core" it will blow out any CPU ,, haswell initially, AMDs even newer offering with 8 physical core.