AMD competition for Core i3 (Gamers thread)

[ilkhan]

Yeah but most RTS games don't need high frame rates to be playable (and the game developers know this)

But if I had SLI or Crossfire and the game I was playing truly used four threads then I would be go with the quad core.

Depends on the RTS. Sim speed and FPS are independant in most RTS games, and sim speed is more important of the two.

I'll keep harping on SupCom because thats what I play, a quad doesn't help as much as clock speed. A good GPU helps to make it pretty (and playing maxed out with 1920x1200+1600x1200 is fairly demanding), but pretty is independent of playable.

The choice of 750 vs 660 will be an interesting one.

 

[jvroig]

Well Intel is pretty confident to price a dual core higher than an AMD quad core so it is obvious we haven't reached that point yet.

Intel's pricing has nothing to do with whether we've reached the point of "quad-cores being generally better than dual cores even in gaming" or not (but yes, we have). Intel's pricing is simply a result of them figuring out that lowering their sticker prices will actually mean castrating their profits far more than just letting AMD's cheaper offerings get a few more sales. With the volume they are shipping, and their far superior market share, they'd even probably let the price of a popular single core CPU they have (if they had any, but they don't anymore) remain higher than AMD's new cheap quad core.

 

[cbn]

Depends on the RTS. Sim speed and FPS are independant in most RTS games, and sim speed is more important of the two.

I'll keep harping on SupCom because thats what I play, a quad doesn't help as much as clock speed. A good GPU helps to make it pretty (and playing maxed out with 1920x1200+1600x1200 is fairly demanding), but pretty is independent of playable.

Not only that but these four thread games they keep telling me that are so hard on CPUs play @ low frame rates regardless of whether the Processor is dual core or quad core.

GTA 4 is a good example to me we still are in the dual core age. If quad core was so much better frame rates would have been doubled. But that didn't happen.

Instead we are talking 33% improvement with quad core over dual core in that game.

 

[cbn]

The choice of 750 vs 660 will be an interesting one.

Does turbo mode need to be disabled when overclocking?

 

[jvroig]

GTA 4 is a good example to me we still are in the dual core age. If quad core was so much better frame rates would have been doubled. But that didn't happen.
Instead we are talking 33% improvement with quad core over dual core in that game.

33% is plenty already. Framerates would double only if: 1.) The CPU was the only significant limiting factor in the first place (perhaps the GPU is already way too powerful for the scenario), and 2.) The game is optimized for quad core.

Of course, then you would say that that is exactly your point, not all games are optimized for quad or more cores. Fair enough. But despite that, gaining 33% is huge. Why would you not want 33% better framerates, considering that today the price difference of dual cores versus quad cores is getting smaller and smaller? It's almost a no-brainer, it's all over except for the shouting. Quad-cores have won. More cores have won. We don't even have to get into the argument of whether this outcome is ideal or not, since the fact of the matter is we've reached the limits of what the GHz wars can get us (not very far, so we've learned), so now it's the core/thread wars.

And in real life usage, are you just playing the game with nothing else running in the background? Or perhaps, in the middle of a gaming session your antivirus kicks in for it's daily scan? Having more than just two cores helps a lot.

 

[MrK6]

Not only that but these four thread games they keep telling me that are so hard on CPUs play @ low frame rates regardless of whether the Processor is dual core or quad core.

GTA 4 is a good example to me we still are in the dual core age. If quad core was so much better frame rates would have been doubled. But that didn't happen.

Instead we are talking 33% improvement with quad core over dual core in that game.

No need for SLI or CF then either, since they don't scale 100% either, right? How in your world does it work that increasing the number of processors will somehow equate to perfect scaling? The fact that a Q6600 at stock (2.4GHz) can keep up with an E8500 @ 3.6GHz in this game shows just how far you're off the mark.

 

[cbn]

How in your world does it work that increasing the number of processors will somehow equate to perfect scaling?

They should give close to perfect scaling.

But for whatever reason software designers are not dividing their games up into four equally sized threads. Left 4 Dead 2 does a better job at optimzing for quad core than GTA IV....but then again this game runs very fast even on a dual core.-->http://www.pcgameshardware.com/aid,...PU-benchmarks-Phenom-II-very-strong/Practice/

Maybe Valve is thinking ahead and planning to use that same game engine with tiny handheld devices?

 

[cbn]

No need for SLI or CF then either, since they don't scale 100% either, right?

It does scale close to 100% (as far as FPS goes) provided the processor isn't too slow. Increase resolution, detail settings and AA/AF to achieve this.

 

[ilkhan]

Does turbo mode need to be disabled when overclocking?

Ill give the obvious answer of: sometimes.

 

[Martimus]

They should give close to perfect scaling.

But for whatever reason software designers are not dividing their games up into four equally sized threads. Left 4 Dead 2 does a better job at optimzing for quad core than GTA IV....but then again this game runs very fast even on a dual core.-->http://www.pcgameshardware.com/aid,...PU-benchmarks-Phenom-II-very-strong/Practice/

Maybe Valve is thinking ahead and planning to use that same game engine with tiny handheld devices?

Scaling is limited by a number of factors, and the software design is only one of those factors. A good thing to read up on is Amdahl's Law: http://en.wikipedia.org/wiki/Amdahl's_law

 

[Fox5]

They should give close to perfect scaling.

But for whatever reason software designers are not dividing their games up into four equally sized threads. Left 4 Dead 2 does a better job at optimzing for quad core than GTA IV....but then again this game runs very fast even on a dual core.-->http://www.pcgameshardware.com/aid,...PU-benchmarks-Phenom-II-very-strong/Practice/

Maybe Valve is thinking ahead and planning to use that same game engine with tiny handheld devices?

As I've said, triple core is the sweet spot.
If you notice, triple cores perform almost identically to quad cores. Game developers are apparently not finding it difficult to split a game up into 3 primary threads, give them a bit longer until they reach 4 or more.

FWIW, the next generation of game engines are being designed to scale to n-cores, however the nature of games means there's a limit to scaling. Generally, you can't stay equally parallel throughout an entire sequence of operations, so your potential speed up collapses on you in any sequence of events with less parallel things to do in each step.
33% is probably the close to the maximum speed up you can expect from going from dual to quads in games. IIRC, the scaling for a sequence of tasks depends on how quickly the parallelism collapses, but is a logarithmic curve so the benefits from more cores quickly reaches diminishing returns.

Quad core processors exist because we can't get that 33% increase from clock speed.

 

[cbn]

33% is probably the close to the maximum speed up you can expect from going from dual to quads in games. IIRC, the scaling for a sequence of tasks depends on how quickly the parallelism collapses, but is a logarithmic curve so the benefits from more cores quickly reaches diminishing returns.

Quad core processors exist because we can't get that 33% increase from clock speed.

This is going to be problematic then. Maybe games will be focusing more on graphics in the future. Mini DTX style mainboards for gamers? One Big GPU and a tiny CPU on the same PCB? For whatever reason graphics seems to have a much easier time working in parallel. (The reason why we almost see perfect scaling with quad fire/crossfire despite the mirrored memory)

Speaking of CPU scaling here is GTA 4 http://www.pcgameshardware.com/aid,...ark-review-with-13-processors/Reviews/?page=2

Notice how Athlon 5000+ X2 (2 cores @ 2.6 Ghz) scales almost twice as fast as Athlon 4000+ (1 core @ 2.6 Ghz).

So at the moment it looks like getting two cores to work in parallel isn't that difficult? Or is it because two cores have an easier time sharing multiple threads? For this same reason wouldn't a quad core scale better if the total thread count were eight or higher? Wouldn't this improve the chances of all four cores working in parallel?

 

[cbn]

FWIW, the next generation of game engines are being designed to scale to n-cores, however the nature of games means there's a limit to scaling. Generally, you can't stay equally parallel throughout an entire sequence of operations, so your potential speed up collapses on you in any sequence of events with less parallel things to do in each step.

I think I know what you are talking about with this term "staying parallel". You mean threaded instructions happening simultaneous to another set of threaded instructions right?

If there aren't enough instructions happening simultaneously then benefits of additional cores goes wasted. If one set of threaded instructions is no longer used by the game another threaded set needs to take its place in order to preserve the maximum potential of the processor.

 

[cbn]

As I've said, triple core is the sweet spot.
If you notice, triple cores perform almost identically to quad cores. Game developers are apparently not finding it difficult to split a game up into 3 primary threads, give them a bit longer until they reach 4 or more.

Good point.

It could be different genres of games will also emerge as software programmers strive to become more creative with the additional CPU cores.

 

[Fox5]

I think I know what you are talking about with this term "staying parallel". You mean threaded instructions happening simultaneous to another set of threaded instructions right?

If there aren't enough instructions happening simultaneously then benefits of additional cores goes wasted. If one set of threaded instructions is no longer used by the game another threaded set needs to take its place in order to preserve the maximum potential of the processor.

Let's say you have a game with 4 primary steps:
The first step has 8 things happening, and we'll divide by half for each next step.
xxxxxxxx
xxxx
xx
x

With a quad core processor, that first step can be made 4x as fast, the 2nd 4x as fast, but no scaling for the 3rd over a dual core, and nothing for the 4th over a single core:
xx xx xx xx
x x x x
x x
x

This will eventually happen for any app, it just depends on how quickly it happens.

Even for a simple example like adding a bunch of numbers together:
If I have the numbers 1 through 10 (and thus a limited data set), I would add them together like this on a single core processor:
1 + 2 = 3
3 + 3 = 6
6 + 4 = 10
10 + 5 = 15
15 + 6 = 21
21 + 7 = 28
28 + 8 = 36
36 + 9 = 45
45 + 10 = 55

Of course there's little tricks to do this faster, but ignoring them, this is how a single core processor would do it. (one trick: 1+10 = 11, 2 + 9 = 11 and so on, you have 5 sets of 11, so just multiply 5 * 11, if you can program something to recognize this)
On a multi-core processor, you could do:
1 2 3 4 5 6 7 8 9 10
1+2 3+4 5+6 7+8 9+10
3+7 11+15 19
10+26 19
36+9
55

Not only does going multicore sometimes mean not using tricks, it can mean using a less efficient algorithm in general. Not in the two examples above, both took 9 operations, however the single core took 9 steps, while the multicore only took 5. So 5x the processing power gave us less than a 100% speed up, and this is how most algorithms are, before scaling difficulties from the lack of instantaneous communication come into play. Some algorithms are even recursive by nature, and can't be parallelized at all (assuming working on the same data).
Some algorithms, in order to do them in parallel, require a less efficient implementation entirely.
A contrived example would be multiplication or division. What I just did was 11*5, but if addition only takes 1 cycle, while mutiplication takes...6, you achieved a speed up by changing the code to parallel additions, but just barely.

Edit:
When dual core processors first game out, you needed at least a 50% advantage in clock speed to make a single over a dual worthwhile. Now you need 100%. The same will eventually hold true for quad over dual. Games will become designed around the limitations and strengths of multi core processors, even if there's a 'more efficient' way to do things on less cores, the preference will be to program to run best with more hardware, not less.
The only thing that will hold this back is communication bandwidth. The Core 2 Quad design wouldn't have scaled very well going to more cores, but the phenom and i7 should scale pretty well. It's feasible that someday we could be at the level of parallelism shown above, with each thread's granularity small enough to do individual additions or multiplications.

 

[toyota]

This is going to be problematic then. Maybe games will be focusing more on graphics in the future. Mini DTX style mainboards for gamers? One Big GPU and a tiny CPU on the same PCB? For whatever reason graphics seems to have a much easier time working in parallel. (The reason why we almost see perfect scaling with quad fire/crossfire despite the mirrored memory)

Speaking of CPU scaling here is GTA 4 http://www.pcgameshardware.com/aid,...ark-review-with-13-processors/Reviews/?page=2

Notice how Athlon 5000+ X2 (2 cores @ 2.6 Ghz) scales almost twice as fast as Athlon 4000+ (1 core @ 2.6 Ghz).

So at the moment it looks like getting two cores to work in parallel isn't that difficult? Or is it because two cores have an easier time sharing multiple threads? For this same reason wouldn't a quad core scale better if the total thread count were eight or higher? Wouldn't this improve the chances of all four cores working in parallel?

you keep talking about quad cores as if they are all the same. the i7 blows your little theory out of the water in GTA 4. even at just 2.66 the 920 i7 is much much faster than the 3.16 E8500 and still quite a bit faster than the 2.8 Core 2 Quad. http://www.pcgameshardware.com/aid,...ead-of-Core-2-Quad-in-CPU-benchmarks/Reviews/thats also with just a 4870 so a much faster card will give the i7/i5 quads an even bigger advantage especially if they were overclocked.

if you are going to be building a gaming computer in 2010 then there is no point in going with anything lower than an i7 or i5 quad if you want the BEST performance.

 

[MrK6]

They should give close to perfect scaling.

But for whatever reason software designers are not dividing their games up into four equally sized threads. Left 4 Dead 2 does a better job at optimzing for quad core than GTA IV....but then again this game runs very fast even on a dual core.-->http://www.pcgameshardware.com/aid,...PU-benchmarks-Phenom-II-very-strong/Practice/

Maybe Valve is thinking ahead and planning to use that same game engine with tiny handheld devices?

Nope, not at all.

Scaling is limited by a number of factors, and the software design is only one of those factors. A good thing to read up on is Amdahl's Law: http://en.wikipedia.org/wiki/Amdahl's_law

Exactly. One has to remember that the CPU is only part of the machine that produces the final product (said frames at said rate). You need to be incredibly CPU limited and have excellent software implementation before you can see dramatic scaling over cores in a gaming environment.

It does scale close to 100% (as far as FPS goes) provided the processor isn't too slow. Increase resolution, detail settings and AA/AF to achieve this.

As stated, because you're creating an extremely GPU-bottlenecked environment. Even then, you never see 100% scaling, and many times it's much less.

 

[cbn]

the i7 blows your little theory out of the water in GTA 4. even at just 2.66 the 920 i7 is much much faster than the 3.16 E8500

You are comparing something with newer microarchitecture and turbo mode to something that doesn't have these things.

you keep talking about quad cores as if they are all the same.

Well not all dual cores are the same either.

Core i3 dual core with hyperthreading is going to be far quicker at dual threaded and quad threaded tasks than something like AMD 3800x2 or AMD Opty 165.

The point I am trying to make is that it appears quad threaded gaming programs haven't reached maturity yet. If this progression continues like this at what point does it make sense to be more graphically biased with hardware selection? Graphic code progressing faster than CPU code?

Good example: Whenever a new Video card is released it is capable of being fully utilized (provided AA/AF/resolution, detail settings can be increased). We can't say the same thing for CPU cores. Increasing the CPU die sized via increases in cores does not yield the same linear increases. Maybe this is why Intel and AMD chose to add L3 cache instead of cores for X given die size? In a nutshell microarchitecture >> number of cpu cores.

 

[cbn]

Nope, not at all.

When dual core processors first game out, you needed at least a 50% advantage in clock speed to make a single over a dual worthwhile. Now you need 100%. The same will eventually hold true for quad over dual. Games will become designed around the limitations and strengths of multi core processors, even if there's a 'more efficient' way to do things on less cores, the preference will be to program to run best with more hardware, not less.
The only thing that will hold this back is communication bandwidth. The Core 2 Quad design wouldn't have scaled very well going to more cores, but the phenom and i7 should scale pretty well. It's feasible that someday we could be at the level of parallelism shown above, with each thread's granularity small enough to do individual additions or multiplications.

Notice the word "eventually".

 

[cbn]

As stated, because you're creating an extremely GPU-bottlenecked environment. Even then, you never see 100% scaling, and many times it's much less.

http://www.anandtech.com/video/showdoc.aspx?i=3658&p=1

In this review two HD5770s scaled over 100% compared to a single HD5870 (see Far Cry2, Hawx, Resident Evil 5, Batman)

 

[MrK6]

Notice the word "eventually".

And, what's your point?

http://www.anandtech.com/video/showdoc.aspx?i=3658&p=1

In this review two HD5770s scaled over 100% compared to a single HD5870 (see Far Cry2, Hawx, Resident Evil 5, Batman)

In none of those benchmarks did the 5770 scale 100%, and they actually just proved everything I said above when creating a GPU-bottlenecked environment. So, again, what's your point? You still have yet to illustrate that you actually understand how scaling works and how CPU scaling manifests itself in a gaming environment.

 

[cbn]

And, what's your point?

In none of those benchmarks did the 5770 scale 100%

Two HD5770s got higher frame rates than a single HD5870.

HD5770 is basically HD5870 split in half right?

If that is not scaling then I don't know what is.

For a definition of scaling read this--->http://forums.anandtech.com/showthread.php?t=334565

 

[cbn]

And, what's your point?

It appears to me that microarchitectural advances are outpacing core scaling. I want to know what AMD has planned to counter this?

 

[Hyperlite]

Two HD5770s got higher frame rates than a single HD5870.

HD5770 is basically HD5870 split in half right?

If that is not scaling then I don't know what is.

For a definition of scaling read this--->http://forums.anandtech.com/showthread.php?t=334565

that is a gross, massive oversimplification that makes zero sense. Based on that logic, dual GPU cards should scale 100% with their single core parent.



**SPOILER ALERT**

they don't.

 

[cbn]

Based on that logic, dual GPU cards should scale 100% with their single core parent.

I realize there are internal losses...but still I am suprised to see this happen --->dual GPUs scaling 100% to their parent card http://www.anandtech.com/video/showdoc.aspx?i=3658&p=1

People will talk about driver issues but if this were really true I think we would see the reverse trend happening. (unless the driver problem was at the individual card level for specific games and something about the Crossfire driver was able to correct this)

Another way of looking at the issue: GPU makers can continually release stronger and stronger GPUs every year and the potential appears to exist (even with a degree of CPU dependence) for full utilization. The CPU story looks different. Intel and AMD (at the gaming level) have not been able to double CPU power at the same intervals as the GPU makers. Example: Core i7/Phenom II adds more die size hogging L3 cache instead of CPU cores.