XeCPU performance metrics?

[Heinrich4]

About article translate i dont understand right this part :


"Though some say the performance of the Xbox 360 CPU is not very good, according to Capcom, the performance of a single core of the Xbox 360 CPU is 2/3 of the Pentium 4 with the same clock speed. When SMT is fully exploited, about 4 times larger performance can be observed. In terms of PC it's comparable with 4 SMT threads in a dual-core Pentium 4 Extreme Edition 840 (3.2GHz)."

So Xenon cpu have same or compare performance with P4 D 840 3.2GHz?

(i see IBM talking SPU with 25.1Gflops and never see any article talking about maximum performance of all 3 core Xenon)

 

[Carl B]

It's their opinion using a readily-understood comparable as the reference point. But "opinion" must be stressed, as others might be able to extract more, and they themselves are clearly in a better place than some others.

They're not talking about Flops though, but real-world performance, so it's a dangerous route to look at it via the Flops of a P4 if that's the route you're going. The XeCPU's max theoretical Flops are a known figure incidently.

 

[Heinrich4]

It's their opinion using a readily-understood comparable as the reference point. But "opinion" must be stressed, as others might be able to extract more, and they themselves are clearly in a better place than some others.

They're not talking about Flops though, but real-world performance, so it's a dangerous route to look at it via the Flops of a P4 if that's the route you're going. The XeCPU's max theoretical Flops are a known figure incidently, but this thread is better sticking to Capcom and Framework vs going down that oft-traveled road. The comparison itself is one independent of Flop considerations to begin with.

Agreed,but except Joker454 and this article i never see concrete information about real world performance of xenon cpu or Gflops processing and in my opinion thesis very important to compare or analize possibilities on each archteture console at same Engine.

 

[Carl B]

Agreed,but except Joker454 and this article i never see concrete information about real world performance of xenon cpu or Gflops processing and in my opinion thesis very important to compare or analize possibilities on each archteture console at same Engine.

The XeCPU is at ~116 GFlops, so if that's the number you're looking for there you go. But as for "real world" performance of either architecture in a games console, it's not surprising that we don't read much in terms of comparing them to other more mundane architectures. There's no particular reason for that... just rather, why would developers be dwelling on that?

 

[Darkon]

The XeCPU is at ~116 GFlops, so if that's the number you're looking for there you go. But as for "real world" performance of either architecture in a games console, it's not surprising that we don't read much in terms of comparing them to other more mundane architectures. There's no particular reason for that... just rather, why would developers be dwelling on that?

Wasn’t it 70 + Gflops ?

 

[Carl B]

Wasn’t it 70 + Gflops ?

76.8 GFLops actually... and yes that's a truer number; the ~116 GFlops includes the potential of the FPU, which really shouldn't be considered. BUT, whatever.

 

[Heinrich4]

76.8 GFLops actually... and yes that's a truer number; the ~116 GFlops includes the potential of the FPU, which really shouldn't be considered. BUT, whatever.

Essentially the point is - no more Flops/Cell vs XeCPU talk in this thread.

Sorry off topic.

But VMX and 2 x FPUs can operate in paralell?

76.8GFlops VMX + 38.4Gflops Fpu are incredible number for real world performance for Xenon, unless someone solved(threads etc) a memory problens controllers for this cpu.

 

[Heinrich4]

The XeCPU is at ~116 GFlops, so if that's the number you're looking for there you go. But as for "real world" performance of either architecture in a games console, it's not surprising that we don't read much in terms of comparing them to other more mundane architectures. There's no particular reason for that... just rather, why would developers be dwelling on that?

Im understand your point,but if many games or engines PC like can be port easily in x360 i think (despite console are closed box for devs,sdks,videogame etc) very interesting now how part of hardware contribute for this and with conterpart in pc universe in some way... and my guess is Xenon cpu is one of great part of this....and if a great article of Framework Engine MT of this topic compare SMT core Xenon and P4 D 840 3.2GHz maybe wee can search some information about that.

 

[ERP]

We've run a variety of tests internally accross console and PC CPU's, but they're really not very useful, other than in terms of expectations. It depends wildly on the type of test.

When your writing a game, you use what you have, I'm not going to spend anything like the same amount of time submitting polygons on a 360, but my statemachines that have wildly errratic memory accesses arent going to run as fast.

The sames basically true of GPU's, they are what they are in a closed box.

 

[Heinrich4]

We've run a variety of tests internally accross console and PC CPU's, but they're really not very useful, other than in terms of expectations. It depends wildly on the type of test.

When your writing a game, you use what you have, I'm not going to spend anything like the same amount of time submitting polygons on a 360, but my statemachines that have wildly errratic memory accesses arent going to run as fast.

The sames basically true of GPU's, they are what they are in a closed box.

Thanx for answer,but your results aproached in some way of Capcom/PcWatchimpress article?

 

[pjbliverpool]

About article translate i dont understand right this part :


"Though some say the performance of the Xbox 360 CPU is not very good, according to Capcom, the performance of a single core of the Xbox 360 CPU is 2/3 of the Pentium 4 with the same clock speed. When SMT is fully exploited, about 4 times larger performance can be observed. In terms of PC it's comparable with 4 SMT threads in a dual-core Pentium 4 Extreme Edition 840 (3.2GHz)."

So Xenon cpu have same or compare performance with P4 D 840 3.2GHz?

(i see IBM talking SPU with 25.1Gflops and never see any article talking about maximum performance of all 3 core Xenon)

Do you have a link to the article?

 

[AlNom]

Do you have a link to the article?

This thread is a split from the Lost Planet article translation that One did awhile back.

 

[ERP]

Thanx for answer,but your results aproached in some way of Capcom/PcWatchimpress article?

IMO It's a meaningless statement, it depends far too much on the workload to say CPU-X == CPU-Y.

 

[Heinrich4]

Do you have a link to the article?

You can see this here:

http://www.watch.impress.co.jp/game/docs/20070131/3dlp.htm

And translated here:

http://babelfish.altavista.com/babe...tch.impress.co.jp/game/docs/20070131/3dlp.htm


http://forum.beyond3d.com/showpost.php?p=919807&postcount=116

http://www.neowin.net/forum/lofiversion/index.php/t534590.html

 

[Heinrich4]

IMO It's a meaningless statement, it depends far too much on the workload to say CPU-X == CPU-Y.

How much workload Cpu-X (Xenon) and Cpu-Y(Cell) working better in sustained processing flops and integer(in how situation)?

 

[ERP]

Depends more on the data layout than on the integer/float mix

 

[Entropy]

Depends more on the data layout than on the integer/float mix

Halleluja.
For decades, the art of performance programming has been the art of managing memory.
For example, assume that your calculation requires reading two operands, yielding one result. That's a single FLOP. It requires 4*3=12 bytes of bandwidth if single precision FP.
So, if the BE has 25GB/s of bandwidth, that's enough to sustain 2 GFLOPs per second, or roughly 1% of its theoretical peak. (The 360 has significantly lower bandwidth to memory, shared with the GPU to make matters worse, but lets keep such comparisons out of this for now.)

How much computational performance you can squeeze out of any modern computer depends entirely on how well you manage to structure your problem to fit the memory hierarchy, and avoid interlocks.

I stopped counting MIPs or FLOPs two decades ago. Data flow is everything.

 

[Arwin]

Halleluja.
For decades, the art of performance programming has been the art of managing memory.

I'm with you so far

That's a single FLOP. It requires 4*3=12 bytes of bandwidth if single precision FP.
So, if the BE has 25GB/s of bandwidth, that's enough to sustain 2 GFLOPs per second, or roughly 1% of its theoretical peak. (The 360 has significantly lower bandwidth to memory, shared with the GPU to make matters worse, but lets keep such comparisons out of this for now.)

This though, seriously. Things are MUCH more complex on either systems, so much so that this comment is totally pointless?

How much computational performance you can squeeze out of any modern computer depends entirely on how well you manage to structure your problem to fit the memory hierarchy, and avoid interlocks.

And now I'm with you again.

 

[Entropy]

This though, seriously. Things are MUCH more complex on either systems, so much so that this comment is totally pointless?

And yet, didn't we see thread after thread devoted to counting FLOPs when the consoles were introduced? And the people doing it weren't the totally clueless - rather, it seemed to be those who had at least some technical background and/or interest. And do we not still see the same FLOPs numbers trotted about as soon as GPGPU is the topic?

To me this implies that a lot of people are both interested enough to care (and thus be able to learn), but not experienced enough to have a feeling for relative importance of various issues.

For example, if you are more sophisticated, you might add to your knowledge of instruction execution speed, the knowledge of main memory bandwidth, cache bandwidth and average access latency. Now we have already moved beyond the scope of these forums. Then we have to consider how the system deals with shared resources*, particular in multiCPU environments, and the tools at hand to manage conflicts. And then...and the application... and the data organisation... and the tools...

The rabbit hole just keeps getting deeper. What we can do who have been down a few of them, is to warn against the understandable desire for simple answers, and perhaps try to show why that desire is futile.

Ent

* I remember when the 360 and PS3 CPUs were first compared here and the overwhelming opinion of people was that three cores connected to cache was a dream of simplicity vs. the SPUs with their local store. Whereas I on the other hand thought that the deterministic environment afforded by the local store combined with generous communication data paths seemed like an organisation much less likely to burst programmer blood vessels. So not only is the picture complex and application specific, but how you interpret it depends on where you're coming from.

 

[Heinrich4]

And yet, didn't we see thread after thread devoted to counting FLOPs when the consoles were introduced? And the people doing it weren't the totally clueless - rather, it seemed to be those who had at least some technical background and/or interest. And do we not still see the same FLOPs numbers trotted about as soon as GPGPU is the topic?

To me this implies that a lot of people are both interested enough to care (and thus be able to learn), but not experienced enough to have a feeling for relative importance of various issues.

For example, if you are more sophisticated, you might add to your knowledge of instruction execution speed, the knowledge of main memory bandwidth, cache bandwidth and average access latency. Now we have already moved beyond the scope of these forums. Then we have to consider how the system deals with shared resources*, particular in multiCPU environments, and the tools at hand to manage conflicts. And then...and the application... and the data organisation... and the tools...

The rabbit hole just keeps getting deeper. What we can do who have been down a few of them, is to warn against the understandable desire for simple answers, and perhaps try to show why that desire is futile.

Ent

* I remember when the 360 and PS3 CPUs were first compared here and the overwhelming opinion of people was that three cores connected to cache was a dream of simplicity vs. the SPUs with their local store. Whereas I on the other hand thought that the deterministic environment afforded by the local store combined with generous communication data paths seemed like an organisation much less likely to burst programmer blood vessels. So not only is the picture complex and application specific, but how you interpret it depends on where you're coming from.

Thanx a lot for share your knowlegde,but the point is this: " So Xenon cpu have same or compare performance with P4 D 840 3.2GHz?"

You could talk about article of Pcwatchimpress impressions/experience of Capcom with cpu Xenon?

(sincerely, since now thanx in advance)