[Ars][Unconfirmed] PS4 to have an x86 AMD CPU

[Arzachel]

No, you can't say the same.

PC developers DON'T spend nearly as much optimizing the code because they don't need to. It is the luxury of having powerful hardware.

I still don't understand what are you getting at? You're arguing against a straw man, I said that static hardware targets make optimization much easier, how is that at odds with what you've said?

 

[taltamir]

For it to be a strawman it must be a misrepresentation of your argument. Such that I am arguing against something we both agree on but I claim is your argument, when it isn't.

You said optimizing for a console can be done more easily (I did not contest this; although I do think it is being grossly exaggerated). But you also said that developers optimize for consoles because it is easier to optimize for them.

I contradicted the latter claim, saying they optimize for consoles because they must, not because its easier. They prefer not to spend the money doing any optimization.

And then you said

You could say the same thing about any kind of software developer, what is your point?

Defending your claim that they want to optimize for consoles because consoles are easy to optimize for...
With the totally unrelated claim that PC developers also need optimizing (I don't see how that is possibly a defense of the other claim).

I disagreed with the claim (PC developers don't need to optimize nearly as much) and I disagreed with it being a defense of the other claim.

 

[Arzachel]

But you also said that developers optimize for consoles because it is easier to optimize for them.

I contradicted the latter claim, saying they optimize for consoles because they must, not because its easier. They prefer not to spend the money doing any optimization.

And herein lies the strawman:

console developers can do much more optimization because they only have to target a few combinations of hardware

I stated that more in-depth optimization is easier as in requires less resources. That was a single statement.

Where did you get:

But you also said that developers optimize for consoles because it is easier to optimize for them.

?

 

[IntelUser2000]

That's not saying much, the GPU's in this console generation aren't exactly stellar either. RSX is either a 7800 with half the ROPs and memory bandwidth or a 7600 with double the pixel pipes depending on how you look at it.

I meant that in the relative sense obviously. GPUs in the Xbox and the PS are relatively modern while the CPUs used are nowhere near it.

On the PC they just set minimum/recommended requirements that are higher and let inefficient code be compensated for with hardware power.

Yes. Newer hardware is better than software optimization, but the original point was you can't exactly compare PC based hardware to a console based hardware.

 

[Fox5]

I meant that in the relative sense obviously. GPUs in the Xbox and the PS are relatively modern while the CPUs used are nowhere near it.



Yes. Newer hardware is better than software optimization, but the original point was you can't exactly compare PC based hardware to a console based hardware.

There's probably a ~10x gap between the PC gpus and the console gpus at this point. I'm not sure the same gap exists between the cpus...
Say an xbox 360 or ps3 cpu core is ~ equal to an Atom proccessor per mhz, the 360 still has 3 with SMT (which benefits in order architectures quite a bit more than out of order architectures) at 3.2Ghz, and the PS3 at 6 at the same speed. The clock speed advantage could still put their cpus on par with Intel's ULV chips (in the low 1Ghz range), which is a respectable amount of performance, and a closer gap than the GPUs. Of course, you can scale down graphics a lot easier than CPU processing needs.

 

[nismotigerwvu]

There's probably a ~10x gap between the PC gpus and the console gpus at this point. I'm not sure the same gap exists between the cpus...
Say an xbox 360 or ps3 cpu core is ~ equal to an Atom proccessor per mhz, the 360 still has 3 with SMT (which benefits in order architectures quite a bit more than out of order architectures) at 3.2Ghz, and the PS3 at 6 at the same speed. The clock speed advantage could still put their cpus on par with Intel's ULV chips (in the low 1Ghz range), which is a respectable amount of performance, and a closer gap than the GPUs. Of course, you can scale down graphics a lot easier than CPU processing needs.

Not even close man, these CPU's were dogs new, let alone now. I wouldn't shocked if atom proved to be faster.

http://groups.google.com/group/alt.games.video.sony-playstation2/msg/62ff83d96ea78ea9?hl=en

That was posted back at launch.

"Right now, from what we've heard, the real-world performance of the Xenon
CPU is about twice that of the 733MHz processor in the first Xbox.
Considering that this CPU is supposed to power the Xbox 360 for the next 4 -
5 years, it's nothing short of disappointing. To put it in perspective,
floating point multiplies are apparently 1/3 as fast on Xenon as on a
Pentium 4. "

 

[taltamir]

But you also said that developers optimize for consoles because it is easier to optimize for them.

I contradicted the latter claim, saying they optimize for consoles because they must, not because its easier. They prefer not to spend the money doing any optimization.

And herein lies the strawman:

console developers can do much more optimization because they only have to target a few combinations of hardware

I stated that more in-depth optimization is easier as in requires less resources. That was a single statement.

Where did you get:

?

Its basic english, this IS what you said. You are trying to BS your way out of it and you can't.
You most certainly did not just state that it is easier with no context whatsover nor any intent nor any further suggestions.

when you say

Heavily threaded, module-conscious software utilizing XOP and FMA4 would work pretty well, considering that console developers can do much more optimization because they only have to target a few combinations of hardware.

You are making several arguments
1. It is easier to optimize a fixed hardware.
2. Consoles get more optimizations.
3. Consoles get more optimizations because they are fixed hardware and that is easier.

You most certainly are NOT making just claim #1 in a vacuum. The very notion is preposterous.

Furthermore, the only OTHER possible explanations as to why they optimize consoles more are:
1. For no reason at all, they just like burning money - preposterous
2. Because they HAVE to optimize it for it to run on such shitty hardware.

And finally, when I threw you a bone and suggested #2 (aka, that you DID just make this statement in a vacuum and that you DID NOT mean anything by it like the other points) you explicitly objected to it; proving that you DID indeed intend to make those other statements. (see post 73)

 

[Joseph F]

256MB of RAM reserved for games 256MB of ram reserved for the OS.

I'm sorry for my OCD, but the PS3 has 256 MB of XDR system memory connected to the Cell, and 256 MB of GDDR3 connected to the RSX. While the Cell and RSX are connected via a high-speed proprietary bus, the Cell can only read the RSX's RAM at a rate of 16MB/s.
(Though, it can write to it at a higher speed, as the chart found here: http://www.***********/playstation3.html#PS3_RSX_GPU shows)

Processor	256MB XDR	256MB GDDR3
Cell Read	16.8GB/s	16MB/s
Cell Write	24.9GB/s	4GB/s
RSX Read	15.5GB/s	22.4GB/s
RSX Write	10.6GB/s	22.4GB/s

 

[taltamir]

I'm sorry for my OCD, but the PS3 has 256 MB of XDR system memory connected to the Cell, and 256 MB of GDDR3 connected to the RSX. While the Cell and RSX are connected via a high-speed proprietary bus, the Cell can only read the RSX's RAM at a rate of 16MB/s.
(Though, it can write to it at a higher speed, as the chart found here: http://www.***********/playstation3.html#PS3_RSX_GPU shows)

Processor	256MB XDR	256MB GDDR3
Cell Read	16.8GB/s	16MB/s
Cell Write	24.9GB/s	4GB/s
RSX Read	15.5GB/s	22.4GB/s
RSX Write	10.6GB/s	22.4GB/s

While interesting, how does this contradict me?

 

[Joseph F]

While interesting, how does this contradict me?

Whoops, I forgot to mention that the PS3 OS only reserves something like 30-40MB nowadays while running a game, I think. I'll try to dig up the exact figure, if you're interested.
Edit: It was 50MB reserved back in 2010, though they've probably optimized it more by now.
http://www.dailytech.com/Sony+Shrinks+PS3+OS+Gives+Devs+70MB+of+Extra+Memory/article17775.htm

 

[Blitzvogel]

Not even close man, these CPU's were dogs new, let alone now. I wouldn't shocked if atom proved to be faster.

http://groups.google.com/group/alt.games.video.sony-playstation2/msg/62ff83d96ea78ea9?hl=en

That was posted back at launch.

"Right now, from what we've heard, the real-world performance of the Xenon
CPU is about twice that of the 733MHz processor in the first Xbox.
Considering that this CPU is supposed to power the Xbox 360 for the next 4 -
5 years, it's nothing short of disappointing. To put it in perspective,
floating point multiplies are apparently 1/3 as fast on Xenon as on a
Pentium 4. "

Bwahaha, I wouldn't believe that for a second.

I like probably many others would liken the Xenon in general performance to a tri-core Atom @ 3.2 GHz, since IIRC, each Xenon core is dual-issue, with SMT. Xenon's big advantage (versus Athlon x2s, Pentium-Ds back in the day), was it's 128 bit VMX units on each core, back when 64 bit FPUs were the norm on the early dual cores.

 

[IntelUser2000]

There's probably a ~10x gap between the PC gpus and the console gpus at this point. I'm not sure the same gap exists between the cpus...

You can't compare the gains made by a specialized ultra parallel one with a general purpose one. 20% gain in GPU is almost nothing, in CPU its almost revolutionary.

Come to think of it, maybe its not that unbalanced. It should be a derivative of Bulldozer cores anyway(Whether its Trinity, Steamroller, or a cut down version of them with less caches for example).

 

[IntelUser2000]

Bwahaha, I wouldn't believe that for a second.

I like probably many others would liken the Xenon in general performance to a tri-core Atom @ 3.2 GHz, since IIRC, each Xenon core is dual-issue, with SMT. Xenon's big advantage (versus Athlon x2s, Pentium-Ds back in the day), was it's 128 bit VMX units on each core, back when 64 bit FPUs were the norm on the early dual cores.

They probably meant that for single threads.

Also the PC chips have two 64-bit FPUs while Xenon had one 128-bit. That puts SP throughput at:

3.2GHz x 4 FP ops/cycle x 3 cores = 38.4GFlops

Pentium 4/Athlon X2: Assuming a middle grade, 2GHz x 4 FP ops/cycle 2 cores = 16GFlops

It's easy to increase flops if you make a chip for that purpose. You can stick 32 Pentium cores running at 1.2GHz to make a 1TFlop capable one like with Knights Ferry.

 

[taltamir]

Whoops, I forgot to mention that the PS3 OS only reserves something like 30-40MB nowadays while running a game, I think. I'll try to dig up the exact figure, if you're interested.
Edit: It was 50MB reserved back in 2010, though they've probably optimized it more by now.
http://www.dailytech.com/Sony+Shrinks+PS3+OS+Gives+Devs+70MB+of+Extra+Memory/article17775.htm

I went and looked up my source again and reread it.
http://www.computerandvideogames.co...ire-large-time-commitment-claims-fallout-dev/
I believe I have msunderstood it the first time. System =! OS. it could be non graphics aspects of the engine for example.

 

[IlllI]

so if the new ps4 switches to x86 and dx11 hardware, what does that mean for backward compatibility with ps3 games?

 

[toyota]

so if the new ps4 switches to x86 and dx11 hardware, what does that mean for backward compatibility with ps3 games?

I thought it was already made clear if that was the case that it would not be backwards compatible.

 

[Fox5]

Not even close man, these CPU's were dogs new, let alone now. I wouldn't shocked if atom proved to be faster.

http://groups.google.com/group/alt.games.video.sony-playstation2/msg/62ff83d96ea78ea9?hl=en

That was posted back at launch.

"Right now, from what we've heard, the real-world performance of the Xenon
CPU is about twice that of the 733MHz processor in the first Xbox.
Considering that this CPU is supposed to power the Xbox 360 for the next 4 -
5 years, it's nothing short of disappointing. To put it in perspective,
floating point multiplies are apparently 1/3 as fast on Xenon as on a
Pentium 4. "

Well, they're still heavily threaded at the least.
Atom about matches Pentium 4 in IPC, but I remember early comments from Capcom putting the Xenon cores at about 2/3rds of a Pentium 4 in IPC. Still, we don't know how well they were using threading, Atom benefits greatly from hyperthreading, to the point where it's almost like having another cpu, Xenon may also. And Cell has actual extra cores...plus, it doesn't take away from the monstrous floating point performance of the two cpus, even if that's not really how games used them. (MIPS seem to be a more accurate comparison for cpus in gaming than FLOPS, imo)
But this also gives credence to the idea that the next Xbox may use an ARM core...and the idea that PS3 could use a Brazos successor.

 

[Tuna-Fish]

I like probably many others would liken the Xenon in general performance to a tri-core Atom @ 3.2 GHz, since IIRC, each Xenon core is dual-issue, with SMT.

In the speed of executing instructions on the CPU, they are probably very close. However, the memory subsystem of the Xenon is a total dog. The caches are slow, the memory pipeline is full of various hazards, and memory latency is just absurdly awful. These hurt really badly in normal object oriented code.

This is the reason Xenon (and the PPU in cell) do so badly in some workloads and so well in others. If you take normal OO code made for a desktop system and run it on a cell, it will very probably lose to an Atom running at a third of the frequency. Simply because the memory pipeline is just so rotten. However, it is possible to write code that works well on the Xenon -- so long as you can treat your memory as a tape instead of doing random access, you can get some really nice throughput out of those vector units.

Unfortunately, there are a lot of workloads that simply don't fit that model. That's why all the rants of how the cpus cannot do AI properly came out when the cpus were released. And they are true -- if you are doing the kind of branchy, memory-limited code that game script and AI are wont to do, the cpu is going to spend 80% of it's time stalled waiting for memory, and that's if you can fit everything in the cache. Once you spill to the main ram, just forget it. You can't fit all the script you need per frame in 1M cache? You're going to have to ship with less script.

 

[Tuna-Fish]

Atom benefits greatly from hyperthreading, to the point where it's almost like having another cpu, Xenon may also.

It does. This is a symptom of the frontend and the memory subsystem being unable to keep the core fed. If your core is always going flat out, adding another thread on it doesn't do you any good. But when it spends most of it's time twiddling it's thumbs waiting for something, putting more threads on it very efficiently extracts more performance.

 

[pelov]

Sorry but I must be missing something here...

Floating point numbers for the CPU within a console are important why? =P There's a reason we have that funny GPU thing that tends to cost more and is far more proficient in that workload. While FP tasks will inevitably be offloaded to the CPU I still think you guys are making the wrong point here. The CPU, whatever it is, will likely have moar coars and/or threads and fair better in integer than FP.

 

[blckgrffn]

Think of being able to make games run well without having really talented guys to optimize your code.

If they can translate more generic hardware into shorter software development life cycles, that will be the biggest news out of this migration, I would think. That will trump well developed current generation engines and backwards compatibility concerns, IMHO.

The big publishers have to just see the $$$ when discussing this, specs be damned.

 

[IlllI]

I thought it was already made clear if that was the case that it would not be backwards compatible.

thanks for such an enlightening and profound reply.

 

[Joseph F]

I went and looked up my source again and reread it.
http://www.computerandvideogames.co...ire-large-time-commitment-claims-fallout-dev/
I believe I have msunderstood it the first time. System =! OS. it could be non graphics aspects of the engine for example.

Exactly.

 

[toyota]

thanks for such an enlightening and profound reply.

well what answer would you like for your question? this article as well as others mention that backwards compatibility would likely be out if this AMD cpu rumor is true.

 

[Blitzvogel]

In the speed of executing instructions on the CPU, they are probably very close. However, the memory subsystem of the Xenon is a total dog. The caches are slow, the memory pipeline is full of various hazards, and memory latency is just absurdly awful. These hurt really badly in normal object oriented code.

This is the reason Xenon (and the PPU in cell) do so badly in some workloads and so well in others. If you take normal OO code made for a desktop system and run it on a cell, it will very probably lose to an Atom running at a third of the frequency. Simply because the memory pipeline is just so rotten. However, it is possible to write code that works well on the Xenon -- so long as you can treat your memory as a tape instead of doing random access, you can get some really nice throughput out of those vector units.

Unfortunately, there are a lot of workloads that simply don't fit that model. That's why all the rants of how the cpus cannot do AI properly came out when the cpus were released. And they are true -- if you are doing the kind of branchy, memory-limited code that game script and AI are wont to do, the cpu is going to spend 80% of it's time stalled waiting for memory, and that's if you can fit everything in the cache. Once you spill to the main ram, just forget it. You can't fit all the script you need per frame in 1M cache? You're going to have to ship with less script.

Hence why any Cell based computing system (outside of the PS3) has been paired with some sort of command processor. Of course, I've heard nothing but complaints about the 1 MB cache for the Xenon, but to be perfectly honest, I've thought the 3x VMX units to be more interesting, since obviously, Xenon was certainly geared to be capable of all the whiz bang physics on top of general game orchestration and scripting without being too substantial in size. I would certainly argue that the Xenon is the better solution for it's designed purpose than the Cell, since it is better balanced and more developer friendly for devs more familiar with x86 CPU design. It's quite impressive what devs continue to do with it, despite being close to 7 years old.