RSX pixel shaders vs Xenos pixel shaders

[ROG27]

The idea that both machines will be very close in performance is a fairly sound expectation when taking into account what is currently known/expected.

But, I guess the point that causes a lot of people to second guess this conclusion is the fact that the PS3 will ship a full year after the xbox360.

For many it is just a little too hard to believe that MS have delievered a machine of smilar power to the PS3 a full year before Sony. Or in the case of this thread, perhaps they find it a little unbelieveable that the silicon nvidia is providing Sony Q4 2006 will be only roughly par with what ATI provided MS Q4 2005.

I guess the hidden objective of this thread and many others like it is to answer how did Microsoft manage such a coupe! What were the tradeoffs. And is there really something we are not missing?

Lack of more expensive CPU and new HD medium (bluray) were the tradeoff.

 

[Gholbine]

I'm not sure why people seem to put the Xbox 360 and the PS3 on equal terms when it comes to performance. In terms of the GPU, as far as I can tell, things end up being pretty close to even. RSX has a raw performance advantage, Xenos has an efficiency advantage, and the argument can't really progress much beyond that.

When it comes to the CPU, however, it seems to me that the Cell is more powerful than the Xenon by a not insignificant margin. The PS3 also seems to have a lot more internal bandwidth than the Xbox 360, considering NUMA offers each device full speed access to a pool of memory (one being XDR which is a little faster than GDDR3), 8 seperate pools of local memory compared to 1 small pool for 6 threads (all of which have to compete with another for execution hardware) and larger bus between CPU and GPU. Am I way off base here, or is this just not in the realm of 'equality'?

I've heard arguments that the Xbox 360's power will be easier to tap, but I beg to differ... The Cell is a more difficult CPU just as the Xenos is a more difficult GPU. We've seen games running on the Xenos, and quite frankly, it struggles with some of the most basic things we've been experiencing on all PC games for some time. This is presumably because developers are still wrapping their heads around the daughter die, predicated tiling, etc. I expect to see exactly the same thing with the Cell in the PS3. Early performance is not up to scratch because devleopers are still grappling with the SPEs and the massive parallelism involved with them. So in this sense, how is the Xbox 360 any easier to program?

The toolset Microsoft have provided is impressive, from what I've heard. This will no doubt aid in development on the platform, but is it really a significant improvement on what Sony is offering? Standard C++ with CG shaders, IBM compiler technology, Collada and OpenGL ES. Is the difference between such offerings enough to offset raw performance advantages the PS3 seems to have?

Someone tell me what I'm missing here (and don't say "a brain" ).

 

[Acert93]

I guess the hidden objective of this thread and many others like it is to answer how did Microsoft manage such a coupe! What were the tradeoffs. And is there really something we are not missing?

Part of the tradeoff was less development time with real silicon. But there is no magic in the industry. While there can be small jumps within a process node as the process matures, the big advances come with process changes. Current processes have ~2 year windows (with GPUs transitioning slower than CPUs). This means if Company A releases a product on a new process immediately upon availability, Company B cannot magically jump ahead 6 months later.

We are seeing this in some of the PS3 rumors. If the PS3 was slated for late Spring 2007 there would be little doubt CELL and RSX would be produced on the 65nm process. Instead they are most likely going to be produced at 90nm.

Power consumption, heat production, yields, etc are hurdles every chip maker deals with. With smart design you can make a superior product, and if you are willing to spend more money you can make compromises in regards to yields (either by making faster and/or larger chips). So there is some wiggle room, but no magic.

I think ERP started this thread off well listing individual areas where chips performed well/poorly. There is no universal winner across the board. This is especially true as game designs can be very different. Instead of focusing on what chip has better flops or cache size, it is more insightful to see how chips benefit certain game designs and techniques. We are long past the days of fixed function hardware, expecting simple "all the time" answers is just unrealistic.

 

[Acert93]

I'm not sure why people seem to put the Xbox 360 and the PS3 on equal terms when it comes to performance. In terms of the GPU, as far as I can tell, things end up being pretty close to even. RSX has a raw performance advantage, Xenos has an efficiency advantage, and the argument can't really progress much beyond that.

What about featureset differences? If you are able to do the same effect, but at 1/2 the cost due to a feature, that adds a significant wrinkle to the "raw numbers". FP10 blending and filtering is a good example, or being able to texture from every vertex shader with a latency of a pixel shader.

What about better performance in certain areas? e.g. RSX has a higher texel fillrate; Xenos has much better pixel shader dynamic branching.

Again, the desire to just bottle them up instead of exploring the differences and how they can impact various game designs, especially future ones, goes beyond, "Oh they are about the same" because RSX and Xenos are very divergent architectures.

When it comes to the CPU, however, it seems to me that the Cell is more powerful than the Xenon by a not insignificant margin.

So you are saying CELL will out perform Xenon is every situation since it is "more powerful" by a significant margin?

 

[patsu]

I guess the hidden objective of this thread and many others like it is to answer how did Microsoft manage such a coupe! What were the tradeoffs. And is there really something we are not missing?

This is going off-topic. If we were to believe the Internet and media...

On the hardware side, High-Defs was cited as the official reason for the delay (6 months ?).

On the software side, the Cell compiler and PS3 SDK needed time to mature according to some report. Backward compatibility is probably still being refined.

On the network side, Sony needed time to jumpstart a XBL-equivalent (or better ?). I recall in some interviews, Kaz also hinted that they wanted to take PS2 and PSP users online too. The focus seems to be on PS3 now.

On the corporate side, they needed to align multiple units.

All in all, Sony spent more time to get to the starting line as a consolidated company effort. They were/are also behind MS in the software and network areas. Their scope is bigger too because Playstation is positioned as the integrated outpost to distribute media.

The good things are: The marketing folks have more time to prepare for the launch. The product folks have more time to package PS3. We are likely to get the same stuff at cheaper price too.

Other unknowns are: The price point, the hard disk, the controller, the power supply, plus how far can 1st-party developers take PS3 at launch.

 

[patsu]

When it comes to the CPU, however, it seems to me that the Cell is more powerful than the Xenon by a not insignificant margin. The PS3 also seems to have a lot more internal bandwidth than the Xbox 360, considering NUMA offers each device full speed access to a pool of memory (one being XDR which is a little faster than GDDR3), 8 seperate pools of local memory compared to 1 small pool for 6 threads (all of which have to compete with another for execution hardware) and larger bus between CPU and GPU. Am I way off base here, or is this just not in the realm of 'equality'?

Yes, the Cell potential seems huge based on its flexible architecture, specs and various live demos. IMHO, it is possible for Cell to outrun Xenos significantly in certain problem classes (and we will get to see them one day). Unfortunately it seems that

(1) The developers (and the SDK ?) needed more time to experiment and parallelize existing/new codebase

(2) Efficient programming on Cell is "hard" due to lack of OoO, more complex memory layout (recode pointer structures to arrays ?), and various latency hiding requirements

(3) Even then, code that can be sped up by Cell constitutes only a subset of the codebase. So the speedup is bounded.

(4) All products have finite development time and resources. So developers may not get all the needed time to optimize for Cell.

(5) It is also possible that the next visual leap of performances require more than the computing differences between PS3 and Xbox360, so the amount of added power in PS3 -- in the grand scheme of things -- may not be sufficient to paint a discernible differences for most people.

Granted, some of these problems are not unique to Cell. It still mean that we will have to wait a few years to see the picture pan out. I hope Sony open up the Linux platform for people like us to program PS3 (for non-games).

Ultimately if we can have very different word processors like Microsoft Word and Apple Pages on "regular" CPUs today, I optimistically believe that we should be able to see different flavor of games on "off-beat" CPUs like Xenon and Cell.

 

[Gholbine]

What about featureset differences? If you are able to do the same effect, but at 1/2 the cost due to a feature, that adds a significant wrinkle to the "raw numbers". FP10 blending and filtering is a good example, or being able to texture from every vertex shader with a latency of a pixel shader.

What about better performance in certain areas? e.g. RSX has a higher texel fillrate; Xenos has much better pixel shader dynamic branching.

Again, the desire to just bottle them up instead of exploring the differences and how they can impact various game designs, especially future ones, goes beyond, "Oh they are about the same" because RSX and Xenos are very divergent architectures.

I acknowledge that they're very different architectures. All I'm saying is that there's really no definitive answer to be had, other than they are close enough to each other such as the difference is insignificant. Is that not the accepted viewpoint?

(RSX is a divergent architecture? I would have thought this was much more true for Xenos.)

So you are saying CELL will out perform Xenon is every situation since it is "more powerful" by a significant margin?

You're right in saying that you really can't condense the performance of a CPU into a single phrase, but I am simply talking about the grand scheme. Yes there's things that Xenon is in a better capacity to execute (namely branch-heavy code), but I think there's just a lot more raw resources there in the Cell than the Xenon. Tapping these resources will be signifincatly more difficult I'd imagine, which is the obvious tradeoff in this instance.

Would it be accurate to say that the Cell and the Xenon are trying to do essentially the same thing, except that the Xenon is a cheaper bandaid solution where the Cell is a more expensive ground-up solution? That's how things appear when you consider cache-locking, VMX-128, etc.

 

[aaaaa00]

Would it be accurate to say that the Cell and the Xenon are trying to do essentially the same thing, except that the Xenon is a cheaper bandaid solution where the Cell is a more expensive ground-up solution?

No it wouldn't be accurate at all. Bandaid solution?!

 

[one]

I guess the hidden objective of this thread and many others like it is to answer how did Microsoft manage such a coupe!

Isn't it all about clever efficiency? Microsoft had the balls and resources to take the risk in some areas.

Let me summarize the common arguments:

1) RSX may have higher theoretical shader power than Xenos, but Xenos is more efficient becauase of Unified Shader. So developers can tap the real world performance of Xenos which is comparable or superior to RSX.

2) RSX may have higher theoretical shader power than Xenos, but Xenos has several traits that are not available in RSX, such as dynamic branching, MEMEXPORT and eDRAM, so the final output of Xenos is comparable or superior to RSX.

3) Transistor budget and semiconductor process are almost the same between RSX and Xenos.

In addition to those relevant to this topic, there are other parallels for this efficiency:

1) Cell may have higher FLOPS than the Xbox 360 CPU, but the Xbox 360 CPU is general-purpose CPU. So developers can tap the real world performance of the Xbox 360 CPU which is comparable or superior to Cell.

2) Sony developer tools are utter crap. Microsoft developer tools win hands down. The resources saved by better tools can be used for more optimization in other areas.

3) Sony network infrastructure is utter crap. Xbox Live infrastructure wins hands down. The resources saved by better network infrastructure can be used for more optimization in other areas.

Yet some argue that even for Microsoft which is known for its software developerment experiences and good developer support the launch time is special and these efficiencies have been almost lost. But it's unlikely that all are lost, facts such that Xenos is a Unified Shader architecture wouldn't go away even if it was in the launch time.

Without all these advantages Microsoft is said to own exclusively, I bet PS3 launch games will be disastrous unlike those of Xbox 360, and even worse than some games in the PS2 launch. Otherwise, something is wrong.

 

[scooby_dooby]

Excellent posts Patsu.

 

[Gholbine]

No it wouldn't be accurate at all. Bandaid solution?!

By 'bandaid' I mean they have taken an already well-established CPU architecture and modified it for improved FLOPs performance, as well as added instructions into an already packed-full ISA. On the other hand, the SPEs, which are the bulk of the Cell, are designed from the ground up to be FLOPs oriented, with their own instruction set, etc.

 

[aaaaa00]

By 'bandaid' I mean they have taken an already well-established CPU architecture and modified it for improved FLOPs performance, as well as added instructions into an already packed-full ISA.

Hardly. That's not a band-aid, that's good engineering -- only reinvent the wheel when you're forced to.

On the other hand, the SPEs, which are the bulk of the Cell, are designed from the ground up to be FLOPs oriented, with their own instruction set, etc.

For another perspective, I'm going to make the argument that CELL is in fact the "band-aid" solution.

Here goes:

<hypothetical, not-completely serious argument>
CELL's design is forced by not having the transistor budget to put 8 real CPU cores in there.

If Sony had the transistor budget, there's no developer in the world that wouldn't instead have asked for an 8 OOOe core traditional SMP crossbar design with lockable L2 cache.

As fabrication technology improves and transistor budget goes up, I expect for the vast majority of computing tasks, the CELL type architecture ("one real core surrounded by cut-down SPEs") will disappear in favor of a design with a number of more sophisticated cores in an SMP configuration. The whole notion of cut-down SPEs will only be reserved for the few applications where the transistor budget prevents you from dropping a full core in there instead.

To put it more clearly, I see CELL as a short-term "band-aid" solution that works around the fact that we can't fit what we really want onto a chip that we can afford to build today.

Whereas Xenon looks closer to what we want a future multicore design to look like -- unfortunately with only 3 available cores and no OOOe -- because that's all that will fit with the technology we have today at the price we're willing to pay.
</hypothetical, not-completely serious argument>

 

[!eVo!-X Ant UK]

Sony network infrastructure is utter crap. Xbox Live infrastructure wins hands down. The resources saved by better network infrastructure can be used for more optimization in other areas.

i love it when people comment on things that hardly anything had been released for.

 

[Kryton]

Okay, you're obviously a fanatic but, please don't just lie to get your point across. A few 'problems' with your argument taken from an entirely neutral view:

Let me summarize the common arguments:

1) RSX may have higher theoretical shader power than Xenos, but Xenos is more efficient becauase of Unified Shader. So developers can tap the real world performance of Xenos which is comparable or superior to RSX.

2) RSX may have higher theoretical shader power than Xenos, but Xenos has several traits that are not available in RSX, such as dynamic branching, MEMEXPORT and eDRAM, so the final output of Xenos is comparable or superior to RSX.

3) Transistor budget and semiconductor process are almost the same between RSX and Xenos.

1) No one has done Unified Shaders before so how can you make such sweeping statements about efficency aside from basing it off benchmarks (from ATi) you also dismiss (from NVidia). Until someone does an independent test this point is mute even for all the argument here either way.

2) Yet again, no one has done anything like MEMEXPORT before. Dyanmic branching already exists in the G70 series (performance might not be great but it certainly is there), get your facts right. eDRAM someone did last generation, can't think who? Perhaps the company who didn't this generation who might understand the tradeoffs?

3) Whoop-de-do. A singing birthday card has as many transistors as a standard calculator. Which is more efficent?

In addition to those relevant to this topic, there are other parallels for this efficiency:

1) Cell may have higher FLOPS than the Xbox 360 CPU, but the Xbox 360 CPU is general-purpose CPU. So developers can tap the real world performance of the Xbox 360 CPU which is comparable or superior to Cell.

2) Sony developer tools are utter crap. Microsoft developer tools win hands down. The resources saved by better tools can be used for more optimization in other areas.

3) Sony network infrastructure is utter crap. Xbox Live infrastructure wins hands down. The resources saved by better network infrastructure can be used for more optimization in other areas.

Yet some argue that even for Microsoft which is known for its software developerment experiences and good developer support the launch time is special and these efficiencies have been almost lost. But it's unlikely that all are lost, facts such that Xenos is a Unified Shader architecture wouldn't go away even if it was in the launch time.

Without all these advantages Microsoft is said to own exclusively, I bet PS3 launch games will be disastrous unlike those of Xbox 360, and even worse than some games in the PS2 launch. Otherwise, something is wrong.

1) Neither Xenon or Cell is general purpose; in fact no CPU is (what is general purpose code?). They are extremely specialised to deal with multi-threaded vector arithmetic (i.e. the 128 VMX registers in Xenon and the amazing floating point performance of an SPU) and have dropped OO execution to gain this capability. Branch performance on Cell may be worse but, because the domain is so closed (games) the hints may actually prove to be better than the extra silicion needed to boost it.

2) Sony is using standards and technology that was used in PS2. SN Systems tools are well known to PS2 devs and I'd be suprised if a developer had never used GCC before. All platforms are using established toolsets not just MS with VisualStudio packaged as XNA (VS is mainly a PC oriented tool that I doubt console-only developers would have touched). The rest of Sony tech from Nvidia (OpenGL ES with CG vs. DirectX with HLSL) and IBM (just another compiler) is also pretty standard stuff.

3) Here I can only relate to the Sony figures but, if the network is so crap why do they have more players online? It doesn't take a genius here to notice that MS uses a peer-to-peer system for hosting games which costs them nothing aside inital setup (PC games have been free because, they too, are esentially the same) - all that subscription money pays for is centralised statistics! Sony don't have to do anything special aside making a decent GUI, say it's free and XBL will stop looking so attractive.

 

[Shifty Geezer]

Okay, you're obviously a fanatic...

Great, mature, intelligent start to your post...

1) No one has done Unified Shaders before so how can you make such sweeping statements about efficency aside from basing it off benchmarks

Just look at the principles of the tech, and it's obvious there are going to be efficiency advantages. When Xenos needs to do all vertex, no pixel work, it was 48 vertex pipes to use. When a conventional GPU needs to do all vertex work, it has a handful of vertex pipes and the pixel pipes sit idle.

2) Yet again, no one has done anything like MEMEXPORT before. Dyanmic branching already exists in the G70 series (performance might not be great but it certainly is there), get your facts right. eDRAM someone did last generation, can't think who? Perhaps the company who didn't this generation who might understand the tradeoffs?

You're just being silly. One was outlining the potential benefits of the different architectures. What has Sony having used eDRAM before got to do with that?

3) Whoop-de-do. A singing birthday card has as many transistors as a standard calculator. Which is more efficent?

I presume both, as they're very specialised hardwares where every transistor contributes to it's limited function. World's apart from GPU efficiency considerations where usage patterns can mean large parts of the transistors sit idle or not, depending on the architecture.

1) Neither Xenon or Cell is general purpose; in fact no CPU is

What a ridiculous thing to say. There's no such thing as a general purpose CPU! My Athlon is not specialised at running an particular type of code, but turns it's hand to all sorts. In what way is it specialised and not general purpose?

(what is general purpose code?).

Read the forum and the discussions already had in this area.

2) Sony is using standards and technology that was used in PS2. SN Systems tools are well known to PS2 devs and I'd be suprised if a developer had never used GCC before. All platforms are using established toolsets not just MS with VisualStudio packaged as XNA (VS is mainly a PC oriented tool that I doubt console-only developers would have touched). The rest of Sony tech from Nvidia (OpenGL ES with CG vs. DirectX with HLSL) and IBM (just another compiler) is also pretty standard stuff.

I think you need to research this info a bit more. A couple of obvious faults (IIRC) is that SN ssystems isn't yet working for PS3, or if it is was only released recently. From dev comments it's not a patch on MS's tools either. PS3 doesn't use OpenGL ES either, but a custom varient OpenGL PS or somesuch. Similar, but it'll still have differences I think. And the fact you have a standard compiler doesn't make something easy to program for. For that you need understanding of the hardware, which Cell makes more difficult than Xenon.

3) Here I can only relate to the Sony figures but, if the network is so crap why do they have more players online?

Because they outnumber XB's 5:1? Percentage wise PS2 has a far smaller user base than XB. Of course, it's debatable whether this can be attributed to the online service or other factors.

 

[Farid]

Thread Locked: Redundant Discussion

So, since this thread is both a versus thread and a derailed thread (from RSX PS VS Xenos PS to Cell VS Xenos), I think that we can safely say that this thread deserve some rest.