Is HDR "Free" for the Xenos?

[BenQ]

It's my understanding that AA is considered "Free" because of the way the daughter die is compartmentalized aside from the rest of the GPU. You can turn on AA and it will have little, if any, affect on the rest of the system/gaming performance.

Is that correct?

But what about HDR?

I know the ROP's on the daughter die are responsible for HDR rendering, but is everything that's done on the Daughter die considered "free"?

A quote from the Xenos article.....

"The ROP's can handle several different formats, including a special FP10 mode. FP10 is a floating point precision mode in the format of 10-10-10-2 (bits for Red, Green, Blue, Alpha). The 10 bit colour storage has a 3 bit exponent and 7 bit mantissa, with an available range of -32.0 to 32.0. Whilst this mode does have some limitations it can offer HDR effects but at the same cost in performance and size as standard 32-bit (8-8-8-8 ) integer formats which will probably result in this format being used quite frequently on XBOX 360 titles. Other formats such as INT16 and FP16 are also available, but they obviously have space implications."

I don't fully understand what all that means. Is Dave saying that 32 bit HDR will be free, and that 64 bit HDR can be done aswell, but with a fillrate penalty?

How "Free" is HDR on the Xenos?

 

[jvd]

I don't fully understand what all that means. Is Dave saying that 32 bit HDR will be free, and that 64 bit HDR can be done aswell, but with a fillrate penalty?

not with 4xfsaa . But it will not increase the hit that 4x fsaa gives . So fp10 hdr +4x fsaa should provide 5% or less performance hit . fp 16 will take fillrate , how much isn't clear

 

[BenQ]

I don't fully understand what all that means. Is Dave saying that 32 bit HDR will be free, and that 64 bit HDR can be done aswell, but with a fillrate penalty?

not with 4xfsaa . But it will not increase the hit that 4x fsaa gives . So fp10 hdr +4x fsaa should provide 5% or less performance hit . fp 16 will take fillrate , how much isn't clear

So you're saying that fp10 HDR and 4X's MSAA will have a total performance hit of 5% or less?.... because I would consider that to be "free."

 

[jvd]

fp10 should be the same as int 10 . So there shouldn't be any hit . That is what the big deal is in regards to it .

 

[The GameMaster]

The "Special" FP10 HDR mode is basically free as it has the same performance and memory metrics as the standard FP8 mode and as the 4xFSAA is also basically free then in that context... yes XENOS provides "Free" HDR + 4xFSAA. There are certain limitations to that special FP10 HDR mode... but... developers have a lot of options to which mode they can use.

Going to simplify this... FP8 is what is standard today and is the baseline for performance and memory usage, the standard FP10 mode has a small hit to performance and requires a bit more memory storage. The standard FP16 mode has an significant hit to performance and requires quite a bit more memory storage. The "Special" FP10 mode on XENOS allows for the benefits of FP10, but with restrictions on the alpha channel (transparency) with the same performance and memory storage as the standard FP8 mode. Additionally... XENOS can perform all of these modes with FSAA active unlike the Geforce 7800/RSX which can not (it can do FSAA or it can do HDR, but not both at the same time).

It's up to the developer what is going to be used and each mode has advantages and disadvantages... again part of XENOS's flexibility.

*REMOVED INACCURATE SECTIONS* I misstated what those modes are exactly.

 

[BenQ]

When you talk about fp8/fp10/fp16 are you talking about a mode that is specifically for HDR?

 

[DemoCoder]

Normal FP8 Mode : INT8 per color channel INT8 for alpha (8x8x8x8)

There is no FP8.

Standard FP10 Mode : INT10 per color channel INT10 for alpha (10x10x10x10)

This would require 40bits per pixel instead of 32-bits. I doubt this mode exists, and it is not INT10.

Special FP10 Mode: INT10 per color channel INT2 for alpha (10x10x10x2)
Standard FP16 Mode : INT16 per color channel INT16 for alpha (16x16x16x16)

These are not "INT10" and "INT16". They are float formats, with mantissa and exponents, and IEEE float style semantics.

 

[Shifty Geezer]

EDIT : This was addressed to furod (sp?) whose post was deleted while I wrote. :?

It depends what you mean by free. Xenos is capable of rendering at 720p without AA, and then rendering that same scene with 2xAA with negligable performance hit.

I see you've arrived on this forum with a view that the 'Free HDR' is a misnomer, much the same way I ranted about the 256 GB/s figure being misrepresented. I think that your arguments are being knocked back because of the way you're presenting yourself. As a forum trying to maintain intelligent discussion against a pressure of troublemakers seeking to be inflammatory, newbie posts especially are viewed skeptically. And a misinformed newbie presenting a POV as fact comes across as being a troublemaker spreading FUD and deliberately trying to mislead. Arriving on a forum populated but (in part) quite smart and knowledgable people who accept that AA is free and telling them they're wrong isn't likely to go down well

It can often come down to the difference between use of passive and assertive language. Basically it's better to arrive asking questions then making statements! If this is a point you contend, try asking a question how AA is free on Xenos given the costs that you perceive, and you'll get answers.

 

[fouad]

EDIT : This was addressed to furod (sp?) whose post was deleted while I wrote. :?

It depends what you mean by free. Xenos is capable of rendering at 720p without AA, and then rendering that same scene with 2xAA with negligable performance hit.

I see you've arrived on this forum with a view that the 'Free HDR' is a misnomer, much the same way I ranted about the 256 GB/s figure being misrepresented. I think that your arguments are being knocked back because of the way you're presenting yourself. As a forum trying to maintain intelligent discussion against a pressure of troublemakers seeking to be inflammatory, newbie posts especially are viewed skeptically. And a misinformed newbie presenting a POV as fact comes across as being a troublemaker spreading FUD and deliberately trying to mislead. Arriving on a forum populated but (in part) quite smart and knowledgable people who accept that AA is free and telling them they're wrong isn't likely to go down well

It can often come down to the difference between use of passive and assertive language. Basically it's better to arrive asking questions then making statements! If this is a point you contend, try asking a question how AA is free on Xenos given the costs that you perceive, and you'll get answers.

Ok thanks, I will consider what you are saying...
And saying that the free AA on xbox360 is a relative term, depending on what we mean, is totaly true ! and this is my opinion, because if free AA means free 4x AA at 480i or 480p than this is true, there is free anti aliasing on xbox360.
and if 2x AA at 720p done with little hit performance is considered as free AA than yes the xbox360 has free AA !

BUT the problem is that people dont understand this, they dont understand that free AA on xbox360 is more of a marketing compaign than anything else, because the way free AA on xbox360 works will lead us to consider even RSX doing free 2x AA at 720P !!!

Even microsoft knew this, proof ?
http://www.bit-tech.net/bits/2005/06/10/richard_huddy_ati/1.html
Richard Huddy said this in an interview :
"We practically have AA for free on the PC anyway right now. If the difference between 1280x1024 with no AA and 1280x1024 with 2x AA is 90 FPS and 70FPS, who wouldn’t turn the AA on? The performance hit isn’t going to be noticeable to most gamers – and with an X800 or X850 those kind of frame rates are common place.â€￾

 

[Shifty Geezer]

That's a marked difference to Xenos's situation. As PC software gets more complicated those frame rates are going to drop. The G70 slides from nVidia that included the statement 'AA for Free' showed that on older, simpler software, this is true - AA didn't affect performance. But on modern games and future games, AA penalities were as high as ever.

The point with Xenos is no matter how mighty the shaders you run, or how complex the models, adding 2xAA makes no difference. This is isolated in a seperate functional unit of the hardware and isn't tied into the shader performance, unlike conventional systems.

So on a PC card, a game from 2 years ago will run at 99% FPS at 2xAA, but a game in 2 years from now will run at 60% FPS, say. Whereas on Xenos, it'll run 99% with 2xAA both the older games and new games.

This difference between this system (eDRAM backbuffer) and something like the G70 is that ATi used a third of available transistor space for AA and other smapling functions. This means less space for shaders. If G70 did the same, it would mean lower peak shader abilities in all situations. By not having this eDRAM, G70 can have more pixel shading power per pixel - stronger than Xenos without AA, but weaker with AA

The picture for Xenos approach vs. G70 approach is complicatd though by Xenos also using unified shaders which, though having fewer of them, occupying less space than G70's shaders, may be able to equal performance through efficiency. If so, the end result is the shader performance of a big budget conventional card, while AA has no impact on this rendering performance (at least at limited AA amounts).

 

[overclocked]

Even if you read the great article Dave did on Xenos its still based on papers and words from ATI how good it is.
I tend to be cautious because if you can judge the performance of a part from the numbers and papers both Matrox Parhelia and NV30 would have rocked. 8)

 

[London Geezer]

Hey i won't tolerate Parhelia bashing on here!!!!

Anyway, as i said, according to NVIDIA and ATI, we've had "free" AA since the Geforce3... So excuse me for not being overly optimistic about having "free AA" in the next gen. It will work, it will be fast, but it's never gonna be free. Until Jesus comes down from the sky and makes it free.

 

[bloodbob]

So you're saying that fp10 HDR and 4X's MSAA will have a total performance hit of 5% or less?.... because I would consider that to be "free."

So if something takes 50% of less I could call it free x 10?

 

[Shifty Geezer]

Anyway, as i said, according to NVIDIA and ATI, we've had "free" AA since the Geforce3... So excuse me for not being overly optimistic about having "free AA" in the next gen. It will work, it will be fast, but it's never gonna be free. Until Jesus comes down from the sky and makes it free.

The term 'Free' is what's misleading. If they called it 'hard-coded' would that settle the issue? It's not free as it's cost the system. It's cost them transistors that could be used for other areas. (yep, we've had this debate before )

They had a choice - all shader rendering power, or 66% shader power, 33% AA power.

It's like...I dunno...Sony could claim Cell has 216 GFLOPS for free. That is, they designed a system and where it has x integer performance, it also has that FLOP performance. But it wasn't free as if they went without the FLOPS they could have added more Integer OPS.

Think of i thte other way round. On a G70 you have x rendering power and run it add 100%. When you add AA, you're still running the ard at 100% though frame rate drops. But on Xenos, with AA your using 100% of the card, but without AA your using 66% of the card. Not having AA costs you efficiency of the hardware. They've prebuilt that performance to the point it's handled nicely, but to maximise card efficiency you need to turn it on. So it's not really free in that respect, but an inbuilt feature that MS requested.

Howveer, on the other side of the term, AA is free because you can use it without affecting the frame rate. It's free in terms of rendering potential, but not free in terms of what it costs to implement.

I can see this being a bone of contention that no-one can agree on

 

[Acert93]

The term 'Free' is what's misleading. If they called it 'hard-coded' would that settle the issue? It's not free as it's cost the system. It's cost them transistors that could be used for other areas. (yep, we've had this debate before )

Run Shifty, come back after the rant

/begin rant

I never did get involved in that flaming thread, but I might as well say it here:

The individuals argueing, "It is not free!!!" and called it misleading have intionally took what ATI said completely out of context and twisted it to be fodder for the console debate. All ATI ever said was that from a performance perspective that 2x AA had no performance hit and 4x AA had a 1-5% performance hit and could realistically be called "free" from a performance perspective.

AA was never said to be "free" from a transistor budget perspective. That was a made up arguement by Sony apologists to downplay the effects of the eDRAM--which have been significant on the forum.

There has been no end to 2 subject themes on B3D since E3:

1. Downplay the CELL SPEs in the PS3
2. Downplay the eDRAM / Unified Shaders in the Xbox 360

Sony and MS fans are equally guilty. Ironically, if people stopped and thought about how unique each design was and how their design structure is built around these philosophies I think people would appreciate them more and how each is unique and plays to their strengths.

Hint: Note where the memory controller is on each system. One is CPU-centric, the other GPU-centric. The CPU-centric design has a revolutionary CPU, the GPU-centric design has a revolutionary GPU.

That is not to say that a thoughtful individual, like Shifty, cannot bring up realistic concerns about a design (e.g. unified shader performance or ease of use, tiling performance hit, will FP10 have significant artifacts, how realistic is it for developers to cleanly serialize their code to fit into 7 hardware threads that will perform adequately on the SPEs, and so forth).

But as Shifty has noted, a lot of discussions have veered away from what was actually said, or the more evaluative stance of information released, to more:

MS Fan: Poo Poo on Sony news/media
Sony Fan: Poo Poo on MS news/media

That includes leap from, "AA is free from a performance perspective on C1" to "AA is not free! It took transistors". That was a silly leap in logic that ATI/MS never made. IMO it is fair to discuss the question like thus:

- "Was 80M transistors a good tradeoff to prevent AA, Z, Alpha, and other framebuffer tasks from being a system bottleneck when those transistors cost potential shader realeastate/performance?"

That prevents fan- boi wars. While the following is 1. twisting facts, and 2. invites an arguement and is really not a stance to promote a useful discussion:

- "AA is not free on the C1"

Some call it semantics, but then again, semantics is what this forum is all about

/end rant

That's a marked difference to Xenos's situation. As PC software gets more complicated those frame rates are going to drop. The G70 slides from nVidia that included the statement 'AA for Free' showed that on older, simpler software, this is true - AA didn't affect performance. But on modern games and future games, AA penalities were as high as ever.

Pretty fair summary of the G70 info. No one knows exactly what tweaking has gone into RSX, but at 550MHz and linked to CELL it should be a strong performer--much stronger than the G70. My guess is between a closed box and a tweaked design the AA performance will still have a hit at higher resolutions on modern games but less so than G70 shows.

I know, talk about the master of the obvious here

I can see this being a bone of contention that no-one can agree on

Why agree on something when we can spend the next 6 years argueing about it? Calling a "truths" before a system even launches? pffffft!!



IMO both companies have divergent philosophies, but with similar foundations. The PPC CPUs and NV/ATI GPUs are the similar foundations. Ditto 512MB memory, sound processing on the CPU, and so forth. The big difference, IMO, is that MS's strengths may show up sooner on the screen, at least in areas regarding the framebuffer (AA, stencil shadows, Z, alpha blends). The unified shaders, and how much vertex work to offload to the CPUs, and the ins and outs will take more time. Sony will get a good shot in the arm from a fairly standard GPU, but wont get the immediate perks the eDRAM offers. Devs will master that with time (some even at launch), but where the PS3 will really begin to shine is when developers begin compiling libraries of code that is threaded to work well on the SPEs. My guess is last gen PS3 games will look so much better than 1st gen PS3 software we wont recognize them coming from the same system. It is going to take developers time to learn what code can be designed to make the best use of the system, and then to serialize it for the 7 SPEs, but once that is accomplished the games will look like night and day.

Sony and MS are really putting forth a great first step this generation with the best hardware, compared to the market, we have ever seen. Like all console launches there are hurdles, concessions, and dissappointments, but overall I do not think we could have asked for anything more. MS and Sony fans are going to be in for a treat over the next 5 years.

And the divergent designs will allow all of us to argue which was better for the next 10 years!

 

[Johnny Awesome]

Excellent post Acert.

 

[Shifty Geezer]

And the divergent designs will allow all of us to argue which was better for the next 10 years!

Once again I disagree with you utterly. I'm sure there's at least 15 years worth of arguing possible out of this

Seriously though that, somes it up nicely Acert. I'm inclined to write a summary of the system overviews as I think people have lost sight of the systems amongst specific numbers and sound bites. Given they both have a similar transistor budget to spend, and both have the same problems regards getting the most out of the hardware, you can look at the architecture as different approaches. It could crudely be summed up in terms of brute power (PS3) versus efficient design (XB360).

 

[scooby_dooby]

Ya, but the potential shading power of the unified shaders is still relatively a mystery is it not?

Or do we know for sure the unified shader appraoch will not be a match for the the 24 shader pipelines in the RSX?

This may be the case, I've seen you guys count numbers before but they are way over me head, so I'm just wondering is pretty much accepted the RSX's standard shader approach will out-perform the unified shaders on the Xenos?

 

[BenQ]

I don't understand why some of you have issue with the notion of AA being "Free" for the Xenos.

Obviously "free" doesn't mean that it's done magically and without transistors. "Free" obviously means that it can be done without affecting system performance.

Even if it's not 100% free, and rather 95% free....

1. Atleast it's predictable in the sense that regardless of what the rest of the system is doing and regardless of the complexity of the game, it's will always be 95% free.

2. IF it IS truly 95% free, then that's close enough to free for me to be comfortable saying that it IS "Free." To be able to claim a mere 5% performance hit under ANY circumstances on ANY game no matter how complex is FAR FAR more than any other GPU can claim.

 

[Shifty Geezer]

Or do we know for sure the unified shader appraoch will not be a match for the the 24 shader pipelines in the RSX?

If this is in respones to my crude summing (and it is crude), I'm talking system architecture, not just GPUs. PS3 has more of stuff - more processors, more FLOPS, more BW between components, whereas XB360 has more stuff-saving features so it doesn't need as much stuff, like not needing as much BW as a high demand of BW is shipped to the eDRAM.

Regards shader performance it's still unknown. Without real-world Xenos measurements and RSX insight ti'll remain that way.