Do you think full hardware scheduling in the shader cores is worthit?

[Anarchist420]

I do. It makes programming easier for both drivers and gamedevs... there is less need to optimize for specific applications and an opt for one app may break another... they'll have to have too many driver releases when they could've just gone with Hyper Q (or whatever it was called) in all GK parts.

second paragraph: I suppose a downside is that the transistor space could be used for more raw/peak power via more shader cores, but full hardware scheduling may help sustained (expected) performance more and it is a superior use of resources.

What I mentioned in the first paragraph should also illustrate that "compute" features do not necessarily exclude "gaming" features... it's really a false dichotomy and I wish everyone would quit trying to say stupid shit like "nvidia improved gaming performance by stripping out the compute features". Nvidia didn't really do much to increase the quality of gaming in the future because if they had gone without such a severe reduction in double precision performance then that could mean better forward+ rendering.

I think AMD's approach to the future was wiser, but if double precision compute won't be used with forward+ on the 7970's GCN cores, then the 7970 fails.

I know I'm extremely biased in the pro-IQ/Compatibility direction, but someone has to be in a world where too many people are satisfied with the crap they're given.

 

[Blitzvogel]

Isn't hardware scheduling one of those things that is extremely difficult to do since AMD and Nvidia don't work via a common ISA like x86, hence much of the reasoning for Larrabee?

 

[nforce4max]

I agree and would love to see this become mainstream as well standard but with Nvidia it will never happen. The only way AMD can make this work is to get Microsoft, Intel, HP (Matrox), and a few other smaller companies like VIA that owns S3 on board to establish some common compute standards. If all else Sumsung and try to push Radeon GCN cores in third party ARM procs. Nvidia likes setting its own proprietary standards and lock everyone else out to maintain a very strong monopoly on the compute market as well workstation market.

 

[BenSkywalker]

I'm still curious as to what you want to see DP used for in real time rendering?

The level of accuracy DP offers is *far* beyond what our displays are capable of showing us, you would need something that could display billions of colors- and have the capability of differentiating between them using your eyes on a per pixel basis. I'm open to hearing what DP is useful for, but I am not seeing its' benefits for real time 3D.

As far as the hardware scheduler- absolutely not. One of the most profound advantages GPUs have over x86 is they don't have to deal with clunky and cumbersome legacy support. While the way we do things today may work great today, five years from now we may be using a very different approach and we would then be looking at either breaking backwards compatability or ballooning die sizes to support multiple methods.

 

[Pottuvoi]

I'm guessing something like object placement rotations, perhaps z-buffer in space exploration genre..
I certainly do not see any common use for fp64 for pixel shaders, fp32 is quite a lot already.

 

[raghu78]

Do not forget GCN is an architecture which is going to drive the next generation of consoles. Hardware scheduling in the shader cores means you can expect the best possible performance at a slight die cost. thats a very reasonable trade-off. Making things easier for game developers means developers can focus their time and effort on raising the bar in graphics technology and in other areas of game development. With a clean sweep of the next gen consoles AMD is looking at GCN being the foundation for cutting edge console game development for a long time 2013 - 2020. AMD's decision to build an architecture with good compute performance is also very sensible because more and more games are making heavy use of compute shaders.

 

[Anarchist420]

I'm open to hearing what DP is useful for, but I am not seeing its' benefits for real time 3D.

Emulation. What do you think about extended precision?

For example, a depth buffer of 32 bit linear FX may not be enough, even if logarithmic, because it may not have enough to hold the biggest picture with enough accuracy (at least not for the best IQ) or it may be too linear... 64 bit (double precision) would be more values so there is linearity and enough to hold the entire scene without having to get creative.

I think 40 bits of precision for fixed and floating point would be sufficient.

Billions of color displays are becoming more common. The extra ten bits could be used for good alpha blending. Additional cycles or perhaps more/better instructions would then allow for good edge AA or an emulation of well-patterned 4x RGSS.

32 bits can be enough if there is a good instruction set and if the programming is done right, but we have too many lazy programmers.

 

[Pottuvoi]

Billions of color displays are becoming more common. The extra ten bits could be used for good alpha blending. Additional cycles or perhaps more/better instructions would then allow for good edge AA or an emulation of well-patterned 4x RGSS.

We get the extra ten bits from a going to fp32, no game currently uses fp32 frame buffer.

For AA and blending a proper programmability of ROPs would be a lot more usable than fp64 and is something that programmers have been asking for a long time.

 

[BenSkywalker]

For example, a depth buffer of 32 bit linear FX may not be enough, even if logarithmic, because it may not have enough to hold the biggest picture with enough accuracy (at least not for the best IQ) or it may be too linear... 64 bit (double precision) would be more values so there is linearity and enough to hold the entire scene without having to get creative.

When discussing GPUs, FP16 is 64bit, FP32 is 128bit. What you are arguing, if you know it or not, is that 128bit isn't enough and we need 256bit

Billions of color displays are becoming more common.

No LCD is close to that, no matter what they claim(if they say they are, they are liars). The human eye can't pick up that many distinct colors anyway, we top out around 32 million(very few displays produce close to that).

The extra ten bits could be used for good alpha blending.

Alpha blending is a hack we needed when we didn't have the resources to handle the geometry. We have the resources now, they need to go away.

Do not forget GCN is an architecture which is going to drive the next generation of consoles.

I have seen nothing that supports this rumor- even the basis for the rumors if it is assumed is 100% accurate is what dev kits have in them. The 360's dev kit had a R4xx derivative this far out from launch, it shipped with R500- the PS3 dev kit was using a software based rasterizer for quite a while.

AMD's decision to build an architecture with good compute performance is also very sensible because more and more games are making heavy use of compute shaders.

Is that coming out next generation? Good is always relative when talking about performance, their compute performance looks good only when compared to the pathetic current nVidia generation. We have already seen that two of their current top chips barely edge out a single GK110- and use ~50% more power to do so. When we are looking at Q4 2013 the compute performance of their current parts is going to be somewhere between 'terrible' and 'lol'.

I'm guessing something like object placement rotations, perhaps z-buffer in space exploration genre..

For something like a space exploration game with so little OD just dump the Z buffer altogether. You could either go to W or even use painter instead. I know we take certain things as a given today because we've been doing them for so long, but it isn't like we don't have other approaches we could use that are superior in certain situations.

 

[raghu78]

Is that coming out next generation? Good is always relative when talking about performance, their compute performance looks good only when compared to the pathetic current nVidia generation. We have already seen that two of their current top chips barely edge out a single GK110- and use ~50% more power to do so. When we are looking at Q4 2013 the compute performance of their current parts is going to be somewhere between 'terrible' and 'lol'.

Consoles don't use 550+ sq mm graphics chips. they use very highly power and die size optimized designs. You forget that GK110 is a 550+ sq mm die. AMD HD 7970 is a last gen product. At 1 Ghz the HD 7970 has a SP performance of 4 TFLOPS and DP performance of 1 TFLOP. GK110 is the next gen Nvidia product which will power the GTX 780. Lets compare the HD 8970 with GTX 780 for perf/sq mm and perf/watt when both are available. If you want to compare perf/watt for gaming optimized products you should look at GTX 760 Ti vs HD 8870. so don't make judgements without any products.

 

[BenSkywalker]

If you want to compare perf/watt for gaming optimized products you should look at GTX 760 Ti vs HD 8870. so don't make judgements without any products.

/facepalm

That is *exactly* what I was saying