Extremely Power Hungry D3D10 GPUs

[Geo]

Indeed we do, and the reality is that ATI have a gaping hole in their product line-up, and NVIDIA have more of my money than ATI!

Why do I feel like we're the blind men who having examined the elephant fully have come to the conclusion that. . oh, NV is kicking ATI's butt in the mid-range. Again. For two years now.

Hey, fair enuf. They got my money last gen with 6800GT, because I liked the features better. First thing I did with my X1900XT was crank the clocks, so I guess I'm the wrong guy to look to for sympathy on the topic subject.

 

[SugarCoat]

Indeed we do, and the reality is that ATI have a gaping hole in their product line-up, and NVIDIA have more of my money than ATI!

X1800XT is certainly just as good if not better then a 7900GT, before you were arguing performance per watt, but now you're talking about a performance hole that doesnt exist. X1900GT and X1800XT are more then enough to compete with the 7900GTs and its right at the same price point if not a few dollars cheaper (street prices). Especially the 512mb X1800XTs which are still floating around at fantastic prices (sub $300). If anything ATIs products in that price range are being hurt by misinformation spread by statements like yours. The 7900GT is in no way alone in its performance bracket nor is it the best performance per dollar.

 

[pakotlar]

I'm not sure how the discussion turned to current gen DX9 cards, but Let me derail us a bit further We are extremely lucky to have two close competitors on the market. If anyone remembers the speed with which 3DFX released parts, and the performance jumps between them, we should be thankful that we have what we do. And what is with the complaints about power consumption? Unless it affects your ability to run it off of your current PSU, who cares how much power it dissipates. Nvidia marketing is getting to some of you I think. Performance/watt or performance/mm^2 isn't very important when the larger, power hungrier part is in the same price bracket (or below) that of the competitor, and delivers faster performance (or sometimes equal performance, but still has benefits in some departments like IQ). The extra heat isn't a selling point, but it isn't a detractor IMO unless it somehow affects your experience negatively. I'm running a x1900xtx and have no problems with heat or noise, and I enjoy the benefits in IQ that it brings. That being said, I've also run a 6800 series card, and that was good too.

edit: And the comment about wanting 70w maximum TDP. What are you talking about? Sure, you can buy that part next gen, but it won't be high end. To expect power consumption to go down and performance to go up is completely unrealistic. Sure, I want a CPU that generates 1W of heat and runs @ Conroe speeds, but that isn't realistic is it? Until my computer starts heating my room up to an asinine degree, I think I could care less how much power it eats as long as the performance is there. Current gen DX9 cards are almost all (from midrange up) consuming more than 70W, but that isn't stopping the people who want performance from buying those cards. Later!

 

[tobbe]

Unless it affects your ability to run it off of your current PSU, who cares how much power it dissipates.

These guys do for example
http://forums.silentpcreview.com/viewforum.php?f=19
as does several posters above too apparently.

 

[EasyRaider]

To expect power consumption to go down and performance to go up is completely unrealistic. Sure, I want a CPU that generates 1W of heat and runs @ Conroe speeds, but that isn't realistic is it?

Conroe does give higher performance at lower power than current CPUs. So why is it unrealistic for VPUs?

Word about noise and power draw gets around. I know that many "regular" people choose NVidia because of it. (Not that they care so much, but to many it seems like the main distinguishing factor.)

 

[SugarCoat]

Conroe does give higher performance at lower power than current CPUs. So why is it unrealistic for VPUs?

Word about noise and power draw gets around. I know that many "regular" people choose NVidia because of it. (Not that they care so much, but to many it seems like the main distinguishing factor.)

Noise... 7900GTX and X1900XTX are damn near identicle in dBA output in typical 3D and Max 3D. The 7900GT like the 7800GT actually have louder profiles do to the size of the cooler working at a higher rpm because the fans are smaller. The noise argument is a little bit of a joke because nine times out of ten the person bringing it up hasnt heard the card they're calling louder, be it Nvidia or ATI. Its a bit sad that this is what its come to. The ATI camp saying, no dont get that cause its got worse IQ. The Nvidia camp saying no dont get that cause your PSU will explode, your power bill will bloat and its REALLY noisy. I'm not sure if this is much better then the days of cheating drivers in that these bickering arguments are just as baseless and silly. It should always be about personal preference, not some bias or otherwise misinformed opinion.

I will comment that for quite a few years people gave the nod to ATI for physical card size, power consumption, and noise output. Nvidia was considered on the worse end of the stick for those issues. But it was never implied as heavily as it is these days that any of those should be the deciding factor of one over the other. Are people getting more picky or just grasping at staws?

 

[pakotlar]

Conroe does give higher performance at lower power than current CPUs. So why is it unrealistic for VPUs?

Word about noise and power draw gets around. I know that many "regular" people choose NVidia because of it. (Not that they care so much, but to many it seems like the main distinguishing factor.)

Sure, Conroe does do that, but it does not lower TDP by much, and much of that difference can be attributed to the process differences (lower voltage, 65nm) as well as design changes. Conroe is wonderful, but there is a limitation to how much you can lower TDP through design changes. You can expect that any architecture based on Conroe (aka same stage config, maybe added width) to raise TDP. Conroe is vastly different than P4. To expect next-gen gpu's to run at a significantly lower TDP than they do now is unrealistic because there will be no fundamental differences between the generations like there is with Conroe and p4 (namely 32stage pipeline vs 14, much lower clocks). Next gen GPU's will fundamentally do 2 things: increase # of ALU's, TMU's, and ROP's. There is no way that increasing these things, along with the additional logic required for DX10, will result in a lower TDP without a process change (which may happen, though this does not guarantee that we will see figures lower than today's TDP on high end cards). That is all I'm saying.

Trust me, I'm all for lower power dissipation, as that has benefits. But expecting it is unrealistic if you want the performance envelope to continue to be pushed the way that nvidia and ATI are doing it. It is usual for the first iteration of a next-gen product to be hot and power hungry, and then revisions of that architecture to become more power-efficient.

 

[pakotlar]

These guys do for example
http://forums.silentpcreview.com/viewforum.php?f=19
as does several posters above too apparently.

Yep, this is something that I overlooked. Although those interested in silent pc's (like those at the website mentioned above) usually go for aftermarket solutions, which in this case will negate any issues of noise, as current GPU's are not so hot as to be unhandelable by large, slow rotating fans offered in some aftermarket coolers. Water is also very popular, and is more than capable of handling the ~100W TDP on these cards.

Also, HTPC users have to consider size and heat, that is true.

 

[KimB]

Part of the tradeoff, yes? We know they've admitted they made AF transistor tradeoffs going from NV3x to NV4x (and NV4x is where we still are). I certainly haven't heard anyone suggest that HDR+AA is transistor-free either.

Oh, I'll claim that, right now. Not completely and utterly transistor-free, but damned close to it.

 

[Geo]

Oh, I'll claim that, right now. Not completely and utterly transistor-free, but damned close to it.

And therefore they haven't done it previously because?

 

[KimB]

I think they did some market research and concluded that most of their customers didn't care if it wasn't going to hurt performance (and it would, with compression algorithms designed for FX8 framebuffers). So they put it at low priority and decided to wait until they could deliver the performance before offering the feature.

 

[FUDie]

I think they did some market research and concluded that most of their customers didn't care if it wasn't going to hurt performance (and it would, with compression algorithms designed for FX8 framebuffers). So they put it at low priority and decided to wait until they could deliver the performance before offering the feature.

You state all this hypothesis as if it were fact. Do you know that nvidia's AA color compression doesn't work with FP16 data? Also, regarding your statement about HDR + AA costing no transistors, did you actually sit down and figure out how many transitors it takes to blend FP16 data vs. FX8 data?

-FUDie

 

[L233]

And what is with the complaints about power consumption? Unless it affects your ability to run it off of your current PSU, who cares how much power it dissipates.

1. noisy heatsink/fan on the video card
2. Heat -> requires better case ventilation -> more noise
3. energy costs

I like my PCs to be as silent as possible. Dustbuster v2.0 aka X1900XT's fan isn't an option and I sure as hell won't void warranty by using an after-market cooling solution. Made that mistake once with my 6800U and was pretty much screwed when the damn thing crapped out a few months later.

 

[KimB]

You state all this hypothesis as if it were fact. Do you know that nvidia's AA color compression doesn't work with FP16 data? Also, regarding your statement about HDR + AA costing no transistors, did you actually sit down and figure out how many transitors it takes to blend FP16 data vs. FX8 data?

Because it wouldn't make sense if the support's not there. It wouldn't be a big change, though. Basically just the different storage size of the pixels would require different addressing to memory, which shouldn't be a big deal in terms of die size or work.

Also, blending functionality is already there, and should be completely orthogonal with FSAA, compressed or no.

 

[FUDie]

Because it wouldn't make sense if the support's not there. It wouldn't be a big change, though. Basically just the different storage size of the pixels would require different addressing to memory, which shouldn't be a big deal in terms of die size or work.

Far more trivial than blending the AA samples.

Also, blending functionality is already there, and should be completely orthogonal with FSAA, compressed or no.

No, I was referring to the blending (filtering) done when the AA buffer is downsampled.

-FUDie

 

[KimB]

Ah, right, the NV4x can do the FSAA downsampling at scanout, thus this would require more transistors in the DAC to keep that same functionality.

But it would always be possible to just do the downsampling via the texture samplers (or whatever hardware is used in MIP map generation), and do it at buffer swap instead of at scanout (which might be better for large framebuffers, anyway).

 

[FUDie]

Ah, right, the NV4x can do the FSAA downsampling at scanout, thus this would require more transistors in the DAC to keep that same functionality.

Quite a few more: FP16 math is not cheap.

But it would always be possible to just do the downsampling via the texture samplers (or whatever hardware is used in MIP map generation), and do it at buffer swap instead of at scanout (which might be better for large framebuffers, anyway).

But that would require that the texture unit be able to read the compressed color data, or that you do a separate pass to decompress the data. A separate pass would cost performance the other would cost more transistors.

-FUDie

 

[KimB]

Quite a few more: FP16 math is not cheap.

Sure, compared to integer math. But if it's a small portion of the die anyway, who cares? This move may or may not be significant, and even if it is, dropping this functionality wouldn't be a huge loss in terms of performance.

But that would require that the texture unit be able to read the compressed color data, or that you do a separate pass to decompress the data. A separate pass would cost performance the other would cost more transistors.

Or that there is a separate decompresser unit that sends decompressed data wherever it needs to go. I do suppose you're right, whether or not there's a transistor cost to provide multisampling with a compressed framebuffer on FP16 would depend upon how the architecture is built now.

But I still contend that nVidia could provide uncompressed multisampling of FP16 framebuffers with next to zero transistor cost. But the performance wouldn't be good.

 

[FUDie]

Sure, compared to integer math. But if it's a small portion of the die anyway, who cares? This move may or may not be significant, and even if it is, dropping this functionality wouldn't be a huge loss in terms of performance.

You have to fetch twice as much data thus you need to handle twice the latency. Seems expensive to me.

Or that there is a separate decompresser unit that sends decompressed data wherever it needs to go. I do suppose you're right, whether or not there's a transistor cost to provide multisampling with a compressed framebuffer on FP16 would depend upon how the architecture is built now.

But I still contend that nVidia could provide uncompressed multisampling of FP16 framebuffers with next to zero transistor cost. But the performance wouldn't be good.

They could also expose AA via super-sampling, but they choose not to and it's not because they aren't able to.

-FUDie