Most Significant Graphics Technology for 2005

[bigz]

apart from the fact that both IHV's bleeding-edge parts were not available in any volume, and still aren't if we're realistic now.

I keep hearing that, but I still also keep remembering how after only about 20 minutes of comparing I was able to order & get delivered an X800 pro VIVO I wanted within 42 hours.

I know the real XT PEs are rare, but it wasn't hard to get a hold of a card and make your own.

Much the same with the 6800GT/Ultra too... but if you wanted to go out and buy the bleeding-edge video cards and not overclock/mod them, you had a bit more of a problem on your hands.

 

[whql]

The low end inflection point is "base" DX9 functionality. (SM 2.0). When nVidia and ATI move SM 3.0 into the low end is about as technologically important a gargle blaster.

What about floating point filtering/blending? This is only currently available on SM3 hardware.

You really think this is going to be important on low end hardware? Look at the huge performance hit there is in Far Cry on a 6800 GT/Ultra, its not going to be a good on low end hardware is it? Pixel performance and bandwidth would need to double for the performance to stand still, and thats not likely to happen anytime soon at the entry level.

Its not available on all sm3 hardware either - none of the 6200's have it.

 

[rwolf]

Most Significant Graphics Technology for 2005: 90nm!

That would be my choice as well.

 

[Joe DeFuria]

The low end inflection point is "base" DX9 functionality. (SM 2.0). When nVidia and ATI move SM 3.0 into the low end is about as technologically important a gargle blaster.

What about floating point filtering/blending? This is only currently available on SM3 hardware.

You really think this is going to be important on low end hardware? Look at the huge performance hit there is in Far Cry on a 6800 GT/Ultra, its not going to be a good on low end hardware is it? Pixel performance and bandwidth would need to double for the performance to stand still, and thats not likely to happen anytime soon at the entry level.

Its not available on all sm3 hardware either - none of the 6200's have it.

My thoughts exactly.

And forget about the extra "features" of SM 3.0, I'm thinking that for low-end hardware, even having FP32 support would basically be a mistake. Too little "reward" for the cost of implementation.

I think the low end should skip SM 3.0 altogether, and just go from basic DX 9 support (SM 2.0) and then jump to WGF support when ready.

 

[KimB]

You really think this is going to be important on low end hardware? Look at the huge performance hit there is in Far Cry on a 6800 GT/Ultra, its not going to be a good on low end hardware is it? Pixel performance and bandwidth would need to double for the performance to stand still, and thats not likely to happen anytime soon at the entry level.

Its not available on all sm3 hardware either - none of the 6200's have it.

1. There's something wrong with FP blending/filtering on the 6800 parts. In some benchmarks, the 6800 Ultra is beaten by the 6600 GT. In fixed implementations, the performance hit for using a FP16 framebuffer should be somewhere in between 2x AA and 4x AA.

2. From what I understand, the 6200 is a 6800 with parts of the core disabled, and as such may not be representative of a future low-end part.

 

[KimB]

I think the low end should skip SM 3.0 altogether, and just go from basic DX 9 support (SM 2.0) and then jump to WGF support when ready.

Except nVidia isn't going to do this. They don't have any architecture that supports only basic DX9 functionality, and they're not going to redesign their architecture for that purpose. It makes much more business sense for nVidia to produce a low-end NV4x part than to create a whole new architecture.

It also makes more business sense for nVidia because low-end parts are primarily OEM parts, and as such the support of new features means more than for retail parts.

 

[Xmas]

1. There's something wrong with FP blending/filtering on the 6800 parts. In some benchmarks, the 6800 Ultra is beaten by the 6600 GT. In fixed implementations, the performance hit for using a FP16 framebuffer should be somewhere in between 2x AA and 4x AA.

I only know one benchmark where that's the case, and I'm not sure that was a correct comparison. The performance hit should be beyond that of 4xAA because of the missing compression.

2. From what I understand, the 6200 is a 6800 with parts of the core disabled, and as such may not be representative of a future low-end part.

6200 without TurboCache is a NV43 with some parts disabled. 6200 with TC is a NV44. Neither card supports FP16 filtering and blending.

 

[digitalwanderer]

OMG, I can't believe we all forgot the most important thing to come out: smartshaders!

 

[Joe DeFuria]

I think the low end should skip SM 3.0 altogether, and just go from basic DX 9 support (SM 2.0) and then jump to WGF support when ready.

Except nVidia isn't going to do this. They don't have any architecture that supports only basic DX9 functionality...

Not sure what you're point is.

The fact that nVidia doesn't have a part with "basic" DX9 functionality doesn't make an introduction of a SM 3.0 part in the low end (assuming they do this at some point), "revolutionary" or even important.

..., and they're not going to redesign their architecture for that purpose. It makes much more business sense for nVidia to produce a low-end NV4x part than to create a whole new architecture.

Since they don't have a fP24 part, you may be right. But that would only mean that nVidia made the wrong choice in not offering a "basic" DX9 part to begin with.

It also makes more business sense for nVidia because low-end parts are primarily OEM parts, and as such the support of new features means more than for retail parts.

I disagree.

Support for new marketable features means a lot to OEMs. About as much as brand recognition, compatibility / stability, and cost.

I still don't see SM 3.0 being any big marketing check-box feature touted by OEMs.

For example, put two otherwise identical systems next to each other: one markets SM 3.0, and one markets PCI-Express....which one do you think gets sold more?

 

[KimB]

I only know one benchmark where that's the case, and I'm not sure that was a correct comparison. The performance hit should be beyond that of 4xAA because of the missing compression.

You're assuming primarily bandwidth limitation. I'm assuming that the chips are going to be more limited by memory space and ROP's. But I suppose you're right in respect to the 6200, so it doesn't really matter.

 

[KimB]

Since they don't have a fP24 part, you may be right. But that would only mean that nVidia made the wrong choice in not offering a "basic" DX9 part to begin with.

Why would offering nothing but DX9 parts that offer more than simple PS 2.0 be a bad thing?

For example, put two otherwise identical systems next to each other: one markets SM 3.0, and one markets PCI-Express....which one do you think gets sold more?

I don't get your point. All of the sub-$200 parts in the NV4x lineup have been PCI-Express native. I expect this will continue as cheaper parts are offered.

 

[KimB]

OMG, I can't believe we all forgot the most important thing to come out: smartshaders!

Smartshaders are novelty toy, nothing more.

 

[Sinistar]

OMG, I can't believe we all forgot the most important thing to come out: smartshaders!

Smartshaders are novelty toy, nothing more.

sort of like ps 3.0 an a 6800.

 

[Geo]

OMG, I can't believe we all forgot the most important thing to come out: smartshaders!

Well, if we were talking about 2004 rather than 2005, I'd say the most important thing was NV getting back in the game.

The most *interesting* (to me, at least), was that it seemed pretty clear that pre-release both companies were feeling pretty smugly certain that they were going to deliver a "clean kill" this round. . .and each got a nasty surprise.

I still wonder if ATI really intended to open with a 16-pipe part, or if they were intending that for their "refresh" as they stocked up on those low-yield parts and got their hand forced by the shock that NV40 delivered. I don't think they believed that NV could jump from 4 pipes to 16 in one generation. They certainly were very intent on making NV go first on releasing what they had up their sleeve.

 

[kemosabe]

My thoughts exactly.

And forget about the extra "features" of SM 3.0, I'm thinking that for low-end hardware, even having FP32 support would basically be a mistake. Too little "reward" for the cost of implementation.

I think the low end should skip SM 3.0 altogether, and just go from basic DX 9 support (SM 2.0) and then jump to WGF support when ready.

I tend to agree from the business perspective, though this begs the question of what ATI is going to do with its next low-end part. Somehow I doubt they'll stick to the RV370 core for all of 2005, despite its current OEM popularity in the PCI-E segment. Do they go with a SM2.0 RV410-based chip, or will there be a SM3.0 R520-based offering coming next year? Despite what we may argue, ATI brass might be tired enough of having the SM3.0 marketing propaganda being thrown in their faces that they may not want to risk losing any of that freshly captured OEM business because of it.

On a related front, since there apparently doesn't seem to be a non-low-k 0.09u process at TSMC, would these entry-level parts continue to be fabbed on 0.11u or does UMC have a lower-cost 0.09u process that might be more attractive? Any thoughts, Dave?

 

[MuFu]

TSMC's 0.09u process might actually be viable from an economic viewpoint - all the fabs are producing 12" wafers and on top of that there will be more dice for a given wafer size than 0.11u. They have the capacity and early yields are reportedly good. Although Nexsys appears to be performance-oriented, on the grand scale of things so are most of the fab processes selected for producing graphics ASICs. Besides, it's a "family" of processes we're talking about and I'm sure there are options that cater for more mainstream performance and yield targets.

ATi seem to be extremely confident in it, having taped out at least one large ASIC at this node already (and R500 must be on its way soon, if it hasn't taped out already), with no prior experience from a less complex die.

Just found this re UMC's 0.09u process if anybody's interested (might have been posted already): CHIPWORKS SEES UMC'S DIFFERENT APPROACH TO 90NM TECHNOLOGY. IS IT 90NM?

 

[Dave Baumann]

I tend to agree from the business perspective, though this begs the question of what ATI is going to do with its next low-end part. Somehow I doubt they'll stick to the RV370 core for all of 2005, despite its current OEM popularity in the PCI-E segment.

I have a strong suspicion that over the course of the R(V)5xx generation you'll see a similar number of R(V)5xx ASIC's as you see for the Radeon Xxx0 series.

On a related front, since there apparently doesn't seem to be a non-low-k 0.09u process at TSMC, would these entry-level parts continue to be fabbed on 0.11u or does UMC have a lower-cost 0.09u process that might be more attractive? Any thoughts, Dave?

80nm and 90nm UMC will be options later on.

 

[digitalwanderer]

OMG, I can't believe we all forgot the most important thing to come out: smartshaders!

Well, if we were talking about 2004 rather than 2005, I'd say the most important thing was NV getting back in the game.

I was joking about smartshaders, and I agree that parity 'tween ATi & nVidia is the most signifigant thing this year.

 

[MuFu]

I tend to agree from the business perspective, though this begs the question of what ATI is going to do with its next low-end part. Somehow I doubt they'll stick to the RV370 core for all of 2005, despite its current OEM popularity in the PCI-E segment.

I have a strong suspicion that over the course of the R(V)5xx generation you'll see a similar number of R(V)5xx ASIC's as you see for the Radeon Xxx0 series.

On a related front, since there apparently doesn't seem to be a non-low-k 0.09u process at TSMC, would these entry-level parts continue to be fabbed on 0.11u or does UMC have a lower-cost 0.09u process that might be more attractive? Any thoughts, Dave?

80nm and 90nm UMC will be options later on.

I can understand why they might go for 90nm@UMC for entry-level RV5x0, but would have thought that an 80nm "sub-node" might be something reserved for memory production only (and barely publicised as a result). Are 80nm chips something you've seen in your crystal ball, Dave?

 

[kemosabe]

Fab favours? There would surely have to be some incentive for UMC to hold onto ATI's mainstream business.