What are the key tradeoffs ATi may be considering with R420?
- Thread starter g__day
- Start date
YeuEmMaiMai
Regular
I believe that ATi will continue along the path of giving us the best bang for the $$$. I do not see them making too many changes to their ansio and fsaa routines. I believe that they may indeed make the R420 full FP32 just to be able to say that they have it along with the speed to effectively use it. A good chunk of the R300 already uses FP32 internally so I do not see any problems adding it to the rest.
ATi has already stated that the intend to reuse as much of the R3X0 as possible in future designs........
ATi has already stated that the intend to reuse as much of the R3X0 as possible in future designs........
Luminescent
Veteran
What if Ati makes it's pixel alu's fast/vast enought to implement texture filtering via fragment programs? This way, they could concentrate on the hardware's future-proof shading speed rather than waste transistors on legacy filtering that will be eventually phased out.
YeuEmMaiMai said:
I don't concur - I think more pressing features and performance will be top of mind until they switch to 90nm on later 2004 and ready themselves for DX10 in mid 2005. They have to make design trade-offs, a full fp32 bit shader pipeline I presume would eat up a hell of alot of additional transistors for little marginal gains with today's software and 3d APIs. Its simply not best bang for buck. But full VS 3.0 amd PS3.0 compliance and heighten performance would be king hits. So I reconned that will be where they invest - rather than introducing any marketing feature - think "cinematic rendering" or "final fantasy in real time" - before its ready for adoption. I think NVidia showed them very clearly not to do that!
more intersting insights scattered by ATi here http://www.elcoyote.dk/ATi.html
[maven said:
How odd, I had earlier today submitted a detailed summary on how many vector ops it takes to do a simple trilinear read/blend. :? Because I don't want to type again, it takes about 60 operations. The key point is that something like this works great in a highly specialized low-precision, parallelized and pipelined environment rather than having to go through the fragment shader. If it did have to go through the PS, then we would be able to calculate a filtered texel, one every two clocks.
g__day said:
The most I got out of that was:
rick bergman
We generally don't discuss unreleased products but rest assured, the ATI team is cranking away on a broad range of products. We will be releasing an entire family of products in the first half of 2004. While I can't give any specifics, ATI has been the clear technology and performance leader for quite some time and our next generation won't disappoint our fans.
They also mention moving to 90 nm in Q4 2004 (value edition of R420 I suppose to give them experience for a 90 nm R500 in Q1 2005 or possibly slightly earlier).
ANd they infer they will stick closely to Directx Next spec - which I infer means for R420 they won't go readically beyond DX9 specifications - hence no fp32 pixel shaders as a prioirity directing how they affect their transistor budget.
ANd they infer they will stick closely to Directx Next spec - which I infer means for R420 they won't go readically beyond DX9 specifications - hence no fp32 pixel shaders as a prioirity directing how they affect their transistor budget.
My bold. 1st half 2004 means:
1. we'll see at least one r420 value card
2 It won't be 90nm
I think we'll have something like 9500* with full vertex shaders, one "quad" turned off if it is 12X1 or 2 quads if it is 16X1. Considering it is a big die they must have a considerable quantity of bad chips and they have the technology to reuse them. Hell, for R420 they could have designed a 16X1, switched off the "worst" quad by default, sell it as a 12X1 and have a great yield.
That would make the low-level card a 4X1 so 9600 is here to stay. maybe update it to ps3.0. they don't have a reason to move it to 90 nm unless they would fit a vs 3.0 unit, which is not quite a stringent requirement on a low-level card.
Edit: same die.
nVidia was saying very, very similar things before the launch of the NV30, and nVidia had a much longer history of high-performance, full-featured products than ATI currently has.geo said:
nVidia still has a longer history of high-performance, full-featured products, products that have, on more than one occasion (as ATI has), surprised and delighted 3D graphics enthusiasts.
YeuEmMaiMai said:
I agree with you about the "not changing their FSAA implementation that much" part. But i would be surprised if they didn't do anything about the "angle problem" in their aniso implementation. I also doubt that they'll add FP32 support since it's not required for VS/PS 3.0. Maybe for specific operations, not that i'm in the position to tell you what those operations would be though
Bjorn said:I also doubt that they'll add FP32 support since it's not required for VS/PS 3.0. Maybe for specific operations, not that i'm in the position to tell you what those operations would be though
Have they reached the point where what is left on the FP32 front is all-or-nothing, or can they usefully (i.e. make a difference in IQ) extend FP32 parts of their architecture without finishing the job?
(Note, I know nothing about R420's details as yet)
Personally, I'm going to be highly surprised if there is any development on the parts that are outside of the PS/VS3.0 specification. Given that this wasn't even on the roadmap 12 months ago I'm scratching my head to see how they can upgrade the shaders, increase performance and move a massively large design to a smaller silicon process in 12-18 months in the first place.
I can see a case for minimising the risks as much as possible, which may include not changing FSAA and filtering properties. There is probably room to play with their AA implementation in the first place (larger caches hence more samples), however I'll be surprised if we see much more.
Personally, I'm going to be highly surprised if there is any development on the parts that are outside of the PS/VS3.0 specification. Given that this wasn't even on the roadmap 12 months ago I'm scratching my head to see how they can upgrade the shaders, increase performance and move a massively large design to a smaller silicon process in 12-18 months in the first place.
I can see a case for minimising the risks as much as possible, which may include not changing FSAA and filtering properties. There is probably room to play with their AA implementation in the first place (larger caches hence more samples), however I'll be surprised if we see much more.
YeuEmMaiMai
Regular
I have a problem with what the NV3X offers performance wise once you enable all of the supposidly superior techniques. I along with many other people would MUCH RATHER have ATI's trade-offs and get AWESOME PERFORMANCE compared to Nvidia's trade-off with much inferior performance. It is really sad that nVidia hyped this FP32 crap and yet they cannot even deliver it at half of the speed ATi does FP24
Chalnoth said:
I'm an idiot consumer. I want things to be loud, colorful and shinyAiluros said:
I honestly hope that that isn't just a tad of nostalgy; I want from both sides the fiercest possible competition. I'm a smart consumer; I like things fast and cheap
I want to be on the winners' side so I can crack jokes about the losers in the pub
