ATI Radeon HD 4000 Series discussion

[Piotrsama]

If it's in production already then the first wafers are done ~2-3months from now. Then couple weeks for assembly, stockpiling and distribution...

Maybe, but I read (can't remember where ) that getting enough GDDR5 at such a fast speed might be a delaying factor.

[AliG]

If it's in production already then the first wafers are done ~2-3months from now. Then couple weeks for assembly, stockpiling and distribution...

I'd say June at the earliest to see it out at newegg, you're forgetting we're waiting on gddr5, the rv770 has probably been ready for a while, ati's probably just making sure everything is top notch before final assembly

[Sly Fox]

This has probably been covered at some point, but nonetheless...

Are these cards just a minor evolution from the 3XXX series (IE: G80 to G92) or could these cards be a large performance boost?

I'm just getting so tired of cards coming out today being... Hardly even a worthwhile upgrade from my old 8800 GTS 640 mb.

I remember when a new generation of GPU's meant suddenly all the games that were virtually unplayable on the last generation suddenly ran at 100 FPS with full AA/AF

Those days need to come back

[Scimitar]

Continuous floating point performance?
Non-hardware reliant AA?
There are a few more but that would just lead this topic even more astray.

Neither one of those is a hinderance to gaming performance. Why would you even want to do software AA, when you can do it in hardware. The G80 is a well balanced card in the things that matter for actual gaming performance.

[AliG]

This has probably been covered at some point, but nonetheless...

Are these cards just a minor evolution from the 3XXX series (IE: G80 to G92) or could these cards be a large performance boost?

I'm just getting so tired of cards coming out today being... Hardly even a worthwhile upgrade from my old 8800 GTS 640 mb.

I remember when a new generation of GPU's meant suddenly all the games that were virtually unplayable on the last generation suddenly ran at 100 FPS with full AA/AF

Those days need to come back

nvidia's initial offering will be even smaller than g80 to g92, but will eventually be forced to reveal "GT200" (whatever the hell it is, $10 says it doesn't work or the yeilds suck like the original r600 designs ). ATI's offering on the other hand will offer anywhere from 40-60% performance increases in general for performance, close to my guess is around a 100% increase for midrange (the 4670 should be a beast compared to the 3670, in fact I place it around 3850 performance), and a decent low end


Eitherway, I don't think we'll be seeing another g80 for a while. The way the market's been behaving dictates that, people are spending more than enough money on g92 for nvidia to invest huge money on replacing it. Not to mention pc gaming has been on a big decline simply because it costs too much to update your pc every 2 years but a consol will last you at least 3-4 and is a lot cheaper too (that and the fact that most pc games these days are crappy consol ports instead of being designed for pc hardware). But once nvidia and ati get into a performance war like intel and amd, then we'll be wowed again

[Donnie27]

If ati wants the market share enough, they'll make a q2 launch happen, which means summer at the latest. I wouldn't be surprised to see the cards in June like they are saying simply because 55nm+the extremely good yields the rv670 has+rv770 is just a suped up rv670=awesome product not too far away

Many sites have been reporting that the rv770's production is coming along amazingly, which means either we have another r600 failure or that we'll be seeing the product right on time like the rv670. I'd be willing to bet on it that the cards will be on time simply for the reasons stated all across this thread.

As for the specs, I'd say they're about right, the rv770 is supposed to be at least 50% faster than the rv670. Judging by the higher clocks, memory bandwidth and 50% more shaders (not to mention 100% more tmus), I'd say that the rv770 may be even more than 50% faster. Of course it all really matters on how well games are coded for it

For that $250 model I'm game! Performance and wattage I don't think I want two of the others sucking 300 Watts

[uwackme]

Nope just crap, as is the NV 9800GTX.

First one to implement 32 ROPs and give us hires w/8xAA in Crysis wins.

Im not buying another card until it has 32 ROPs.

[LordEC911]

Neither one of those is a hinderance to gaming performance. Why would you even want to do software AA, when you can do it in hardware. The G80 is a well balanced card in the things that matter for actual gaming performance.

Ummm... hardware AA IS a hindrance to gaming with certain types of rendering.
Software AA is eventually going to be the future whether anyone likes it or not.

2/3?
Are you referring to the proverbial "missing MUL"? (although then the number would be &#189

It's true that in some contexts nVIDIA says G80 is MADD+MUL. Technically it's true but there's a catch. The "missing MUL" isn't even supposed to handle shader instructions - it isn't even part of the shader core.

Infact, the actual shader ALUs appear to be quite efficiently exploitable as B3D found out:
...Unlike the SPs on R6 of which only upto 2/3 are utilized even in synthetic tests.

Because 2+1 = 3 and 2 outta 3 is two-thirds... aka NOT 1/2.
"SOME" contexts, they have stated that from the get-go as FACT.
Love to see some of these findings of how R600 SPs are only using 2/3 at max in synthetic benches...
Now if you used the words "average" and "real world performance" I might believe you.

Nope just crap, as is the NV 9800GTX.
First one to implement 32 ROPs and give us hires w/8xAA in Crysis wins.
Im not buying another card until it has 32 ROPs.

LMAO! Yes because the number of ROPs on a GPU is all that matters...

[largon]

Because 2+1 = 3 and 2 outta 3 is two-thirds... aka NOT 1/2.
"SOME" contexts, they have stated that from the get-go as FACT.

You're right about the "2/3" in a sense that each MADD does 2 FLOPS. But that's it. For some reason you desperately cling to the idea that the MUL should somehow be counted in as well...

Guess that's like saying the branch unit (6th ALU) in R600 should output measureable FLOPS too.


nVIDIA does not claim G80 has 128 superscalar ALUs, only 128 scalar ALUs. That means they admit each SP is simply just 1x MADD, the MUL is not counted as part of the ALU core. Technically they're correct if they say G80 has 128 MADD+MUL since, sure it does have MADD+MUL. That doesn't mean they all are used for shading. But I said this already. So I have no idea what youre trying to say with the "2/3". The real figure for G80 is likely closer to 2/2. In the end your argument that shaders in R600 are more efficient than in G80 is simply quite silly because it's a given that a scalar GPU architecture is always more efficient than a superscalar one. Superscalar GPUs always face the same problem; shader instructions can not be chopped smaller and smaller to the extent that all sub-ALUs could be used at anywhere near 100% saturation. Scalar chips simply doesn't have this problem.

Love to see some of these findings of how R600 SPs are only using 2/3 at max in synthetic benches...

Sure.

Originally Posted on Techreport
3DMark pixel shader 1600x1200:
R600
223,9 FPS
0,28890322580645161290322580645161 FPS/MHz

8800GTX
329,3 FPS
0,24392592592592592592592592592593 FPS/MHz

->
R600 pushes mere 18% more FPS/MHz, so if G80 would run at near 100% SP saturation in this test then only ~50% of the sub-ALUs in R600 would be in use.

Originally Posted on Techreport
3DMark shader particles 1600x1200:
HD2900
119,4 FPS
0,15406451612903225806451612903226 / MHz

8800GTX
124,6 FPS
0,097185185185185185185185185185185 / MHz

->
R600 pushes 58% more FPS/MHz, so if G80 would run at near 100% SP saturation in this test then only ~2/3 of the sub-ALUs in R600 would be in use.

Now if you used the words "average" and "real world performance" I might believe you.

So you think R600 is sooo awesome because it can pull huge FLOPS in some FLOPS-virus benchmark prog coded just for R600, eh? But when the chip is put to real use - be it 3DMark or games - most of the ALUs sit idle because in real world shader code is not made specifically for R600 and because drivers can't produce enough instructions.

[LordEC911]

Now that this thread is so far off topic, let's try to get it back on track.
The point I was trying to make is that the only part of the R600 architecture that is lacking is the texturing power.
People seem to believe that there is an AA bug that kills performance.
This is not the case, not only was IQ increased over the R580 but so was performance. Shader based AA is also not a cause of R600's poor performance.

With RV770 been an evolutionary step up from the R600, doubling the texture units, increasing the ALU count and tweaking the ROPs a bit (even while keeping the count at 16) should be more than enough for a sub $300 product. Depending on what R700 actually turns out to be, the next quarter should be a good time for ATi until we learn what GT200/G100 is.

[flopper]

Now that this thread is so far off topic, let's try to get it back on track.
The point I was trying to make is that the only part of the R600 architecture that is lacking is the texturing power.
People seem to believe that there is an AA bug that kills performance.
This is not the case, not only was IQ increased over the R580 but so was performance. Shader based AA is also not a cause of R600's poor performance.

With RV770 been an evolutionary step up from the R600, doubling the texture units, increasing the ALU count and tweaking the ROPs a bit (even while keeping the count at 16) should be more than enough for a sub $300 product. Depending on what R700 actually turns out to be, the next quarter should be a good time for ATi until we learn what GT200/G100 is.

hardware and game developing isnt a match and fit all the time.
I will get a ati card this time.

[Scimitar]

Nope just crap, as is the NV 9800GTX.

First one to implement 32 ROPs and give us hires w/8xAA in Crysis wins.

Im not buying another card until it has 32 ROPs.

ROPs play more of a supporting role for a graphics card not a primary role like the shaders. They write processed pixels into the buffer, and handle some AA functions and post processing. You should buy a card that has a good balanced design and not just focus on one particular facet of it.

[Helmore]

ROPs play more of a supporting role for a graphics card not a primary role like the shaders. They write processed pixels into the buffer, and handle AA/AF and post processing. You should buy a card that has a good balanced design and not just focus on one particular facet of it.

And in the R6** the ROPS barely do anything AA related, as this is done by the shaders. Oh and isn't AF mainly done by the TMUs or am I completely off on this one?

[Scimitar]

And in the R6** the ROPS barely do anything AA related, as this is done by the shaders. Oh and isn't AF mainly done by the TMUs or am I completely off on this one?

You're right on that one. AF is by the TMUs. Just looked it up.

[Helmore]

You're right on that one. AF is by the TMUs. Just looked it up.

Which is also the R600/RV670's main weakness. The hit on AA is not really that big on these cards, it's just lacking in it's ability to do AF quick enough.

[AliG]

Exactly why the tmu count was raised and not the rop. Considering that its a superscalar design, the 32 TMUs would have been perfect for the r600 as that would have meant the same 2:1 ratio nvidia used with g80 (which seems to work perfectly). But seeing how the actual stream count has been raised from 64 to 96, I'd say I would have wanted 46 TMUs instead as g80 proved the 2:1 works while anything less offers performance issues.

Regardless, the rv770 will be here months before gt200 and it should bring in a lot of money. Besides, looking at how silent ati and amd were about the original r600 (before the gazillion changes) and k10, and how nvidia is always bragging and leaking out their highend stuff, I'd say they are facing problems with gt200

[Helmore]

I think ATi should keep the Shader/TMU ratio at at least 3:1 as ATi also does the AA in the shaders. Oh and doesn't 480 stream processors equate to 96 shaders? I'm wondering what the amount of SIMDs will be on the RV770 though, but I'm not sure how important that is in terms of efficiency.

[AliG]

But AA is handled by the ROPs in other designs, not TMU, and my bad on the 72, I was thinking 48 (instead of 64) stream due to the 480 shaders and multiplied that by 50% to get 72


Eitherway, performance gains should be huge with the rv770, the extra TMUs will help open up its power a bit and the the higher amount of actual shaders should help a lot since hardly any games are coded to make use of the weird 4:1 ratio

[whocaresbg]

R700 & GT200 specs speculate :

http://forums.vr-zone.com/showthread.php?t=260999

.

[AliG]

dude those are wayyyy off

perhaps gt200 could have 240 shaders, but I doubt it, unless if they are going with a superscalar design similar to that of the r600, 240 is completely out of the question because nvidia 80 tmus wouldn't be enough for a suped up g80 imo (the 8800gtx needs 64 and it only has 128 shaders). Besides no way the die size is over 600mm^2, g80 was rediculous near 500mm^2 and had terrible yields compared to g92 especially the rv670, no way in hell they dare to do that. If they do, ati will simply price the r700 at $400 while gt200 is $700+ (they'll need it be like the ultra's pricing due to the extremely low yields) and even if the r700 is up to 15% slower, people will take two and xfire it any day on an intel platform over a single gt200

besides, why would nvidia go for a 512bit bus when gddr5 is right around the corner? The price of designing and implementing the large ring bus into the die would be no where near worth it, even gddr4 would be enough. Only reason I could see them wanting to use a 512bit bus would be for the 32 ROPs. But either way, I can't picture the core running at even 500mhz with those specs. You'd need stock watercooling for even 400mhz for a die larger than 600mm^2.


As for the rv770, I highly disagree with those too, too many sites have confirmed the 480 shader with 32 TMUs not to mention the use of gddr5 in both ati's and nvidia's performance+high end for this to be correct. Besides, what they have wouldn't be close to a 50% performance increase, in fact that would probably about the same performance with any AF due to extremely low tmu count and low shader count increase

[Helmore]

Maybe you should read the comments, then you'll know that those are just some kind of made up specs.....
Which is just stupid ofcourse

[Metroid]

If Nvidia adopts GDDR3 once again, It will surely be left behind because AMD will adopt GDDR5 as fast as It can. The speed difference between (G3, G5) is too far off. The competiton is here Nvidia and GDDR5 is there to be adopted if not buried on sunshine.

Metroid.

[AliG]

Maybe you should read the comments, then you'll know that those are just some kind of made up specs.....
Which is just stupid ofcourse

of course they're made up, the specs are someone's speculations (speculation being the key word). They can speculate all they want, I just wanted to make it clear they speculated poorly

[LordEC911]

Exactly why the tmu count was raised and not the rop. Considering that its a superscalar design, the 32 TMUs would have been perfect for the r600 as that would have meant the same 2:1 ratio nvidia used with g80 (which seems to work perfectly). But seeing how the actual stream count has been raised from 64 to 96, I'd say I would have wanted 46 TMUs instead as g80 proved the 2:1 works while anything less offers performance issues.

R600 was 4:1 in terms of ALU:Tex
If RV770 really has 96 shaders then it would be 3:1

I think ATi should keep the Shader/TMU ratio at at least 3:1 as ATi also does the AA in the shaders. Oh and doesn't 480 stream processors equate to 96 shaders? I'm wondering what the amount of SIMDs will be on the RV770 though, but I'm not sure how important that is in terms of efficiency.

With 96 shaders it should be kept at 4SIMDs.
The rumored 160shaders was also rumored to be a 5SIMD approach.
The amount of SIMDs are important to efficiency due to the batch size.

[largon]

AliG,
R600 and G80 have the same ALU:TMU -ratio - 4:1.
G80 has *only* 32 TMUs.