ATI Radeon HD 4000 Series discussion

[STaRGaZeR]

You dont get it

I got it the first time you exposed it

I have my reasoning for my estimated score, I would like to see yours.

I'm talking about something I think you haven't taken into consideration in your numbers: the fact that R600 and derivatives lack in texture power the most. And that 3DMark06 is very shader intensive. You can't use 3DMark06 as a good bench to cover all areas of the new design because the enhancements done will be used mainly in gaming perfomance with AA and AF (where TMUs are most needed). I agree with your 25-30% boost in 06, because that is what 06 utilizes most: shaders. You'll have 50% more shaders, so 25-30% is realistic. You'll also have 100% more TMUs. Combining this to the weak perfomance of 3870 in games specially with AA and AF caused by the only 16TMUs, I expect up to 80% here depending of the game of course. But even if you have more of A, B and C you can't predict how A, B, and C will work together. This is pure speculation.

I see that RV770 has a huge potential for games compared to RV670, but in 3DMark I don't think you'll see more than your 25-30%. My bet is that benchers will be a little disappointed with the chip, but gamers will enjoy it a lot if the price is right.

That is my reasoning based on the current RV770 specs rumours. And all of this is comparing 3870 with 4870, I'm not taking NVIDIAs into consideration

[AliG]

eitherway, 3dmark means nothing unless you're a world record bencher like kinc, kingpin, fugger, shamino, etc, that's why the 8800gt is preferred over the 3870 for single card setups. **hopefully** the TMUs will help unleash some of the cards potential and bring it to a respectable gaming card. I have no doubt that it can beat the 9800gt, the real questions are whether it can put some pressure on the 9900gts and how well can the r700 compete with the 9900gtx

[perkam]

I got it the first time you exposed it


I'm talking about something I think you haven't taken into consideration in your numbers: the fact that R600 and derivatives lack in texture power the most. And that 3DMark06 is very shader intensive. You can't use 3DMark06 as a good bench to cover all areas of the new design because the enhancements done will be used mainly in gaming perfomance with AA and AF (where TMUs are most needed). I agree with your 25-30% boost in 06, because that is what 06 utilizes most: shaders. You'll have 50% more shaders, so 25-30% is realistic. You'll also have 100% more TMUs. Combining this to the weak perfomance of 3870 in games specially with AA and AF caused by the only 16TMUs, I expect up to 80% here depending of the game of course. But even if you have more of A, B and C you can't predict how A, B, and C will work together. This is pure speculation.

I see that RV770 has a huge potential for games compared to RV670, but in 3DMark I don't think you'll see more than your 25-30%. My bet is that benchers will be a little disappointed with the chip, but gamers will enjoy it a lot if the price is right.

That is my reasoning based on the current RV770 specs rumours. And all of this is comparing 3870 with 4870, I'm not taking NVIDIAs into consideration

Your point is valid, plus I made a mathematical error, hence my low numbers

Here are the corrected figures based on 25% increase

If you get 10k stock currently with the 3870, I'm predicting you'll get 12.5k stock with a 4870.
If you get 11k stock currently with a 3870, I'm predicting you'll get 13.75k with a 4870.
If you get 12k stock currently with a 3870, I'm predicting you'll get 15k ( ) stock with a 4870.

Now that looks a lot better, and is something we can agree on

Perkam

[Mungri]

Sorry, I cant agree on any predictions based on a card that isnt released yet.

And the 4870's do have seperate shader clocks, I read that somewhere on teh interwebs.

I am Expecting 4850 to be equal to the 8800 GT, the 4870 to be equal to a 9800 GTX. Just because I said so. Therefore it is true ^^

Single 4850 = 13k+ marks in 3D mark 06
Single 4870 = 17k+ marks in 3D mark 06

Lets just wait a month and see who is right :p

[AliG]

I highly doubt that will be the case. If anything, if the 4870 scores 17k+ in 3dmark (which won't likely happen at stock, but very possible for oc), the 4850 would be around 15~16k as they're the same card with slightly different clocks&/memory

[MAS]

something tells me we have to wait only a week or two weeks...

[zerazax]

this might be the quietest release (rumor-wise) for an ATI card in a while...

[Mungri]

I highly doubt that will be the case. If anything, if the 4870 scores 17k+ in 3dmark (which won't likely happen at stock, but very possible for oc), the 4850 would be around 15~16k as they're the same card with slightly different clocks&/memory

Oh ok then...

13-14k for the 4850,
15-16k for the 4870



Those predictions are with the card in my PC though, I.E overclocked.

[Beefy22]

I got it the first time you exposed it


I'm talking about something I think you haven't taken into consideration in your numbers: the fact that R600 and derivatives lack in texture power the most. And that 3DMark06 is very shader intensive. You can't use 3DMark06 as a good bench to cover all areas of the new design because the enhancements done will be used mainly in gaming perfomance with AA and AF (where TMUs are most needed). I agree with your 25-30% boost in 06, because that is what 06 utilizes most: shaders. You'll have 50% more shaders, so 25-30% is realistic. You'll also have 100% more TMUs. Combining this to the weak perfomance of 3870 in games specially with AA and AF caused by the only 56 bit external memory16TMUs, I expect up to 80% here depending of the game of course. But even if you have more of A, B and C you can't predict how A, B, and C will work together. This is pure speculation.

I see that RV770 has a huge potential for games compared to RV670, but in 3DMark I don't think you'll see more than your 25-30%. My bet is that benchers will be a little disappointed with the chip, but gamers will enjoy it a lot if the price is right.

That is my reasoning based on the current RV770 specs rumours. And all of this is comparing 3870 with 4870, I'm not taking NVIDIAs into consideration

You make some goods points... But what about the fact that the Render Back ends(Similar to ROPs) is still just 16 (for both RV670 and RV770)? All that Pixel processing power, So why skimp on rendering the actual pixels? (The last question was not a rhetorical, I'm looking for a good explanation)

- 480 shader units; 50% increase
- 32 texture mapping units; 100% increase
- 16 render back ends; 0% increase

[Mungri]

I think I lost it with the technical side of graphics cards after they got more complicated then memory bandwidth, pixel pipelines, textures per pass, and shaders.

I know nothing else about stuff from after the X1900's lol.

Too complicated now.

[STaRGaZeR]

You make some goods points... But what about the fact that the Render Back ends(Similar to ROPs) is still just 16 (for both RV670 and RV770)? All that Pixel processing power, So why skimp on rendering the actual pixels? (The last question was not a rhetorical, I'm looking for a good explanation)

- 480 shader units; 50% increase
- 32 texture mapping units; 100% increase
- 16 render back ends; 0% increase

With my extremely limited knowledge about GPUs architecture I can't answer to that. But one thing is sure: NVIDIA needs lots of ROPs because AA is done in the ROPs. In RV670 AA is done in the shaders, so they don't need as much. But how many? 16 seems kinda low, and I wonder why they left it like it was before. We have to wait. In R600 and RV670 the 320:16:16 ratio doesn't work well because of the TMUs, will the design be balanced with 480:32:16? Also, the shaders in this architecture are far from being fully loaded when gaming.

[Mungri]

I read on the 4800 spec sheet that they will have 32 ROPs.

[STaRGaZeR]

I read on the 4800 spec sheet that they will have 32 ROPs.

Where? All previous information says only 16 ROPs.

[whocaresbg]

The ROP:TMU ratio always had been 1:1 in the ATI's high-end .

[Scimitar]

So? That can easily change. They aren't linked to each other.

Post that spec sheet.

[LordEC911]

You make some goods points... But what about the fact that the Render Back ends(Similar to ROPs) is still just 16 (for both RV670 and RV770)? All that Pixel processing power, So why skimp on rendering the actual pixels? (The last question was not a rhetorical, I'm looking for a good explanation)

- 480 shader units; 50% increase
- 32 texture mapping units; 100% increase
- 16 render back ends; 0% increase

That is where you are wrong, hopefully...
There is a LARGE increase with the ROPs even while keeping them the same number. They should be tweaking them so that they perform roughly twice as fast while staying at the same clockspeed.
Plus the fact that they, if the independent clocks is indeed false like I think it is, are at a 1050mhz core clock which is a ~35% increase.
This is why I doubt they will be going with separate clock domains, it would hurt the TMU and ROP performance at the expense of a slight increase in shader performance, which isn't the bottleneck in the first place.

If the above is correct we should see a, theoretical, max increase of ~170% for the ROPs.
The same ~170% increase for the TMUs.
A ~103% increase in shader performance.
Plus a 80% increase in memory bandwidth, in regards to the 4ghz GDDR5.

16ROPs is just fine if they do indeed have the 1050mhz clock and are tweaked to double the z performance.

[STaRGaZeR]

Ei, ei, ei, wait a second.

There is a LARGE increase with the ROPs even while keeping them the same number. They should be tweaking them so that they perform roughly twice as fast while staying at the same clockspeed.

What? Twice? You better have some proof of that, I'm calling total BS.

Plus the fact that they, if the independent clocks is indeed false like I think it is, are at a 1050mhz core clock which is a ~35% increase.

I also think the independent clocks are fake. But there won't be a 1050MHz chip, that's for sure.

If the above is correct we should see a, theoretical, max increase of ~170% for the ROPs.
The same ~170% increase for the TMUs.
A ~103% increase in shader performance.
Plus a 80% increase in memory bandwidth, in regards to the 4ghz GDDR5.

Wow, you've done the perfect GPU in just 4 lines. Congrats

Never gonna happen man, be serious and stop dreaming.

[Beefy22]

That is where you are wrong, hopefully...

We'll see... I think this is going to be a great series of cards for AMD. But I think they're going to hit the market over-hyped and quite possibly not live up to the mounting expectations. -Case in Point -> http://www.guru3d.com/news/ati-radeo...hmark-results/

There is a LARGE increase with the ROPs even while keeping them the same number. They should be tweaking them so that they perform roughly twice as fast while staying at the same clockspeed.

If the above is correct we should see a, theoretical, max increase of ~170% for the ROPs.

performing twice as fast while maintaining the same number of RBEs ≠ 170% increase in performance... Nor do I expect to see independent clocks, but I could be wrong.

[AliG]

I think that amd should multiply the core speed by at least 2 to get the shader clocks as g80 (especially g92) proved that extremely high clocked shaders will be more useful for today's games than massive numbers of them. Besides if you have the TMUs and ROPs running at those clocks how could it hurt performance?

[jaredpace]

Imo your reasoning is flawed here. There is no way to predict the effect of a faster bus on the 3670. There is no way you can know whether its lack of TMU/ROPs is going to allow it to have anything more than minimal gains from a faster bus. Additionally what the 4870 should bring is improved real world gaming performance as it seems to address what appears to be the biggest bottleneck in the R600 architecture for many games

is this true?
What was the bottleneck & solution?

[Kai Robinson]

The problem was 16 TMU's. They've doubled them as the solution.

[Eastcoasthandle]

x1800xt
321 million transistors on 90nm
Pixel fill rate 10000 Mpixels/s
Texel fill rate 10000 Mpixels/s
625/1500


x1900xtx
384 million transistors on 90nm
Pixel fill rate 10400 Mpixels/s
Texel fill rate 10400 Mpixels/s
650/1550

x1950
384 million transistors on 90nm
Pixel fill rate 10400 Mpixels/s
Texel fill rate 10400 Mpixels/s
650/2000


HD2900
700 million transistors on 80nm
Pixel fill rate 11900 Mpixels/s
Texel fill rate 11900 Mpixels/s
742/1650


3870
666 million transistors on 55nm
Pixel fill rate 12400 Mpixels/s
Texel fill rate 12400 Mpixels/s
775/2250

(let me know if any of the above needs to be changed)

4870
___ million Transistors on __nm
Pixel fill rate ___________________
Texel fill rate ____________________
___/____


Who's willing to fill in the blankets?

[LordEC911]

4870
___ million Transistors on __nm
Pixel fill rate ___________________
Texel fill rate ____________________
___/____


Who's willing to fill in the blankets?

~850m on 55nm, ~250mm2
16.8GPixel/s
33.6GTexel/s
1050/4000mhz

And for those of you that don't understand what I am saying about the ROPs, go back and read the thread. They will be tweaking the ROPs to perform more efficient.

[NH|Delph1]

1050/4000mhz

1050 is Theo's wishful thinking

You can't trust him on evaluating ATI performance, the guy is blinded by his love for them

//Andreas

[alexio]

4870
1337 million Transistors on 13.37nm
Pixel fill rate 1337 leets/s
Texel fill rate 1337 leets/s
1337/GDDR13-37


Who's willing to fill in the blankets?

Just some speculation, but who knows it's true?