AMD: R7xx Speculation

[Jawed]

I don't discount any of this. I was speaking purely from the perspective of functional unit utilization, referring to ATi's traditionally-lower SP utilization rates (particularly in the R6xx vs. G8x/G9x generation).

In G8x/9x the MUL was idling when not used for transcendental calculations or attribute interpolations. Of course, trying to quantify that is pretty hard...

Jawed

 

[ShaidarHaran]

In G8x/9x the MUL was idling when not used for transcendental calculations or attribute interpolations. Of course, trying to quantify that is pretty hard...

Jawed

Good point. We've been leaving the MUL out when considering the efficiency of G8x/G9x, so if you count all the SP units, you could say ATi has higher utilization rates.

 

[trinibwoy]

Good point. We've been leaving the MUL out when considering the efficiency of G8x/G9x, so if you count all the SP units, you could say ATi has higher utilization rates.

Not sure what numbers you would base that on. How do you know how often the SF/interpolation unit is not doing SF/interpolation work? In general it depends on whether you consider G80's MUL available for general shading and give R6xx's interpolation unit a free pass as a non-programmable ALU.

 

[AlexV]

Not sure what numbers you would base that on. How do you know how often the SF/interpolation unit is not doing SF/interpolation work? In general it depends on whether you consider G80's MUL available for general shading and give R6xx's interpolation unit a free pass as a non-programmable ALU.

The MUL was marketed as available for general shading. R6xx's interpolation unit was never marketed in any way, and even less presented as part of the shader core, AFAIR.

 

[ShaidarHaran]

Not sure what numbers you would base that on. How do you know how often the SF/interpolation unit is not doing SF/interpolation work? In general it depends on whether you consider G80's MUL available for general shading and give R6xx's interpolation unit a free pass as a non-programmable ALU.

No numbers (other than marketing FLOPs).

 

[rwolf]

You guys didn't learn your lesson from the 2900 fiasco.

Same pre launch hype all based on rumors. Everyone looked stupid launch day.

Guess we are all in for a repeat.

This time it looks like the shoe is on the other foot.

 

[Shtal]

40 TMUs ?

ATI R4xx/R5xx/R6xx - 3 generation stayed with 16 TMU's.

I would assume ATI already realized by now "I hope" that TMU's became limited factor.

I got a feeling we will see 40 TMU's + 800SP's, sound about right.

 

[fellix]

ATI R4xx/R5xx/R6xx - 3 generation stayed with 16 TMU's.

...and ATi made it to this very day alive and in well enough shape.

 

[Shtal]

...and ATi made it to this very day alive and in well enough shape.

R600/RV670 ??

 

[chavvdarrr]

ATI R4xx/R5xx/R6xx - 3 generation stayed with 16 TMU's.

I would assume ATI already realized by now "I hope" that TMU's became limited factor.

I got a feeling we will see 40 TMU's + 800SP's, sound about right.

no way, I bet they'll say "16 is enough"

 

[Arnold Beckenbauer]

16 extreme TMUs.

 

[ShaidarHaran]

16 extreme TMUs.

ROFLMAO! You deserve a cookie for that.

 

[Pete]

So OBR said he couldn't post #s: "i am not under NDA but posting of result before NDA is not too fair for AMD ... and i cannot to do it." Nedjo, another poster there, seems to know what's up and said it was pressure from OBR's source that's stopping him from spilling the beans early (1, 2).

But another poster said OBR posted this elsewhere:

http://pctforum.tyden.cz/viewtopic.php?p=7610020#7610020

E8400@3.2GHz, 2GB DDR3@1600MHz, X38 2D - 135W, 3dMark06 - 222W, HD4850 ~ 12 000 3DMarks

this is what OBR posted at another forum

Edit: link fixed.

 

[AlphaWolf]

the link in the quote gives a 404.

 

[Tim]

So, with 800 SP the RV770 is only 50% faster than RV670 (320 SP) ?

No, the 4850 is "only" 50% faster but the 4850 is clocked almost 20% lower, has less memory bandwitdth and lower power consumption than the 3870.

 

[Anarchist4000]

the link in the quote gives a 404.

Try refreshing. I got a 404 the first time, refreshed then the page loaded.

 

[Wirmish]

No, the 4850 is "only" 50% faster but the 4850 is clocked almost 20% lower, has less memory bandwitdth and lower power consumption than the 3870.

... and the HD 4870 is 30% more powerful than the HD 4850... with 25% higher frequency.

So the GDDR5 = only +5% performance ?

 

[AlphaWolf]

... and the HD 4870 is 30% more powerful than the HD 4850... with 25% higher frequency.

So the GDDR5 = only +5% performance ?

your logic isn't very sound.

 

[Lukfi]

Faster memory doesn't yield great performance differences. But the HD 4870 has 20% faster core, not 25%.

 

[ChrisRay]

Doesnt GDDR5 have like very high latencies? Or was that GDDR4..