Radeon 4000-series in June?

Page 2 - Seeking answers? Join the AnandTech community: where nearly half-a-million members share solutions and discuss the latest tech.

AzN

Banned
Nov 26, 2001
4,112
2
0
Originally posted by: Extelleron
Originally posted by: Azn
Originally posted by: Furen
Originally posted by: Azn
It feels like 3870 was just released yesterday.

16 ROP, 32TMU with 480SP doesn't sound too powerful at all. Sure it's better than 3870 but it won't be too dramatic for sure compare to 8800gt. A 8800gts is 16/64tmu and 128SP that is more power than AMD's 320SP shader.

Good thing about this is that the mainstream parts are finally moving to 256bit memory bus.

Well, mathematically speaking, you're going to have 50% more hardware (with twice as many TMU, which is good) and a 50% (ish) higher clock, so around 225% the performance of a single 3870. not bad for a simple mainstream part. Now, you put two of these together and you have crossfire on a stick for the highend.

It sounds absolutely impressive but, IMO, it's impossible. I mean, it's the same process as the 3870 but they were able to cram 50% more logic and increase the clockspeed by a massive 50% while only increasing power draw by something like 30W... yeah right. If the R700 truly has 480 "shader processors" (aka 96 SIMD units) then I'd guess it will run at more pedestrian clockspeeds.


How do you figure 50% faster clock speeds? Memory is whole lot faster but core clock speed is only 25% faster. HD 3870 currently runs @ 775mhz and the new RV770 is supposed to run 1050mhz.

Again 8800gts has same pixel pipes and double TMU as RV770. It will be faster than 8800gts but I doubt it will be noticeably faster. Maybe 25% faster than 8800gts and about 50% faster than 3870.

Never say impossible. You know how small 3870 chips are compared to 8800gts? 55nm was at its infancy when 3870 was made. Now that it has matured it can pump out better yields.

In terms of GFLOPS:

HD 3870 = (320*2)*.775 = 496~ GFLOPS
HD 4870 = (480*2)*1.05 = 1008 GFLOPS
HD 4870 = 2.03X HD 3870 in shader performance

Texture performance should be much more than doubled.

HD 3870 = 16 * 775MHz = 12,400
HD 4870 = 32 * 1050MHz = 33,600
HD 4870 = 2.70X HD 3870

Memory bandwidth should be just about doubled as well. So HD 4870 = 2.03X Shader, 2.7X texture, and 2X memory bandwidth compared to the HD 3870.

You are forgetting pixel performance.

SP isn't the wholly grail of performance and neither is texture fillrate. They all add fps in a game but those kind of math doesn't add into the real world of gaming.
 

Extelleron

Diamond Member
Dec 26, 2005
3,127
0
71
Originally posted by: Azn
Originally posted by: Extelleron
Originally posted by: Azn
Originally posted by: Furen
Originally posted by: Azn
It feels like 3870 was just released yesterday.

16 ROP, 32TMU with 480SP doesn't sound too powerful at all. Sure it's better than 3870 but it won't be too dramatic for sure compare to 8800gt. A 8800gts is 16/64tmu and 128SP that is more power than AMD's 320SP shader.

Good thing about this is that the mainstream parts are finally moving to 256bit memory bus.

Well, mathematically speaking, you're going to have 50% more hardware (with twice as many TMU, which is good) and a 50% (ish) higher clock, so around 225% the performance of a single 3870. not bad for a simple mainstream part. Now, you put two of these together and you have crossfire on a stick for the highend.

It sounds absolutely impressive but, IMO, it's impossible. I mean, it's the same process as the 3870 but they were able to cram 50% more logic and increase the clockspeed by a massive 50% while only increasing power draw by something like 30W... yeah right. If the R700 truly has 480 "shader processors" (aka 96 SIMD units) then I'd guess it will run at more pedestrian clockspeeds.


How do you figure 50% faster clock speeds? Memory is whole lot faster but core clock speed is only 25% faster. HD 3870 currently runs @ 775mhz and the new RV770 is supposed to run 1050mhz.

Again 8800gts has same pixel pipes and double TMU as RV770. It will be faster than 8800gts but I doubt it will be noticeably faster. Maybe 25% faster than 8800gts and about 50% faster than 3870.

Never say impossible. You know how small 3870 chips are compared to 8800gts? 55nm was at its infancy when 3870 was made. Now that it has matured it can pump out better yields.

In terms of GFLOPS:

HD 3870 = (320*2)*.775 = 496~ GFLOPS
HD 4870 = (480*2)*1.05 = 1008 GFLOPS
HD 4870 = 2.03X HD 3870 in shader performance

Texture performance should be much more than doubled.

HD 3870 = 16 * 775MHz = 12,400
HD 4870 = 32 * 1050MHz = 33,600
HD 4870 = 2.70X HD 3870

Memory bandwidth should be just about doubled as well. So HD 4870 = 2.03X Shader, 2.7X texture, and 2X memory bandwidth compared to the HD 3870.

You are forgetting pixel performance.

SP isn't the wholly grail of performance and neither is texture fillrate. They all add fps in a game but those kind of math doesn't add into the real world of gaming.

ROPs are no longer the benchmark of a card's performance. If that were true, the 8800GT would be way slower than the 8800GTS 640MB and the 8800GTS 512MB wouldn't be as fast or faster than the 8800GTX.

Even then, the HD 4870 would still offer 35% more pixel performance and that is barring any architectural improvements that would improve performance per clock.



 

AzN

Banned
Nov 26, 2001
4,112
2
0
Originally posted by: Extelleron
Originally posted by: Azn
Originally posted by: Extelleron
Originally posted by: Azn
Originally posted by: Furen
Originally posted by: Azn
It feels like 3870 was just released yesterday.

16 ROP, 32TMU with 480SP doesn't sound too powerful at all. Sure it's better than 3870 but it won't be too dramatic for sure compare to 8800gt. A 8800gts is 16/64tmu and 128SP that is more power than AMD's 320SP shader.

Good thing about this is that the mainstream parts are finally moving to 256bit memory bus.

Well, mathematically speaking, you're going to have 50% more hardware (with twice as many TMU, which is good) and a 50% (ish) higher clock, so around 225% the performance of a single 3870. not bad for a simple mainstream part. Now, you put two of these together and you have crossfire on a stick for the highend.

It sounds absolutely impressive but, IMO, it's impossible. I mean, it's the same process as the 3870 but they were able to cram 50% more logic and increase the clockspeed by a massive 50% while only increasing power draw by something like 30W... yeah right. If the R700 truly has 480 "shader processors" (aka 96 SIMD units) then I'd guess it will run at more pedestrian clockspeeds.


How do you figure 50% faster clock speeds? Memory is whole lot faster but core clock speed is only 25% faster. HD 3870 currently runs @ 775mhz and the new RV770 is supposed to run 1050mhz.

Again 8800gts has same pixel pipes and double TMU as RV770. It will be faster than 8800gts but I doubt it will be noticeably faster. Maybe 25% faster than 8800gts and about 50% faster than 3870.

Never say impossible. You know how small 3870 chips are compared to 8800gts? 55nm was at its infancy when 3870 was made. Now that it has matured it can pump out better yields.

In terms of GFLOPS:

HD 3870 = (320*2)*.775 = 496~ GFLOPS
HD 4870 = (480*2)*1.05 = 1008 GFLOPS
HD 4870 = 2.03X HD 3870 in shader performance

Texture performance should be much more than doubled.

HD 3870 = 16 * 775MHz = 12,400
HD 4870 = 32 * 1050MHz = 33,600
HD 4870 = 2.70X HD 3870

Memory bandwidth should be just about doubled as well. So HD 4870 = 2.03X Shader, 2.7X texture, and 2X memory bandwidth compared to the HD 3870.

You are forgetting pixel performance.

SP isn't the wholly grail of performance and neither is texture fillrate. They all add fps in a game but those kind of math doesn't add into the real world of gaming.

ROPs are no longer the benchmark of a card's performance. If that were true, the 8800GT would be way slower than the 8800GTS 640MB and the 8800GTS 512MB wouldn't be as fast or faster than the 8800GTX.

Even then, the HD 4870 would still offer 35% more pixel performance and that is barring any architectural improvements that would improve performance per clock.

Of course render backs have effect on how the card performs. IF that was true why even have 16 rops? Why not just have 2 rops?

8800gts fail to G92 GT because 8800gt has nearly 3x texture fillrate and more SP that has more effect than 2400 more pixel fillrate. Same thing could be said about the GTX.