Originally posted by: Extelleron
Originally posted by: chizow
After seeing more firm specs on the RV770 I don't think there's too much good news here. The only good news for ATI is that they'll have the fastest single-gpu card for a month or two until GT200 releases, at which point they'll get lapped again in terms of performance. Than an X2 version might put them in a competitive position again at which point NV will respond with a die-shrink or SLI-on-a-card solution of their own or both. All while maintaining a comfortable lead at the high-end with a $2000 GT200 Tri-SLi solution.
As for 4870, I don't think it'll be much faster than 9800GTX/8800GTX/Ultra in terms of performance. Maybe 15-25% faster, max. 16 > 32 TMUs seem to be the biggest gain here and specifically mentioned as a major bottleneck for ATi R600 parts. Still, that only puts ATI's texture fill-rate equivalent to a 9600GT, not counting any advantages from different vendor design. The rest of the specs seem rather unspectacular with questionable gains, although shaders may also scale well as that seemed to be another weak point of R600. Going from 64 to 96 real shaders, or 320 to 480 super scalar along with unlinked shader clocks should help close any gaps in shader performance in unoptimized games where NV held a lead previously.
This part would've been a great answer to G80/G92 6 months ago when RV670 released, or even a year ago when R600 released. But at this point I think it'll be obvious that its too little too late, mostly competing with G80/G92 and made obsolete again when NV fires back with GT200 later this quarter.
If you look at the specifications vs. the performance of the 3870, then you are dead wrong.
Looking at pure numbers, the HD 4870 is a solid ~2X improvement in just about every area of the GPU over HD 3870. The only area where performance hasn't been improved much is the ROP area; the ROPs are not much of a bottleneck, and with a faster core speed, ATI already has a significant advantage in that area over nVidia.
Looking at the numbers to back up what I said:
In terms of shader performance:
HD 4870 (480 * 2 * 1.050) = 1008
HD 3870 (320 * 2 * 0.775) = 496
4870 =
2.03X 3870
In terms of texture performance:
HD 4870 (32 * 0.850) = 27.2
HD 3870 (16 * 0.775) = 12.4
4870 =
2.19X 3870
In terms of memory bandwidth:
HD 4870 (3880 * 0.032) = 124.2 GB/s
HD 3870 (2250 * 0.032) = 72.0 GB/s
4870 =
1.725X 3870
The 4870 improves in every aspect significantly, and it is more balanced than the current design. Shader performance remains strong, but now the texture performance is there to back it up. The GPU has plenty of power and also plenty of memory bandwidth to keep it fed.
Looking at 3870 reviews... there is no situation that I can find where doubling the 3870's performance does not equal better performance than the 8800 Ultra. In many cases the gain is very significant.
The 4870 X2 should definitely exceed the performance of the 9800GX2 by a wide margin. Obviously GT200 is another story. But how powerful is GT200 really going to be? Considering the current die size of G92, which is huge as it is, how much room does nVidia have to expand on it? I cannot imagine that GT200 would be anything more than 40-50% faster than 9800GX2 if it is a single GPU, and from what I see that would make it slightly faster than the 4870 X2, if that.
The problem with nVidia is of course die size, as I mentioned. From what I have seen, RV770 should be much smaller than nVidia's G92 and exceed its performance greatly. Even nVidia moving to a 55nm process would likely make G92 around equal to RV770 in die size. Considering rumors point to GT200 on 65nm, it would likely be a chip like G80, in the range of ~500mm^2. That's not a GPU that any company wants to produce; it is a lot easier and cheaper to fab (2) 250mm^2 chips than to fab a single 500mm^2 chip. If GT200 is really single GPU, then that will likely be the situation.
As for nVidia responding with a die shrink, they are usually way behind in moving to a new process. AMD moved to 55nm in Nov 07, and nVidia does not have a single 55nm GPU out 5 months later. By the end of this year, RV770 could be shrunk to 45nm if TSMC's process is ready in time.