NEWS: X900 Series Specs

Quote:

---

Originally Posted by Micutzu

I don't think we're going to see R580 or G80 sooner than 8-10 months from now, the market doesn't require that and ATI/Nvidia won't release even if they already have them prepared, just for the competition's sake. The G70's already look like they can clock 500Mhz+ and have great yelds, so i would only expect a 24pipe 7800Ultra clocked at about 500/1400 or so as a response for the R520.

---

While all of that seems logical, only the last part seems probable as R580 really is only a refresh (although looking to be a great refresh) and I do expect it to be launched Q1. I too assume an ultra at 500/1400 to battle R520, as it seems most feasible. Who knows from there though on nvidia's side...I can only gather that they most definately have a gameplan to stay on top, even if it's by inches instead of miles, hence why no ultra is announced until we see what we see what R520 is and can do, and no definate release dates for nvidia's next parts until R580 rears it's head.

Quote:

---

Originally Posted by turtle

Say it's on the 110nm process, and clocks relatively similar to a g70 or g70 ultra. If it has the same amount of vertex shaders, presumbabley 8 extra pipes wouldn't do that much, especially as the % extra pipes added becomes less. (16-->24 is greater than 24-->32)

30%? I don't know about that.

---

The % increase from 24 to 32 is still 33%.

Quote:

---

Originally Posted by turtle

If you look at 7800gt's performance (again, yes, in 03 at stock rez which may or may not be a fair representative bench) against a gtx, it performs at 95% of a gtx at the same clock, that's with one less quad and one less vertex shader...and we all know vertex shaders are important. Also, didn't we try tests in another thread about disabling a quad on a nvidia card, with all vertex shaders intact, and it was only like 3% differance?

---

The difference between 20 and 24 pipes is only 20%. Also in this case we're talking two quads, not one. And yes, 3DMark03 probably wouldn't be a fair representation of the performance difference considering that even 05 is somewhat CPU limited on these cards.

I think you overestimate the importance of vertex units, simply because these days performance is almost always shader-pipe-limited, meaning the vertex units have more than enough power to push today's games. So much more that often times they stand idle compared to the work required of the shader pipes. Even the demands of the ROPs outweigh the vertex units because of current software trends.

The test I ran was with a 6600, and the vertex units were not the same. On one configuration I ran 4 pipes at 400MHz, with 1 vertex unit, and 8 pipes at 200MHz with 2 vertex units. The performance difference was merely 3%, in favor of the 4-pipe config. I performed further testing that I didn't post about, and found out that at higher clockspeeds the 8-pipe config turned out slightly faster (about 5%). This flip-flop performance characteristic suggested that there is no advantage between the configurations.

Quote:

---

Originally Posted by turtle

...but from what i've seen, the quads themselves arn't making a huge differance, it's the tweaking nvidia did to the core on the gf7 that makes it shine over gf6, and the higher clock speeds themselves on the 110nm process coming from 130nm.

---

The quads along with the new tweaks all had a helping hand. Keep in mind the clockspeeds were barely different between the 6800 Ultra and 7800GTX. 400MHz core versus 430MHz core (a difference of about 8%). The memory clockspeeds only went up 9% (although the 16 ROPs were extremely improved). In fact the only real reason NVIDIA went to 110nm was because it was very viable, practical, well matured by that point, and it was necessary to fit 24 pipes at a reasonable die size.

Clockspeeds weren't really a focus for NVIDIA. Their whole goal was to increase performance using the architecture as the driving force. It was sort of a tactic to get people to recognize the importance of a good architecture rather than just raw pipelines and clockspeed. Not only that, NVIDIA wanted the G70 to run cooler and use less power than the NV45. Not an easy feat.

Quote:

---

Originally Posted by turtle

Presubabley a 32pipe card would still be 110nm if it was g70-based. If it's 90nm, then crap, that's a completely different story. The clock speeds would surely make it 30% faster.

---

NVIDIA has already said they were taping out several 90nm parts by the launch of the 7800GTX (see this article). By now those should be fairly close to being ready to launch.

A 32-pipe core on 110nm just wouldn't be possible. You can call this speculation if you like, but it's just going by logic. Already the G70 with 24 pipes has a massive die size, even bigger than the NV40. 32 pipes would make it impossibly big, so moving to 90nm with that part would be a must.

Again, I must emphasize that clockspeeds aren't everything, especially when talking about the sort of architectures NVIDIA employs. Even if your line does see a process shrink, that doesn't automatically mean you're going to shoot up the clockspeeds, as the G70 has proved. A process shrink can be useful for other things, as I've mentioned earlier.