512MB GeForce 7800 GTX

ATI's design approach is pretty clear to me: implement enough parallel units to soak up the available bandwidth. This applies to texturing and rendering (ROPs). So there's no point in implementing 24 or 32 texture pipes if there's only enough bandwidth for 16.

In those terms I like to characterise R5xx as being designed for about 1GB/s per texture pipe and 2GB/s per ROP. It would be interesting to examine this point to see if it's fair...

But ATI is heading down the road of maximum arithmetic per clock (hence 16-1-3-1 - and the inference that the ratio will only go up, e.g. 16-1-4-1 - though a unified architecture may turn up first). Naturally the out-of-order scheduler in R5xx is a key component

NVidia sort of got there first with 6600GT, 4 ROPs instead of 8 (and ATI doesn't seem so interested in splitting ROPs off - at least not yet - though Xenos does have independent ROPs, but that's where the daughter die comes in). And there's been sight of an NVidia patent in which texturing is run decoupled from arithmetic. So NVidia could be close to having fully decoupled ROPs and texture units. In which case G80, say, may be playing the maximised arithmetic game, too.

Jawed

 

If you exclude the texture pipes (which is what Mintmaster did) then they're prolly about the same size. Hence unification - and why unification is so good for SM3 vertex texturing. The pixel pipes obviously have rasterisation and AA sample generation hardware plus batching (and lots of other fiddly details, I'm sure...), but that's fixed function so shouldn't be that large.

Remember that $200/400 price-gap between X1600XT and X1800XL? I expect that the release of R580 (which can't be long after Christmas) will shove X1800XL and X1800XT-256/512 down a couple of hundred dollars (knowing ATI, it'll only be $100, sigh) - and we'll see three or more R580 variants taking over the $400+ high-ground.

The only brake on R520's price-crash would be a simultaneous release of an 8-1-3-1/2 part (RV560?) at the same time as R580.

Jawed

 

This is the crux of the matter, how well can nvidia's 90nm be implemented at the high end.

Will nvidia use the smaller process to go for 32 pipes or simply just move the G70 24 pipes to a faster level? I agree with democoder absolutely that parallelism is what 3d is all about however I doubt clock speed limits will be a limit, not when you are only at 25% of what cpu's can do. I'm talking the physics of transistors here rather than the realities of trying get a gazilion transistors on a gpu chip fast.

Can nvidia shift to 90nm successfully ? Ati have been on the backfoot for a while now but it seems that r580 is "best foot forward", can nvidia keep in step? I have no idea.

 

You´re kidding, right ? Comparing a sub-100M chipset with a potentially very high clocked >300M part on 90nm.
We haven´t even seen prototypes of G72/G73 and G75, yet.

Q1/2006 at the earliest, while R580 will beat 512MB GTX. (Ultra is reserved for 90nm).

 

Are you sure about that ?

 

Well the transition itself went very smoothly for ATI - it was the chip that had the problems.

 

If today's architectures could combine as many floating point as fixed point instructions (in the recent past) there wouldn't need that much of a need to continue scaling units as much.

IMO amount of units is going to scale equally as clock frequency in the coming years; of course not to the same degree probably as current Intel CPUs, but I wouldn't be surprised to see damn close to 2GHz frequencies in the next decade.

Besides DX10 sooner or later will mean for all IHVs unified shader cores; counting not only gets more complex for the layman as me then , but units scale in a totally different manner then.

 

R580 is what the R520 should have been in terms of ALU throughput, necessity or not.

In terms of ALU performance it's rather NVIDIA that's one step ahead with the G70; the question mark would be what route exactly they're planning to follow with the 90nm refresh in early 2006. However if I presuppose that both R580 and G7x@90nm won't have any changes when it comes to ALUs compared to their predecessors I don't necessarily see a disadvantage for NV if they further scale the same units as so far.

If however both sides have made significant changes in the pipeline we're not aware of, all is up in the air. Point being though that IHVs usually do not result to radical changes with "refreshes".

 

No benches yet but Sampsa said this:

 

Well of course it does. The more interesting question is by how much. Hopefully we find out Monday.

 

Just to let you know .. that's close to the price that we pay($1100+) here for one GTX 256MB

US

 

Actually, given its its scheduling, the order in which it was done makes very good sense, IMO. R520 concentrated on SM3.0 and IQ - back in June we were all scratching our butts waiting for some decent titles to really make use of the ALU horsepower available around then, arguably we still are now, so concentrating primarily on IQ for a Spring '05 part would have been a reasonably good call. Going back and beefing up the ALU design for a winter part (R580 should have been released here, or here abouts) also makes sense as winter is usually a time for all the big new titles. The main thing "wrong" with R520 is its timing, but we knew that.

 

I wanna know if he benched @ hi-res AA/AF, else his statements are pretty useless imo.

 

You were scratching your butt?

 

So.. a December, January, February release for the R580

Question.. will it be a paper release like the R520?

 

The 520 wasnt a paper launch. They had XL's avail two days after the date, even in retail stores such as Frys. They had XT's avail three days before the date, and in retail stores such as CompUSA on the release date.

 

Kinda, yeah. (Heh, reading back on that entire conclusion now is quite interesting)

Actually, I wasn't drawing references to R580's current release date.

 

These cards are on sale right now. Link? Heh.. Search!