D9P/G94: 9600 GT on 19th of February
- Thread starter AnarchX
- Start date
leoneazzurro
Regular
According to the clock rate and texturing fill rate numbers it's 64 SP.
Uhm, SPs are independent from the TMUs, you can't claim that with any certainty whatsoever (although it *is* the current consensus, apparently).
vertex_shader
Banned
Looks like as the 8800GT, same length PCB (less layer?), rev2 reference 8800gt cooler :smile:
Wasn't it confirmed somewhere that the "rev2 reference cooler" was infact a custom solution by some manufacturers, not a official design by nVidia?
leoneazzurro
Regular
I obviosly forgot to mention "if the TU/SP ratio is the same as current G9x chip" (4 TAU per cluster). TU in G8x/G9x are not really indipendent from the SP, as there is a fixed amount of those per cluster. Of course, Nvidia could have changed the TAU/SP ratio in the clusters, but it is difficult to believe so: lower ratio would mean this chip has 128 or more SP (too much for its purpose), higher ratio would mean this chip would be highly shader limited for its texturing power, with only 32 SP.
That's assuming the that there's a POT number of ta/tus per cluster - not sure if this really needs to be the case, there could be 3 ta/tus per cluster thus giving 96SPs (though such odd ratios indeed just sound wrong).
A higher ratio certainly would be ridiculous, even now the chip can't achieve its maximum texture throughput (with bilinear filtering, for int8 textures at least) due to being interpolator (shader) limited.
edit: oops my math is off by a factor of 2. G92 has 8 tus per cluster of course, so that hypothetical case would be 6 tus per cluster for a 96SP 9600GT. And it doesn't really add up to 32tus, so better forget it...
Last edited by a moderator:
leoneazzurro
Regular
Hmm.. 3 TAU seems too odd for me, even if it would be possible. I think the 4 TAU per cluster in G92 are not there for filtering purposes only, but for point sampling as well (I think it's the same story of R6XX having four times more "texture Samplers" in the Texture unit blocks than Filtering Units)
64.
Jawed
Legend
OK, who believes this?:
http://www.fudzilla.com/index.php?option=com_content&task=view&id=5052&Itemid=1
http://www.fudzilla.com/index.php?option=com_content&task=view&id=5052&Itemid=1
Jawed
Fuad...
To fight in <$150 with 320mm² would be suicide. G94-400 will be 9600GT.
Well, if the RV635 turns out as weak as it seems in the other thread (other than a die shrink, HD3650 looks very similar to HD2600 XT, i.e., not much of a performance fighter), then the midrange performance crown will still be in Nvidia's hands, forcing AMD to once again fight with pricing alone.
This 9600 GT looks exactly the opposite of the 8600 GT/GTS: 256bit = over designed
Sound_Card
Regular
Well, if the RV635 turns out as weak as it seems in the other thread (other than a die shrink, HD3650 looks very similar to HD2600 XT, i.e., not much of a performance fighter), then the midrange performance crown will still be in Nvidia's hands, forcing AMD to once again fight with pricing alone.
This 9600 GT looks exactly the opposite of the 8600 GT/GTS: 256bit = over designed
HD3850?
HD3850?
I doubt HD3850 can go much lower than 150 dollars in the open market for a whole year (the standard time frame for any mainstream GPU).
666M transistors may be small for a performance part on 55nm, but it's way too complex for a mainstream product, otherwise there would be no RV635, now would it ?
A further low-clocked RV670 obviously wasn't cost-effective enough for AMD to cancel RV635...
The notion that was suggested by Arch was that Nvidia is using a 320mm² gpu to fight a 120mm² gpu. In such a case AMD have a die advantage of over 166%, thats HUGE. And then AMD does not have to "fight" on the pricing side, the ball would be in Nvidia's court.
I think RV635 was already out the door and bug free for AMD to cancel it. They may still use some of the borked RV670 chips to offer a sku that sits below the 3850 and still be price and performance competitive with 8800GS/9600GT.
Sound_Card
Regular
Maybe I live on a different planet, but I recall the HD 3850 holding a mainstream lable. Considering the 8600GTS is floating around $120 to $180, I doubt a 64sp 256bit SKU will be any less, if not cost more. From my perspective, the 9600GT looks like a band aid to hold off the HD 3850 which I doubt will do that well of a job other than being a 20 bill cheaper.
Well, if the 9600 GT is a "band aid" to last a whole year and fight off the HD3850, then what would you call to the 8800 GS 384MB/192 bit, or even the 8800 GT 256MB versions with slower memory ?
Sound_Card
Regular
Well, if the 9600 GT is a "band aid" to last a whole year and fight off the HD3850, then what would you call to the 8800 GS 384MB/192 bit, or even the 8800 GT 256MB versions with slower memory ?
I would call those better band aids. Just not economically efficient enough. Very much how ATi had to use the 1800gto to fight the 7600. Plain and simple, the HD 3850 is just more flexible all around. They are using 512mb SKU's to target the 8800GT 256mb, and they putting major pressure on top of the 8600 and soon to be 9600 by just plainly being much faster for a pinch more.
The HD 3850 seems to have punched NV right in the gut. I was only merly trying to point from what you said that the 9600gt is going to stand alone un disturbed. The HD 3850 is just too close to make that kind of call.
Last edited by a moderator: