AMD: R9xx Speculation

I'm hearing rumors that 100% of certain iconic Parisian skyline entities are for sale & I can cut you a great deal

 

Its not from new owners, but from a guy who sells 480s and has enough time to flash and check them one by one

 

The point I was trying to make is that even *if* all 480's could be flashed to 512 cores their life may be severely shortened. While it is only a speculation, it's the best reason I can think of for nVidia not planing to release the full spec 480 any time soon

 

Better yet: how can you flash a GTX 480's BIOS to turn it into a 512-SP part when there are no 512-SP parts, therefore no such BIOS? Unless you have access to more than the public.

 

I'm not sure, but aren't GeForce cards "non-unlockable" since GF6800? GTX470, GTX465, GTX275, GTX265, G92, G80 etc... none of them can be unlocked via simple bios flash. GTX480 should be an exception?

 

I lolled at the mere mention of 512CC unlock, especially since no-one can even do a proper voltage adjustment on the GTX480 bios, let alone do something like unlocking cores, written in a custom bios.

but then I Googled.. and I noticed that nVidia improved their scores 100%, back in March, you only had a 50% chance to unlock a GTX480 to 512CC.

Is this one of their tricks to make people sway or really wishful thinking on the side of the unknowing?

The only reason GTX465s unlock to GTX470s is because they are just that, 470s sold sold as 465s because for their current price, they just stay nVentory.

 

How does that correlate with 465 having only 1 GiB? And the price difference isn't that big between the two, making 470 far more attractive than 465 at only a 15% price premium you get ~25-30% more perf.

 

So, any rumors about the r9xx?

 

That it's still 4 months out.

 

We want to know nAo!!! Actually how do you know its still 4 months out? It could also be 3 months or 5 months unless you've heard differently, so share!

 

Late October to early November if I read it correctly.. September was too early any way.. so yeah.. it can always be later of course

 

Sorry for off-topic..

What we will get is 336 shaders @ 1350 MHz.

That's 83% of the shader power of the 470.

On memory bandwidth the difference is 115.2 GB/s vs 133.9 GB/s, 86%.

Texture fill rate of 470 is higher though, 37.8 Gtx/s vs 34 Gtx/s.

As the texture fill rate is probably be the biggest bottleneck of 470 the improvement in that may actually put 460 to same performance level, even though shader power and memory bandwidth will decrease.

So I admit that performance-wise, GF104 can indeed replace 470, and I was incorrect before. (though the reason why I was wrong was different than what you said)

 

sorry for more off-topic..

Power/thermals cannot be reason for disabling some of the cores;

More cores at lower clock speed(giving same performance) will consume less power/create less heat than smaller number of cores at higher clock speed.

(power is linear to core count, but dynamic power is f^3 when voltage is adjusted accordingly)

 

GTX480 has 48 ROPs, which run at very low frequency, so their theoretical performance numbers are only slightly above RV870. Decreasing their clock-speed to compensate higher number of SPs could decrease performance quite significantly. GF100 doesn't like high-resolutions and this change could make it even worse...

 

Such is the life of a hardware speculation expert such as not myself. October/November/December hell they could even miss this year entirely! Oh the joys of not knowing... The only certainty is that a lot of the information you have is certainly wrong!

If it helps any, im going to make sure Dave gets his name credited to a certain game development as revenge for all this secrecy.

 

+23,5% theoretical-marks, to be precise.

Isn't that just another way of saying:
- GF100 is very efficient even at mainstream resolutions or
- Cypress is not very efficient at mainstream resolutions?

I mean performance is alot more like you would expect given the theoreticals in higher resolutions, after all.

 

Yes, your formulations are correct, but my point is, that performance decrease caused by reduction of back-end performance could be more significant, than performance gain caused by enabled units (32SPs/8TMUs). So functional units can be really disabled because of power issues.

Btw. it would be interesting to test relation between gaming performance and number of active units... RV870 isn't very effective in this aspect (1440->1600 SPs (+11%) gives only ~2% of gaming performance, so clock speed is more important for RV8xx than mumber of functional units). I believe GF100 is much better in this aspect, but I still believe, that higher clock speed result in better performance gain than more units...

 

GTX 480 is only capable of 30x 32-bit pixels per clock so it actually has less fillrate than both the HD 5870 and HD 5850. Not counting Z and memory bandwidth limitations of course.

 

Something over 20 Gpixels/s fillrate should be more than enough for single monitor and 4M pixels. For full HD lcd monitor and vsynced 60 fps this fillrate is overkill. In real games u are probably limited by bandwith and shaders.
It would be interesting what performance could u reach with radeon 4870 having double shaders and tmu-s. I think we would be suprised how close would it be to 5800 cards.

 

Nextgen mobile chips:
GT425 probably will launch in October. ATI will go 6300,6400,6500. 5 series will gradually be withdrawn from the market after October, replaced by the 6 Series.

Source:
http://translate.google.com/transla...zol.com.cn/38/1191_376981.html&sl=zh-CN&tl=en

Wu Haijun is Chairman of HASEE.