8800 GT (aka G92)

I noticed today that NVidia has launched the Geforce 8800 GT. This is the now legendary G92 chip, which is rumored to support 64bit double precision arithmetic in the GPU, in line with earlier statements from NVidia suggesting 64bit functionality by the end of 2007. However, when one looks online at the specs for this GPU, there is no mention of any 64bit capability.
Can someone shed some light on when 64bit functionality is expected to be released ?

G92 does NOT have double precision support. I cannot comment on future, unreleased products or their release schedules.

Looking at the table of specs on this page can anyone say how this will compare, performance-wise, to the G80s?

For example, I can do about 33% more stuff with a GTX vs. a GTS. The speed up looks to be proportional to the number of stream processors.

Does that mean I might expect a ~17% (1/6) speed up with a GT vs. a GTS?

Probably sm architecture in GT is the same as in GTS, so your theoretical speedup is:

(1500-1200)/1200 * (112-96)/96 = ~45% (you don’t include higher sm clock in your previous post)

PS Using GTX you could reach 50% theoretical speedup again due to higher sm clock.

The shader clock is 1500 MHz for the Ultra, not the GTX. The 33% I gave is what I actually see. I guess the best way to find out is to plug in one of these new cards and see what happens.


I have some questions about G92…


Is supported in current release 1.0 or it’s going to be supported in next version?..


The compute capability in CUDA… Is 1.0 like G80 cards? or 1.1 like 8600 cards? or better?..

Someone can say/check if device to host transfers are improved 2X on these cards due to PCIEx 2.0 on motherboards with these (i.e the bandwith test in CUDA SDK with pinned/locked memory) (now that G92 and X38 Motherboards are avaiable?)


Compute capability 1.1.

Why did nVidia break naming convention on this GPU? The G7* became the GeForce 7000 series, the G8* became the GeForce 8000 series. And then, the G92? The 9000 series? Err…no! Another 8000 series.

This was very disappointing, both as a researcher and a gamer.

I think they only do a +1000 skip if they have another high end model. As 8800GTX is still the fastest, making a 9xxx model would completely confuse buyers.

You’ve already post link on specs above. Look through info provided there (GTX has 1350 sm instead of 1200 gts ). But I agree with you that just testing on your project could provide reliable result :)

I don’t think the G92 name was publicized by NVIDIA, the name is 8800GT. At any rate, it’s just a name. It’s the performance that matters, and that’s pretty good.