Nvidia and the Kepler Architecture

Thursday, September 13, 2012
At PAX, I had the most excellent privilege to view the Kepler architecture (600 series) in action.  It is BEAUTIFUL!  Unfortunately, I was unable to obtain a copy I could unpack and take my own comparison videos of, but the next best thing is words, simple words, no engineering gobbledy-gook this time.  You're right, I'm lying (at least about words being the next best thing), so here you have it, this was my favorite demo to be viewed as it includes all aspects (PhysX, tesselation, and adaptive Vsync) in comparison to not having these things:

See, this is beautiful.  They didn't reveal what the machine without PhysX was running with, but I'd have to assume another 600 card.  I'd imagine it looks just a little bit better than had they run it on a Fermi (500 series) card.

To break down the aspects included, here is what they are.  PhysX is Nvidia's improved physics engine, tesselation which is the process by which a polygonal patch is broken down into higher resolution sub patches, dividing a low polygon count mesh into a higher detail mesh on demand (usually based on distance and angle to the camera)  which means that it's feasible now to, say, round out a character's ears to make them look more realistic.  And adaptive vsync allows for less stuttering on screen during a game.  As I said, I am not going to break this down further, otherwise this article would be difficult to manage!

Overall, the major improvements from the Fermi architecture are the aspects listed above, which are made possible by twice the amount of CUDA cores being present in any derivative of the Kepler architecture, this allows for more threads to be processed at a time, and therefore faster calculations!

Today Nvidia officially announces the 660 and 650 cards.  This is a fantastic opportunity for those more sensitive to price changes in video cards to update, as compared to the 670+ cards the price points are rather low.  You do suffer some functionality from these cards as compared to the higher priced versions, the 650 is not SLI compatible, and the 660 has only slightly fewer CUDA cores than the 660ti.  These seem like they will be good for about a year or more depending on the usage on them, but as a personal opinion I think that for price and functionality the 680 is superior and will at least time proof your machine for a considerably longer time than the 660.

Because I am being a lazy person today, following this text will be a list of the cards' attributes.

GeForce GTX 660ti
GeForce GTX 670
GeForce GTX 680 <-- My personal favorite!
GeForce GTX 690

If you look at these, you should notice that the amount of CUDA cores has deeply impacted the price point on the various models (660 and 650 not listed, but they do have 900+ cores, except 650).  If you look back at Fermi's specs, you'll also see that there will be (as a direct result of ~50% more cores plus the other awesome aspects) approximately a 50% performance increase should you decide to upgrade to Kepler, even if you decide to go for the 650 card.

Happy gaming!