Nvidia launches GeForce GTX 660 (not Ti)

Quizzical

http://www.geforce.com/hardware/desktop-gpus/geforce-gtx-660-oem/specifications

It's OEM-only, which means Nvidia probably doesn't have a ton of them to get rid of.  Still, it's an interesting card.

The card disables two of GK104's eight SMXes, as compared to one for the GTX 660 Ti and GTX 670.  It also disables one memory channel, like the GTX 660 Ti.  It clocks both the memory and the core lower, which allows the card to only need one 6-pin PCI-E power connector rather than two.

Memory is still clocked very highly, however, at 1.45 GHz.  For comparison, AMD's Radeon HD 7970 clocks memory at 1.375 GHz, which is the same as last generation's Radeon HD 6970.  The rest of AMD's lineup other than the 7970 GHz edition clocks memory substantially lower.

It also matches the memory channels properly, with either 1.5 GB or 3 GB of video memory.  I'll take this as a concession of the obvious:  2 GB of video memory on a 192-bit bus is completely stupid.

So you can think of the GeForce GTX 660 as being a GTX 680 with 1/4 of the card disabled, and the whole thing clocked lower.  Unlike the GTX 660 Ti, that should make for a balanced card, and could be a nice product--unlike most OEM-only video cards.  Too bad it's OEM-only, as at the right price, it could be a nifty product.  Of course, the right price would surely have to be less than a Radeon HD 7870, and maybe Nvidia didn't want to go there.

-----

But this again raises the question of, where is Kepler?  Nvidia now has 5 GK104-based cards out, one GK107-based card that is a crippled joke and needs its price to fall by 1/3 to be competitive, and nothing else.  No Teslas or Quadros, either; AMD's GCN-based competitor to both, the FirePro series (FireStream has now been discontinued) has already been officially launched and reviewed, though the cards don't seem to be available for purchase just yet.

It's looking like by the time Nvidia has a full Kepler lineup out, it may have to compete against the Radeon HD 8000 series for most of its lifetime, rather than the 7000 series.  This would be much like how Fermi had to compete mostly against the Radeon HD 6000 series rather than the 5000 series that should have been its contemporary.

Still, Kepler should turn out better than Fermi did.  The cards that are out show that Kepler is a nice architecture, at least at the high end.  GK107 and Cape Verde being the same size gives me some doubts about how nice Kepler would be at the low end, but it should be a lot more competitive than Fermi was.  Unlike Fermi, where the architecture itself was a disaster, the only major problems with Kepler are volume and yields.  That should be fixed eventually, but in the meantime, AMD wins this generation pretty much by default.

Even so, volume and yield problems should be fixed eventually, however many respins it takes.  Even if it requires a base layer respin, that's fixable.  There is real hope for Nvidia that a full Kepler lineup might well finally win a generation for the first time since the GeForce 9000 series beat the Radeon HD 3000 series, or at least get rough parity with AMD.

Unless, of course, the Radeon HD 8000 series is awesome.  I'd expect some substantial improvements over Tahiti at the high end, but not so much below that, as Cape Verde and Pitcairn seem to be pretty good.  But you never know; the Radeon HD 3000 series was, after all, followed by the 4000 series.