GTX 680, Turbo Cores, and Cuda Cores!
Subject: Graphics Cards | March 8, 2012 - 11:59 PM | Josh Walrath
Tagged: nvidia, kepler, gtx 680, GDC
It seems that there have been a few leaks on NVIDIA's first Kepler based product. Techpowerup and Extreme Tech are both reporting on leaks that apparently came from Cebit and some of NVIDIA's partners. We now have a much better idea what the GTX 680 is all about.
Epic's Mark Rein is showing off his own GTX 680 which successfully ran their Samaritan Demo. It is wrapped for his protection. (Image courtesy of Extreme Tech)
The chip that powers the GTX 680 is the GK104, and it is oddly enough the more "midrange/enthusiast" offering. It has a total of 1536 CUDA cores, runs at 703 MHz core and 1406 MHz hot clock, has a 256 bit memory bus pumping out 196 GB/sec, and has a new and interesting feature that is quite a bit like the Turbo core functionality we see from both AMD and Intel in their CPUs. Apparently when a scene gets very complex, the chip is able to overclock itself up to 900 MHz core/1800 MHz hot clock. It will stay there for either as long as the scene needs it, or the chip approaches its upper TDP limit.
These reports paint the GTX 680 as being about 10% faster than the HD 7970 in certain applications, but in others it is slower. I figure that when reviews are finally released the two cards will have traded blows with each other over who has the fastest graphics card. Let's call it a draw.
The GTX 680 should be unveiled in the next week or so, but initial reviews will not surface until later in the month. Retail availability will be relegated until then, but with the issues that TSMC has had with their 28 nm process (it has been stopped since the middle of February) we have no idea how much product NVIDIA and its partners has. Things could be scarce after the introduction for some time.