In early 2008, NVIDIA’s GeForce 9600 GT, armed with a mere 64 shader units, 16 ROPs, 512 MB of memory, and an inviting price-tag, rattled competitor AMD’s Radeon HD 3800 lineup. It allowed gamers to achieve playable framerates with cranked up visual details that were, until then, not possible with graphics cards in its price-segments.
From that point on, NVIDIA realized it could gain a substantial market share in the sub-$250 price-segment, hovering around the $200 price-point, if it creates a GPU that can handle high-resolution gaming with a fair amount of eye-candy enabled. Continuing its legacy, NVIDIA’s GeForce GTS 250, GeForce GTX 460, and Gigabyte GTX 660 2 GB 192-Bit GDDR5 Directx 12 Double Fan Graphics Card are each successful products. In August, NVIDIA launched the GeForce GTX 660 Ti, a GPU that achieved a nice price-performance index in the $250-300 price-range. NVIDIA’s next logical step would be to create a GPU that does the same with the $200-250 price-range. Enter the GeForce GTX 660.
Unlike its “Ti” cousin, the GeForce GTX 660 is not based on the GK104 silicon, from which several other GPUs, such as the GTX 670, GTX 680, and the dual-GPU GTX 690 are derived. The GTX 660 is, instead, based on the new GK106 silicon, which makes its desktop debut today. The GK106 is a physical downscale of GK104, which retains all its features, including component hierarchy, but has fewer numbers of them. The GK106 silicon is smaller, with a die-area of 221 mm² and transistor count of 2.54 billion (compared to 294 mm² and 3.54 billion of the GK104). The GK106 is built on the same 28 nanometer silicon fabrication process. A smaller chip results in reduced power draw. A case in point is that the GeForce GTX 660 needs power from just one 6-pin PCIe power connector; the GTX 660 Ti needs two of them.
As mentioned before, components on the GK106 maintain the same hierarchy as on GK104, and the two provide the same exact feature-set. The chip is based on NVIDIA’s successful GeForce Kepler architecture. While GK104 packs eight graphics processing clusters (GPCs), with a total of sixteen streaming multiprocessor (SMX) units, holding 192 CUDA cores each, amounting to a total of 1,536 CUDA cores; the GK106 packs three GPCs, and five SMX units, totaling 960 CUDA cores. It’s interesting to note that NVIDIA created a GPC with just one SMX unit, if the block diagram is anything to go by. Perhaps the chip really does have six SMX units, but it’s kept out of the block diagram, perhaps to help harvest the chip better.