Introduction

NVIDIA's much talked about Kepler GPU architecture is finally testing waters in the crucial performance market-segment, with the GeForce GTX 670. The new graphics card, as NVIDIA puts it, is built from the DNA of GeForce GTX 680. This launch closely follows that of the GeForce GTX 690, which is the posterboy of the Kepler GPU architecture, and is the fastest graphics card ever made. The GeForce GTX 670 is designed to let NVIDIA capture the US $399 price point, its competitor is meant to be Radeon HD 7950, although its specifications give us hope that the SKU could fry both bigger and smaller fish than that. It could target the Radeon HD 7950, as well as take a shot at scoring a higher price-performance than the much cheaper Radeon HD 7870 GHz Edition.

The GeForce GTX 670 utilizes the same 28 nm GK104 silicon, with one streaming multiprocessor (SMX, amounts for 192 CUDA cores) cut out, resulting in a CUDA core count of 1,344. The GPU clock speeds are also toned down a tiny bit, with 915 MHz core, and 980 MHz GPU Boost frequency, but the memory clock speed (of 6.00 GHz, resulting in 192 GB/s bandwidth) is left untouched. It packs 112 texture memory units (TMUs), and 32 raster operations processors (ROPs). The GeForce GTX 670 retains the entire feature-set of the GTX 680, including 4-way SLI support, and the ability to drive four monitors by a single card (and hence 3D Vision Surround).



A lot rides on the success of GeForce GTX 670, apart from its ability to compete with AMD's offerings. Besides the fact that even at $399, it could make NVIDIA substantial margins, it could allow NVIDIA to carve out even smaller SKUs using the same chip, to capture lower price points, and save both R&D costs and harvest GK104 chips better.