Today NVIDIA launches their new GeForce GTS 450 Series of graphics cards. As the "GTS" naming suggests, the card is more of a mid-range part than a high-performance device. It is NVIDIA's expressed goal to offer a low priced DirectX 11 graphics card that the majority of gamers can afford - and still play the latest games.
NVIDIA's GF106 GPU is based on NVIDIA's Fermi architecture, which was released earlier this year. In order to achieve the product performance and positioning NVIDIA was looking for, they disabled some components of the full GF106 core, resulting in the graphics processor for the GeForce GTS 450. Unlike other Fermi-class cards the changes here are quite limited. Basically NVIDIA removed two memory chips from the design, two chips less means 2x 32-bit less bus interface, so a single 64-bit memory controller got disabled. The memory bus width is closely coupled to the number of active ROPs in the GPU, so one ROP partition is also disabled. Overall this change enables board partners to build more price competitive products because of the reduced cost for memory chips and PCB signal routing. It helps NVIDIA to increase its GPU yields. This also leaves NVIDIA with the option to create a more powerful "GTS 455" SKU at a later time, that has a 192-bit GDDR5 memory interface, with likely 768 MB or 1536 MB of memory.
It's interesting to note that NVIDIA did not disable any of the streaming multiprocessor units (SMs), and all four SMs holding 48 CUDA cores each are enabled, giving the CUDA core count of 192. NVIDIA's reference clocks on the GeForce GTS 450 are fairly low with 783 MHz core, 1566 MHz CUDA cores, and 900 MHz memory, so it comes as no surprise that most board partners are shipping pre-overclocked cards at lauch.
MSI's GeForce GTS 450 uses a custom PCB design and MSI's well-established Cyclone thermal solution. The card comes with an overclock out of the box at 850 MHz core and 1000 MHz memory which is a bit lower than the average of factory overclocks on the GTS 450 models tested today.