GeForce GTS 450: Farewell, G92
There’s something about a game of catch-up that inspires urgency. After struggling to bring its Fermi architecture to market six months after AMD launched its own DirectX 11-class Radeon HD 5000-series, Nvidia’s derivative GPUs are now hitting one after the other, in rapid succession.
I’ll admit—I was worried when the company launched its GeForce GTX 465, based on the massive GF100 with nearly half of the chip disabled. Was it having problems with derivative designs?
Then Nvidia introduced the GeForce GTX 460, based on the GF104 graphics processor. That card impressed me. It was, as our respective review mentioned, what GeForce GTX 465 should have been. Then we tested two GeForce GTX 460s in SLI and gave that configuration another Recommended Buy award—not necessarily for what those two boards did against the competition from AMD, but for the fact that they trounced a GeForce GTX 480 at a comparable price. Suddenly, we were seeing solid value from Nvidia.
Now, two months later, the company has another derivative part ready to rock. Its GF106 is, in essence, one half of a complete GF104. But remember that the GeForce GTX 460 is really seven-eighths of a GF104. So, the GeForce GTS 450 is in all actuality a bit more than one-half of a GTX 460.
Enter The Next Cut-Down Fermi-Based GPU
Remember back to the GeForce GTX 480 launch. The company’s flagship was (and still is) based on a cut-back version of the GF100 GPU. This was a consequence of not getting the yields needed to launch with a 512-shader part. The GeForce GTX 460 was (and still is) based on a cut-back version of the GF104 GPU. This was to avoid showing up the GeForce GTX 470 with too much performance.
Today, the GeForce GTS 450 represents Nvidia’s first Fermi-based card based on an unaltered graphics processor, GF106. Update: it turns out that the block diagram Nvidia sent out isn't true to the silicon. The GF106 used on GeForce GTS 450 actually is deliberately handicapped by Nvidia. The only product, thus far, based on an unaltered GF106 is the GeForce GTX 460M mobile graphics module. Here, one ROP partition is turned off and simply "ignored" by Nvidia's team of block diagram marketers.
GF106 is composed of one Graphics Processing Cluster (GPC), with four Streaming Multiprocessors (SMs). Each SM wields 48 shader cores. These are the same “modified” SMs first seen on GF104, equipped with four dispatch units and eight texture units (instead of two and four, respectively). In total, that gives us 192 shader cores, four Polymorph engines (one per SM), and 32 texture units.
The chip’s back-end is similarly cut down to a pair of ROP partitions, each capable of eight 32-bit integer pixels per clock. Nvidia refers to that arrangement as 16 ROPs. Naturally, each partition is associated with a 64-bit memory path, yielding 128-bits aggregate this time around. That’s half of what the GeForce GTX 460 1 GB has available to it, so we expect to see a substantial performance hit with anti-aliasing enabled.