Generally, we'd open a review like this one by reminding you of the recent history of video cards and graphics chips like this one, setting the proper context for everything that comes next. Today, though, I have a plane to catch, and I anticipate writing a lot of the commentary on the following pages from an oddly crouched position in a coach-class seat while stale air tainted with the faint smell of farts blows in my face.
The source of our rush is a long, intense week spent with Nvidia's new graphics card, the GeForce GTS 450. Priced at around $130, this card is Nvidia's answer to the Radeon HD 5700 seriesif you can call it an "answer" after the competition has been on the market for a full year. Regardless of the timing, though, the GTS 450and the GPU behind itis a potentially attractive proposition for those who lack the resolve and (display) resolution to spend more than the cost of three big-name games on the hardware required to play them. Keep reading for our detailed testing and incredibly rushed text on the GTS 450.
Yep, it's yet another Fermi derivative
The graphics chip behind the GeForce GTS 450 is the third variety of the DirectX 11-class Fermi architecture that Nvidia has brought to the desktop market. As you may know, many chip designs these days are essentially modular, and can be scaled up and down in size and features to meet different goals.
The chip that powers the GTS 450, known as the GF106, is perhaps best thought of as roughly half of the GF104 GPU used in the GeForce GTX 460. Where the GF104 has two GPCs, or graphics processing clusters, the GF106 has just one, so it has half the triangle setup rate of the GF104and just one fourth that of the big daddy, the GF100.
Inside of that GPC are four shader multiprocessor blocks, or SMs, arranged essentially as they are in the GF104. That means each SM has 48 stream processors (Nvidia likes to call them "CUDA cores"; we do not) and a texture block capable of sampling and filtering eight texels per clock. In total, then, the GF106 has 192 SPs and can filter 32 texels per clock. For compatibility reasons, this GPU has the ability to process double-precision floating-point math, but only at one-twelfth the rate it can handle single-precision math, again like the GF104. (The GF100 is much more formidable, but it serves different markets.)
If you look closely at the diagram above, you'll notice that the GF106 bucks expectations for a mid-range graphics chip in a couple of notable ways. Rather than the expected pair of 64-bit GDDR5 memory interfaces, the GF106 has a trio. Correspondingly, it has three ROP partitions, each capable of outputting eight pixels per clock, rather than the two ROP partitions one might expect. The 50% wider memory interface and ROP partitions give the GF106 substantially more potential oomph than competitors like AMD's mid-range Juniper GPU.
Of course, the extra power comes at a price, as you can see with a quick glance at the transistor count and die size numbers on the right. The GF106 is quite a bit larger than Juniper, all told.
Incidentally, since Nvidia doesn't divulge die sizes, we've put asterisks next to some of the figures in the table. We've simply gone with the best published numbers we can find for GF100 and GF104, but since it lacks a metal cap, we were able to measure the GF106 at roughly 15 mm by 16 mm, or 240 mm². We may be off by less than a millimeter in each dimension with our quick sizing via wooden ruler, but we're pretty close.
The larger chip size likely translates into higher manufacturing costs for Nvidia, but it doesn't necessarily translate into higher prices for folks buying graphics cards based on it. We're just showing you this information for the sake of chip-geekery. Following further in that vein, we have some similar-sized pictures of the two chips below, shown next to a U.S. quarter to celebrate American hegemony and also to provide a size reference.
The intriguing thing about the GF106 is that, like all of the Fermi-derived graphics processors to date, we've not yet seen a product based on a fully enabled version of the chip. The GTS 450, as we're about to find out, only uses a portion of the GPU's total power. We're dying to know whether Nvidia has been producing gimpy implementations of its DX11 graphics chips out of necessity (due to manufacturing and yield issues), for strategic reasons (keeping a little juice in reserve), or some combination of the two (and what combination, really, which is the key question). We don't know yet, but we do get to use a lot of parentheses in the interim, which is its own reward.