The upcoming GeForce GTX 260/280 GPUs are based on the GT200 (NV60) core and will be built using a 65 nm manufacturing process at TSMC. However, we heard that not many chips will actually fit on a 300 mm wafer, since Nvidia has come up with a huge die measuring 24 x 24 mm, resulting in a die area size of 576 mm2.
This area is almost 100 mm2 larger than Nvidia’s previous 90 nm high-end GPU (G80) and a consequence of 16 processing blocks (G80 came with nine blocks, eight were enabled for GTX models, six for GTS versions) and a new 512-bit memory controller, which replaces the old 384-bit model (the GTS260 will integrate a 448-bit version).
The current G92 or GeForce 8800GT/8800GTS512/9800GTX/9800GX2 CPUs are built in a 65 nm process and end up at a die size of 330 mm2. The original G80 (NV50) GPU, better known as the GeForce 8800, was manufactured in 90 nm and delivered a die size of 484 mm2.
With its new GPU generation, Nvidia is going to continue on the safe route and plan with enough spare transistors for 240 shader units (actually, 240FP+240MADD). Just like ATI’s graphics parts and Sony Cell processor, this should not be considered odd. Keep in mind that the G92 chip debuted with 112 shader units and, after production had ramped up, Nvidia unlocked all the shader and texture units to create GeForce 8800GTS512, followed by the 9800GTX and GX2.
The same will be the case with the GeForce GTX 280 and 260. The GPU will have 15 processing units (240 shader processors) available on the GTX280, while the GTX260 will come with 12 units for a grand total of 192 shader processors. This may be an indication for the complexity involved in manufacturing such a part - especially if you think about the fact that Nvidia has a tradition of designing its GPUs completely in software and a massive supercomputer system in Santa Clara, California. The company always has been proud of the fact that every GeForce chip you buy essentially is "first silicon."
Our sources state that the manufacturing cost of the GT200 die is somewhere between $100 to $110 per piece. It is pricey and you will be getting a lot more processing logic inside this core than with any other semiconductor part in the short history of the IT industry.
Typically, the CPU die spends about 66-80% of its real estate on cache and the remainder on logic ; in the world of GPUs it is vice versa. Our sources indicated that the number of transistors on the GT200 will settle somewhere in the neighborhood of 1 billion - north of 900 million and south of 1.1 billion.
The production cost of a chip is determined on the amount of chips that fit on a wafer. In light of the fact that a 300mm (12") wafer carried less than 120 G80 GPUs, Nvidia will see even fewer GT200s, apparently about 100, according to what he heard.
We can’t wait to lay our hands on one (or two) of these babies, and we’ll share our impressions as soon as possible.