So far, graphics specialist NVIDIA (NASDAQ:NVDA) has released two high-end graphics processors for PC gamers: GeForce GTX 1070 and GeForce GTX 1080. Both of these are based on the same GP104 chip design, although the one in GTX 1080 is fully enabled while the one powering the GTX 1070 has functional units disabled.
There has been some speculation that the company's next graphics processor, the GeForce GTX 1060, will also be based on an even further cut-down variant of the GP104. Here's why I don't think this is likely.
Basic semiconductor economics
It is very often the case, especially with products made up of a bunch of the same structure repeated multiple times (like graphics processors), that a partially defective chip can still be sold. In fact, I doubt that at this stage of the game the GeForce GTX 1070 is made up of GP104 chips that are fully functional but have been artificially disabled for segmentation purposes.
This is a good way to get more saleable chips from a given semiconductor wafer, which ultimately helps boost gross profit margins.
However, as chip companies get better at building their products, the percentage of the chips from a given wafer that are fully functional, or nearly fully functional, goes up significantly. At some point this strategy breaks down as the company winds up disabling fully functional and relatively expensive-to-make chips to serve a much lower price point.
Instead, it makes more sense for said company to build smaller chips designed specifically for a given target market and, ultimately, price point.
Applying this to the GeForce GTX 1060
According to a table posted on Chip Hell, the GeForce GTX 1060 will pack 1280 CUDA cores, 80 texture units, and 48 render output units. The table also says that it'll be based on the GP104.
For reference, the full GP104 used in the GeForce GTX 1080 packs 2560 CUDA cores, 160 texture units, and 64 render output units. Aside from the render output units, the GTX 1060 looks like half of a GTX 1080.
If this were the case, NVIDIA would essentially have to sell GP104 chips with virtually half of the chip disabled. If GP104 yields are poor right now, then this could make sense. However, with CEO Jen-Hsun Huang going on record to state that yields on these products are "good," this doesn't make a lot of sense.
Why wouldn't it make sense, then?
First off, the GTX 1060 is likely to be priced at or below $300, so it is likely to sell in far greater volumes than the $379-plus GeForce GTX 1070 and the $599-plus GeForce GTX 1080. NVIDIA's manufacturing yields on the GP104 chip would have to be quite bad in order for it to make sense to ship so many chips with virtually half of the chip turned off.
It is much more likely that NVIDIA will bring out a smaller, cheaper-to-build processor to serve the market for graphics cards cheaper than the GeForce GTX 1070. The entire lineup of Pascal-architecture-based graphics cards is now known, and the GP106 is likely to be a much more suitable part to serve the role of a GeForce GTX 1060 than a heavily cut-down GP104 is.