Image source: NVIDIA.
Graphics specialist NVIDIA (NVDA 1.07%) recently announced a new mid-range graphics processor known as the GeForce GTX 1060. The chip is built on Taiwan Semiconductor Manufacturing Co.'s (TSM +0.34%) 16-nanometer manufacturing process, representing a much needed, and long awaited, transition away from Taiwan Semi's venerable 28-nanometer technology.
Let's explore the cost structure of the GTX 1060 relative to its predecessor, the GeForce GTX 960, and the potential associated margin implications to NVIDIA.
Chip cost goes up
The GTX 960 measures in at 228 square millimeters on TSMC's 28-nanometer manufacturing process. The GTX 1060, meanwhile, comes in at 200 square millimeters on TSMC's newer 16-nanometer manufacturing process.
Semiconductor industry analyst Handel Jones estimates the cost of a 16-nanometer wafer to a fabless customer by the end of 2016 at $7,779.22 -- an increase of about 70% from his 28-nanometer wafer cost estimate of $4,577.25. What this means is that the effective cost per millimeter of silicon actually increases generation-over-generation.
In my estimate of the GeForce GTX 1080's cost structure, I used Silicon Edge's dies per wafer estimation tool to figure out roughly how many of each chip NVIDIA will be able to get from a 300 millimeter wafer. For the 1060, that figure works out to 293; for the 960, I get around 256.
If we assume that 80% of the GTX 1060s are saleable, then that's around 234.4 good chips per wafer. For the purposes of this estimate, since Taiwan Semi's 28-nanometer process is so mature, let's say that 90% of the GTX 960s from a given wafer are functional, or around 230 chips.
Taking the respective wafer cost estimates and dividing them by the number of good chips yields cost-per-chip estimates. For the 1060, that's $33.30 per chip, and for the 960, it's $19.90. The raw die cost of the GTX 1060 appears to be higher than its predecessor.
The GTX 1060 packs more memory, too
Since these chips aren't sold as raw chips to end users but, instead, as chips mounted onto boards with many different components, it's useful to look at the generational changes in the surrounding components.
The GTX 1060 comes with 6 gigabytes of GDDR5 memory rated at 8 gigabits per second. The GTX 960 initially debuted with 2 gigabytes of GDDR5 rated at 7 gigabits per second. NVIDIA is likely using denser memory chips with the 1060 relative to the 960, so it's not likely that memory costs have tripled. However, it would appear that generation-on-generation, the memory dollar content has gone up.
Cost structure up, but pricing moves up, too
I believe that in going from the GTX 960 to the GTX 1060, NVIDIA's cost structure moves up. If the company and its partners were aiming to sell the 1060 at the same prices that they sold the 960 for, this would suggest margin compression for the graphics specialist.
However, NVIDIA says that boards based on the GTX 1060 will start at $249, a $50 premium to what the GTX 960 sold at when it first launched. It stands to reason, then, that although the GTX 1060 is pricier to build than its predecessor, the company likely believes that the additional performance will allow it to successfully command higher prices. The higher pricing likely offsets the increased manufacturing costs, which may mean that the 1060 is gross profit margin percentage neutral relative to the 960.
