EETimes: 'Costs cast ICs into Darwinian struggle'

EETimes recently published an interesting article on the rapidly rising costs of semiconductor fabs, process R&D and chip design. The comments in the article, by an employee of Synopsys and an analyst of VLSI Research, echo previous predictions made by TSMC's Chairman and Founder, Morris Chang. Basically, at the 45 and 32nm, nodes many smaller companies are likely to quit the scaling race (or at least scale down designs less aggressively) as their volumes won't be high enough to justify it.

This problem applies to both integrated device manufacturers (IDMs), such as Intel and Sony, and fabless design houses including NVIDIA, AMD's ATI division and Qualcomm. For the former, it means the one-time R&D development costs will be higher, although this is partially offset by the greater number of companies that enter in process research alliances, such as the one between IBM, AMD and Chartered Semiconductors.

Even disregarding process research costs and mask costs, however, the largest 32nm fabs might cost from $5B to $10B; numbers which have never been seen before in the industry. Even comparing fabs of equal capacity would certainly see costs that are a fair bit higher than today's, and this would likely translate into higher wafer prices at foundries for fabless manufacturers, in order for them to amortize their investment.

Furthermore, design costs are poised to rise substantially. While the semantics there aren't too clear, it can be assumed that this implies that a chip of similar 'complexity' will cost more to design than in previous process generations. Predicted costs of $20M to $50M for a given chip design on 45nm could go up as high as $75M on 32nm, according to some pessimistic analysts. Fab and design costs have been increasing for a long time now, but at these levels design dynamics are getting a lot more problematic than previously.

And to finish on more of a GPU-oriented note, one of the most direct implications of this is that the volume and revenue requirements to get a proper return on investment for any given chip will be much higher. As such, this is one of the various factors that will ultimately kill the monolithic ultra-high-end GPU chip, which will then get replaced by a SLI or Crossfire variants of the mid-end or high-end chips. One recent example of that is the GeForce 7950 GX2, and it is rumoured that AMD's R700 will go in the same direction.