Moore’s Law & Semiconductor Pricing
Evolution of Moore’s Law
In 1965, Gordon Moore (then a part of Fairchild Semiconductor) published a paper titled “Cramming More Components onto Integrated Circuits” where he described in detail how unit cost per component falls while the number of components in an integrated circuit increase. As per Moore’s theory, cost per component is inversely proportional to the number of components in a circuit. However with increased components (increased complexity) the yield (number of good circuits per 100 circuits manufactured) reduces, which, in turn, increases the cost per component. Thus the cost is acted upon by two counteractive forces resulting in a minima point: “The number of components in the circuit at which the cost per component is minimum”. As per Moore this minimum point increases by a factor two every year. In 1975 this was revised to indicate an increase by a factor of two every twenty-four months.
To understand let’s consider an example: if cost per component were minimized at 50 components/circuit in 1965, then the cost per component would be minimum at 100 components/circuit in 1966. However a question remains open: Will the area of an integrated circuit double with the doubling of components? The same question is answer by Robert H. Dennard in his famous Dennard Scaling Law: Every successive generation the area per transistor (component) lowers by 50% while the power consumption reduces by 50%. Thus if we combine the two laws: Every year (or two years), the number of transistors (components) per unit area of circuit doubles, while the power consumption remains constant.
In terms of costs it can be interpreted as follows: If an integrated circuit with X components cost $Y to make in a particular year, then the same circuit should cost $(Y/2) to make two years later. Unfortunately, semiconductor customers tend to get fixated at times on the “cost-halving phenomenon”.
How Moore’s Law impacts semiconductor pricing:
When I say “fixated”, I imply holding on to a law without considering the limiting and boundary conditions. It is common to hear comments like: “All that I am asking for is a Moore’s law-based reduction.”
A Moore’s Law-based price reduction makes sense when the following criteria are met:
- The supplier (semiconductor chip maker) is moving to a new process technology every two years.
- The customer accepts a new process technology chip as and when introduced.
- The new process technology chip is identical in architecture and functionality as it previous-generation counterpart.
Rarely does such an ideal condition prevail.
A supplier doesn’t move to a new process technology every two years for all its products. When Wikipedia talks about Semiconductor Fabrication they mention the following table:
At first glance, it seems that in a given year all semiconductor chips migrated to the assigned process technology. The reality, however, is different. While some chips have indeed migrated to 14nm we still have chips where the most advanced process technology is much more primitive. As an example let’s consider SRAMs. Most advanced SRAMs are 65nm today. Apart from commercial limitations, there are technological limitations that prevent chips from shrinking as per Moore’s Law. Hence using Moore’s Law-based reduction models may not work where the process technology has not sufficiently reduced.
Customers don’t always accept new process technology chips as and when introduced. When a chip migrates from one process technology to another, not all customers adopt the new chip instantly. The new chip may be pin compatible to its predecessor, but there would be differences in terms of characteristics. Thus the new chip has to be qualified afresh. Customers do shy away from the hassle especially if their qualification process is lengthy and cumbersome. Suppliers also tend to shy away from insisting much because requalification opens up the socket for competition chips. Thus using Moore’s Law-based reduction models may not work if the chip used is not changing.
A special mention could be made here about some Hi-Rel customers (for example auto customers). The semiconductor chips that are automotive qualified are sold to automotive tier-1 suppliers. Tier-1s in turn use the same chip for multiple OEM (Original Equipment Manufacturer or Car-Maker) projects. Thus changing the chip requires them to go through re-qualification by most or all OEMs. Promoting a new process technology chip to Automotive Tier-1 suppliers is therefore highly challenging. On the other hand, the Tier-1s themselves are under cost-reduction pressure for OEMs (linked to industry expected economies of scale and learning). This makes the business trickier, and it is important to freeze price roadmaps with automotive customers at the very onset of a project.
The new process technology chip is not always identical in architecture and functionality to its previous-generation counterpart. Most often nowadays companies don’t invest on solely migrating chips from one technology to another. Instead, the next generation chip can be a better chip in terms of functionality. Thus comparing it to its predecessor would not be fair.
In case of semiconductor chips, when it comes to new business, the prices are market-driven for commodity products and value-based for proprietary products. Moore’s Law-centric discussion comes in mostly during price negotiations for existing business. When such a topic comes up during negotiation, it can be countered by explaining the limitations of correlating pricing with Moore’s Law.
If the customer is buying older process node parts and a new revision is available, the best counter would be to offer some reduction on the older revision while making an attractive price offer on the new revision. That works two ways – it saves the product margin of the older revision part, and it also incentivizes the customer to qualify the new revision part for new projects. It quite helpful in cementing a long-term relationship with the customer.
If the customer is buying the latest revision part and as a company you don’t intend to move to a more advanced process tech in future, it should be clearly stated with the rationale laid out. It could be due to technology limitations (e.g. poorer yields on newer revisions) for the particular product, or it could be a well thought-out business strategy. If the industry trend for the product family is such then it may be easy to explain. Possibly the TAM (Total Allocated Market) is shrinking, leaving very few suppliers in the market, which makes further R&D for the product carry little business sense. In this case, the customer would prefer avoiding friction to avoid future supply risks.
However, if competitors are migrating to a newer process node you as a company are left with few options. Either you follow suit to avoid serious cost disadvantage (and potential future price wars) or restrict your market aspirations to select high price sockets.