The Shape of Code

People buy a computer today because they have a problem that needs to be solved today, or in the near future. If today’s computer is good enough to solve today’s problem, then any expectation of the next generation being substantially faster is irrelevant.

If a cpu manufacturer (e.g., Intel) knows that the next generation, available in less than 18 months, is going to be substantially faster, how do they price the existing products over the months before and after the next generation is launched? Similarly, for memory chip manufacturing businesses (there used to be many) knowing that the next generation, in less than 18 months, is going to have twice the storage capacity, how the existing products be priced over the months before and after the next generation is launched?

During the 1990s desktop computers regularly became noticeably faster. Computers also regularly became faster in earlier decades, but early manufacturers built their own cpus, rather than selling them to others to integrate into a finished product. Yes, there were manufacturers selling cpus in the 1980s, but the market was fragmented, and any trends that might be present are obscured by market turbulence.

At the start of the 1990s, compiler messages appeared slowly enough that they could be read as they scrolled up the screen, but within a few years it was hard work spotting just a few words before the lines disappeared at the top of the screen. When working on a new machine for the first time, developers would be amazed at the rate compiler messages scrolled up the screen, but this rate quickly became the new normal.

During the 1990s, the performance of Intel’s x86 cpu family increased exponentially, and memory device capacity grew exponentially.

Manufacturer pricing data for cpu/memory chips in the 1990s is somewhat sporadic. However, end-user prices are listed in computer magazine adverts. The paper The economics of PC remanufacturing by G. Ferrer lists monthly end-user prices between August 1995 and January 1997, extracted from adverts in the British magazine Personal Computer World. I extracted prices for the same cpus/memories from the earlier/later editions available online. All monthly prices are averaged over all advertised prices found in each month’s edition.

The plot below shows monthly prices for Intel’s 80486DX2 and Pentium processors running at various clock rates; the extended straight lines are fitted exponential models (code+data):

When first released, higher performing cpus were sold at a premium price, which decreased (exponentially) at around 30% per year. Staying on the leading edge of computer performance required regularly paying premium prices to buy a new computer containing the latest cpu.

In the 1990s memory chips were packaged in SIMMs, of various form factors, which were superseded in the late 1990s by DIMMs, and then the DDR SDRAM we use today. The plot below shows the price of non-parity 72-pin SIMMs chips having a given capacity (code+data):

The decreases extend over a longer period than the cpu prices (presumably because people are adding more memory to older computers that are not being replaced).

The price changes could just as well have been plotted using a linear scale, as the exponential one used. In both cases there are distinct periods where the prices form a roughly decreasing line.

Why are manufacturers able to make such large reductions in the prices of the devices they sell?

Many semiconductor devices are fabricated on a large wafer. When the fabrication process for a particular device is new the yield may low, i.e., only a small percentage (perhaps 10%) of devices are functional. As production experience is gained the device yield increases and the production process is speeded up, i.e., more functional devices are produced for not much more cost.

There are two supplies for Pentiums, Intel and AMD, with Intel dominating, and having a large profit margin. There are many manufacturers of memory devices, it is a commodity market with slim profit margins.

An analysis of the extent to which profit margins and market structure influence pricing is outside my domain expertise.