Applying a centuries-old technique to modern cost estimation.
The following letter was published in the Letters to the Editor in the November 2013 CACM (http://cacm.acm.org/magazines/2013/11/169039).
The viewpoint "The Value of Microprocessor Designs" by Ana Aizcorbe et al. (Feb. 2013) aimed to analyze the value of microarchitectures in isolation, as though they could be mixed-and-matched with various silicon implementation technologies over the years. This is a nonsensical proposition. For example, a Pentium III microarchitecture could not have been realized in 0.8u technology due to insufficient transistors. A 486 microarchitecture could be instantiated on, say, 90nm technology but would have been too slow to be competitive. And the design trade-offs baked into the 486 pipeline, appropriate for the silicon of the 1980s, would not leverage the much larger transistor budgets of the 1990s and later.
These microarchitectures were not independent of one another, as Aizcorbe et al. implicitly assumed. The Pentium III was the same microarchitecture as the Pentium II but with SSE instructions added. Moreover, both came from the original Pentium Pro P6 microarchitecture. The Pentium-M was also a descendant of the P6. The Pentium 4 microarchitecture was substantially different, but, as chief architect of both P6 and Pentium 4, I can testify that the Pentium 4 was not unrelated to P6.
Aizcorbe et al. also said, "The Pentium 4 design provided very little value to Intel" due to "overheating." Incorrect. The Pentium 4 was designed to the limits of air-cooling but not beyond those limits. It did not overheat nor suffer heat-related reliability issues. Had the Pentium 4 been designed with a much different microarchitecture, Aizcorbe et al. suggested Intel might have earned much higher profits. What other microarchitecture? They shed no light on this question. Neither Aizcorbe et al. nor Intel can possibly know what might have happened had the Pentium 4 been designed another way.
They also missed another important real-world issue: fab capacity and opportunity cost. Chip manufacturers can always lower the price to make demand go up, but the bottom line might suffer. Our goal in the Pentium product line was to design chips that would command high prices but also generate high demand. And we did it in the face of a market that evolved dramatically from the 486 days, when only engineers and scientists could afford PCs, through the early 1990s, when home computers were becoming common, through the late 1990s, when the Internet took off.
Aizcorbe et al. did mention such influences but did not take them into account in drawing their conclusions. Each of these market exigencies influenced the design, as well as the final results, in special ways. Microarchitectures are intrinsically bound to their markets, and their implementation technologies and are not fungible per their assumptions.
Bottom line, Aizcorbe et al. observed that Intel did not earn the same profits on all its products over the years. Some, it turns out, paid off more than others. We can agree on that much. But saying so is not news.
Robert (Bob) Colwell
We welcome Colwell's comments but disagree on one major point: Our methodology did not mix-and-match different microarchitectures and silicon implementation technologies, as he suggests. We used only those combinations actually produced by Intel at a given point in time. However, our approach does hinge on the assumption that even when a new microarchitecture is available, Intel cannot retrofit all its current capacity to use it. As a result, old and new microarchitectures are used concurrently in production, allowing us to infer the incremental value of the new microarchitecture.
New Haven, CT
Displaying 1 comment