Social aspects of chip design
michaelnielsen.org
michaelnielsen.org
- chip architects, usually PhD's in computer arch or really senior engineers + people to run performance simulators
- implementors/logic designers: take the architecture, divide it into blocks (instruction fetch unit, memory interface, etc), create a microarchitecture and code it in RTL (register transfer language, Verilog)
- verification engineers: read the specification and write software+tests to exercise corner cases. verification is usually staffed at twice the number as implementors
- layout/backend engineers: synthesize the RTL into logic gates and/or draw custom transistor circuits where speed requires it
- circuit designers: design phase lock loops, clock paths, pad drivers, etc.
I don't know the breakdown at Intel, but a new chip architecture can take on the order of 500 people. The floating point bug would have been in the floating point unit, so probably at most less than 10 of the logic designers would have been aware of the design.
As an aside, what was amusing to me were the estimates of the probability of errors that various voices were making: Byte made one, I remember IBM making one. All of the estimates presumed that the floating point calculations made were somehow randomly distributed, which is very clearly not the case.
Robert Colwell tells the story of the P6 architecture with a collection of anecdotes. It's full of insights and easy to digest at the same time.
For example, many of the issues in processor design stem from miscommunication about how functional units communicate. Since units on a processor can only communicate with their neighbors, they arranged the teams in the office to reflect the actual die layout. This way, teams responsible for a functional unit could always communicate directly with their on-die neighbors!