15.01.2013 Views

U. Glaeser

U. Glaeser

U. Glaeser

SHOW MORE
SHOW LESS

Create successful ePaper yourself

Turn your PDF publications into a flip-book with our unique Google optimized e-Paper software.

Kevin J. Nowka<br />

IBM Austin Research Laboratory<br />

© 2002 by CRC Press LLC<br />

13<br />

Issues in High-Frequency<br />

Processor Design<br />

13.1 The Processor Implementation Performance<br />

Food-Chain<br />

Placement and Routing: Distance and Wire Loads Are<br />

Performance • Fast Wires and Fast Devices: Gain Is<br />

Performance • Gate Design: Boolean Efficiency Is<br />

Performance<br />

13.2 Noise, Robustness, and Reliability<br />

13.3 Logic Design Implications<br />

13.4 Variability and Uncertainty<br />

13.5 Summary<br />

Successful design of high-frequency processors is predominantly the act of balancing two competing<br />

forces: striving to exploit the most advanced technology, circuits, logic implementations, and system<br />

organization; and the necessity to encapsulate the resulting complexity so as to make the task tractable.<br />

This chapter addresses some of the compelling issues in high-frequency processor design, both in taking<br />

advantage of the technology and circuits and avoiding the pitfalls.<br />

Advances in silicon technology, circuit design techniques, physical design tools, processor organization<br />

and architecture, and market demand are producing frequency improvement in high-performance microprocessors.<br />

Figure 13.1 shows the anticipated global and local clock frequency of high-performance<br />

microprocessors from the SIA International Technology Roadmap for Semiconductors. 1,2 Because silicon<br />

technology continuously advances, it is necessary to either define high frequency at each time or define<br />

it in a technology-independent manner. For the remainder of this chapter, high frequency will be defined<br />

in terms of the technology-independent unit of fanout-of-4 (FO4) inverter delay. 3 Figure 13.2 presents<br />

the expected global clock frequency in terms of the ITRS gate delays. 1,2 From this figure, it is apparent that<br />

the local cycle time of high-performance microprocessors is expected to shrink by about a factor of two<br />

in number of gate delays. The ITRS gate delay is approximately a fanout-1 inverter delay, which is roughly<br />

a fixed fraction of one FO4. This cycle time improvement must be provided by improvements in the use<br />

of devices and interconnect, circuits, arithmetic, and organizational changes.<br />

This chapter will concentrate on high-frequency designs, currently defined as less than 18 FO4 inverter<br />

delays for a 64-bit processor and 16 FO4 for a 32-bit processor. These break-points are chosen because<br />

(1) representative designs have been developed, which satisfy these criteria, 4–7 (2) they are sufficiently<br />

aggressive to demonstrate the difficulties in achieving high-frequency designs, and (3) they fall firmly<br />

within the expected targets of the high-performance microprocessor roadmaps.<br />

In the remainder of this chapter, issues related to the design of processors for these high-frequency<br />

targets will be described. The ultimate dependence of the achievable cycle times on interconnect efficiency<br />

on low latency circuits will be discussed, and potential problems will be described.

Hooray! Your file is uploaded and ready to be published.

Saved successfully!

Ooh no, something went wrong!