|
|
|
|
|
|
|
today's design tools, this is not a difficult problem, but it does require more development and simulation time to ensure that these verifications are performed. Because the pipeline operates at a higher frequency due to the super-pipelining, testing the product also requires higher-speed test equipment on the fabrication and production floorsand will possibly require the purchase of higher-performance equipment to meet the testing requirements. For purposes of this study, we will ignore these issues, but they must be evaluated in a real design situation to ensure that engineering does not design a product that cannot be easily or cheaply transferred into manufacturing. |
|
|
|
|
|
|
|
|
One aspect of manufacturing that this discussion has not addressed is that the cost of the processor does not necessarily have a significant effect on the price of the final system, which is driven by both manufacturing and marketing issues. Assuming that marketing has no impact on the initial price estimate, using a 3- to 5-times markup will provide a typical approximation to this initial price. These two processors give $149 to $455 and $310 to $516 differences in price between the original system using the Baseline Mark I processor and the two options for the Baseline Mark II processor. This difference, not more than $500, is clearly not too significant for a $5,000 to $10,000 workstation. Even if one processor cost double the other processor, the result would still be insignificant. The lesson here is that processors have a minor effect on the final price of the system and, although the cost of a given component is important, it may not necessarily be the most important factor in component selection. As a caveat, this is often true only when comparing very similar componentsfor example, comparing a high-performance and a low-performance processor also requires comparing the costs of the supporting design elements, whereas comparing two comparable-performance processors need consider only the costs of the two processors. |
|
|
|
|
|
|
|
|
Another aspect of the selection is in the response of the system to future changes in the overall Baseline systemfor example, improving the physical or virtual memory systems performance while keeping the processor the same. This is a significant problem that should be considered before making a final selection. Looking at the performance calculations, we can specify the performance of the machines as: |
|
|
|
 |
|
|
|
|
Preduced-scale = 1.811 + 0.132Pcache-miss + 0.0076Ppage-fault |
|
|
|
|
|
|
|
|
Psuper-pipelined = 1.448 + 0.168Pcache-miss + 0.0102Ppage-fault, |
|
|
|
|
|
|
|
|
and solving for the break-even performance point, we get |
|
|
|
|
|
|
|
|
Figures 10.4, 10.5, and 10.6 show the performance for these processors, with varying page fault penalties for three cases. Figure 10.4 is representative of performance with the addition of a second-level cache off-chip that reduces the effective first-level cache miss rate from 5 cycles to 2.5 cycles; Figure 10.5 is the current configuration, and Figure 10.6 is a configuration with a lower-performance memory system installed and a first-level cache miss rate of 7.5 cycles. In the first two cases, it is clear that for almost any credible page fault penalty, the super-scalar processor is a clear winnerthe |
|
|
|
|
|
