|
|
|
| Table 10.11 Given area specifications. | | Feature size | | | Defect density | | | Yield | | | Control overhead | | | Aspect ratio overhead | | | Latch overhead | | | Bus overhead | |
|
|
|
|
|
|
Fortunately, area statistics for our baseline model are available. These are given in study 2.3. (Note the added branch adder plus its overhead.) A quick summary is provided in Tables 10.1210.14. |
|
|
|
|
|
|
|
|
Description of Calculations |
|
|
|
|
|
|
|
|
Each module is shown with its corresponding area utilization in "A," where "A" is defined as 1481 rbe and each rbe is 675 f2. The various units that belong in the baseline processor are summed up. The only addition in this implementation is the branch adder, which was added to enhance the CPI. After obtaining the total area for both integer and floating-point units, the overhead for both latches and buses for communication is added. These are 10% and 40% of the units, respectively. As indicated, the baseline implementation requires 82.35 A (80.9 A as in study 2.3, plus 1 A for branch adder and about 0.57 A for branch adder overhead). It is given that the die area is 230 mm2. To account for peripheral area such as I/O pads, 20% of the die area is subtracted, which gives us 184 mm2. Now we must convert the available area from mm2 to units of A: |
|
|
|
 |
|
|
|
|
Area (A) = (area) mm2/ (1481 * 675 * f * f * 0.001 * 0.001). |
|
|
|
|
|
|
|
|
Subtracting the baseline area and deducting 10% for aspect mismatch gives the area available for cache plus directory. A similar analysis is used for superscalar and multiprocessor implementations of the chip. The basic parameters are given in Tables 10.1110.14. |
|
|
|
|
|
|
|
|
Based on the area calculations in Tables 10.1210.14, we explore different cache configurations for the available area. |
|
|
|
|
|
|
|
|
Whenever data needs to be sent from one entity to another, a simple solution is synchronous communication: when one entity is ready to send and the other to receive, the data element is transferred. If more flexibility is required, a buffer needs to be introduced to allow asynchronous communication. A buffer allows processes of similar bandwidth to proceed in |
|
|
|
|
|