We have discussed our power-management techniques primarily in the context of RDRAM architecture, but they are also applicable to other SDRAM architectures that support multiple operating power levels. In Figures 8(a-c), energy consumption is shown when running the same set of workloads discussed above, assuming a 4-node DDR memory configuration using PAVMr2. The cumulative energy is normalized against the default hardware-implemented policy for DDR. We still obtain significant energy savings, but not as much as with the previously-described RDRAM configuration.
There are two reasons for this. First, the power difference is much smaller between the DDR modes that correspond to RDRAM's Standby and Nap modes. Therefore, for DDR, putting nodes in ``Nap'' mode shows a smaller relative energy savings. Second, and more importantly, the notion of a node is coarser-grained for DDR than RDRAM. As discussed earlier, power management for DDR can only be done at the module-level, whereas in RDRAM, power can be adjusted at a device-level granularity, resulting in a much larger number of nodes.
To show the effect that the number of nodes in the system has on energy savings, in Figure 8, we also compare energy consumption assuming 1-node and 16-node DDR configurations. For a 1-node configuration, PAVM basically degenerates to the On/Off policy, since the only node must be active for all processes. As the number of nodes increases and the size of each node decreases, PAVM has finer-grained power management control, and yields greater energy savings. Once the number of nodes is increased beyond a certain point, we would expect decreasing, and possibly negative, marginal returns due to operational overheads of managing a large number of nodes. Finding the sweet spot that provides the maximum energy savings is system-/memory- dependent and beyond the scope of this paper. However, we believe that the 8- to 16-node granularity provided in most RDRAM configurations is not far from this sweet spot for typical mobile workloads. Furthermore, assuming that 4 nodes are available in a DDR system is probably optimistic, since in real systems, we are more likely to see 1-node and 2-node configurations, especially on mobile platforms. The results for SDR is similar to DDR, and due to the space limitation, are not shown here.