In this section we focus on the necessity for throttling the memory bus. The material forms the basis for our discussions in the rest of the paper.

Figure 1: Schematic showing main components in current-generation laptops
click image for larger view
 

Figure 1 shows a high-level view of the main components on a mobile platform built on Intel® Centrino® Duo mobile technology. The system memory gets accessed in almost all activities taking place in the platform. All data transfers to and from the system memory are managed by the Intel® chipset. Thus, if the chipset/system memory is idle, the platform itself is generally in an idle state. But if there is activity on the chipset/system memory, the platform is consuming more power, causing the chipset and memory components to heat up.

As mentioned earlier, mobile platforms have limited cooling capabilities. In the past, memory speeds and capacities generally allowed the system memory subsection to stay within the cooling limits of the platform, and the DRAM case temperatures were below the maximum operating specifications of 85°C. But with the increase in memory capacity and speed, we are now reaching the point where memory thermals are starting to exceed the cooling capabilities of mobile systems. When the cooling budget is exceeded, that means the system is no longer able to cool the memory subsection, and DRAM case temperatures begin to exceed their maximum case temperature specification.

In a recent laboratory study, several different thin-and-light laptop designs were tested with various SO-DIMMS of different memory speeds, capacities, densities, and vendors, using OpenGL Benchmark software. This software has high bandwidth utilization (about 52 percent of theoretical max) and causes DRAM devices to draw constant high power.

Figure 2: Lab data showing memory bandwidth and DRAM case temperatures
click image for larger view
 

In Figure 2, temperature and bandwidth data are dependent on system, memory, and software configuration. The figure shows the results of a small sample of this thermal study. All three systems are thin and light designs: each system is from a different vendor. System A platform layout has one memory module on top and one on the bottom, System B platform layout has both memory modules on top, and System C platform has both memory modules on the bottom.

The results of the above study show that typical 512 MB SO-DIMMs in various thin-and-light notebook designs are well below 85°C and do not appear to be in jeopardy of exceeding their thermal limits. In some cases, however, 1 GB-capacity SO-DIMMs are reaching, and sometimes exceeding, their maximum case temperature specification of 85°C. These results show that memory modules are already operating near their maximum specifications, and as memory power and thermals increase with system capacity and speed, memory modules will begin exceeding their maximum specification with multiple realistic workloads. Small form factor designs are of even greater concern due to their lower cooling budgets and thermally challenged environments. There is clearly a need for a robust thermal management solution.

If left unchecked, DRAM devices will start running above their maximum operating case temperatures, and memory-related reliability issues will begin to crop up. Thus the memory bus needs to be throttled to ensure that the DRAM devices operate within their thermal limits. Memory throttling provides a solution to cool the DRAM devices by reducing memory traffic allowed on the memory bus, thereby reducing the power consumed by the DRAM devices and thus reducing thermal output.

We now describe the two throttling techniques in detail.