Intel EIST SpeedStep

With the load results, there was an approximate 14°C difference in the temperature between 800MHz and 1.86GHz, and a 6-13 Watt difference. There was also an impressionable difference between the idle and load measurements. To see the performance impact upon the various SpeedStep levels, we had timed the LAME compilation process with GCC, timed the LAME encoding process, monitored the average frame-rate inside of Enemy Territory, and measured the Mflops using the Opstone Sparse-Vector Scalar Product benchmark. It is important to note, that if using the default SpeedStep configuration, the frequency and voltages will immediately return to their default values when launching any CPU intensive work. However, we performed these tests to see precisely the performance differences if you were to choose over-riding the levels in order to reduce heat, power consumption, and a cooler/quieter environment. All four of these tests are CPU centric, with three of them being real-world scenarios. Enemy Territory was chosen for the gaming benchmark, as it is largely CPU limited with current-generation graphics cards. The Opstone Sparse-Vector Scalar Product benchmark can be obtained from BlueSailSoftware, and during testing the Pentium 4 optimized build was used. During this portion of the testing, the computer was powered directly by the IBM power adapter.

To differentiate between the traditional SpeedStep/Geyserville Technology and Enhanced Intel SpeedStep Technology, EIST allows for multiple frequency/voltage levels instead of simply two, which was found on the original implementation. As with AMD Cool 'n' Quiet Technology, the increased number of power-states allows for superior performance control with optimal savings when it comes to the power consumption, battery life, heat output, and noise level. With Enhanced Intel SpeedStep Technology, the frequency switch is software controlled, rather than toggling the GHI# pin between two states, and is supportive under Linux (with the 2.6 kernel and cpufreq) and of course Microsoft Windows. Another benefit to EIST is that the states are changed by having the software write to the processor model specific registers (MSRs), which eliminates Chipset dependency. The processor also controls the voltage changes internally to ensure a safe and glitch free transition. As far as the lag goes when switching from a non-demanding task (i.e. browsing the Internet) to a demanding test (i.e. audio encoding) there is up to a 10 micro second delay during which the processor core(s) and L2 cache are unavailable. Enhanced Intel SpeedStep Technology also offers an improved thermal monitor mode. For this mode, when the on-die thermal sensor reaches a high die temperature, the processor has the ability to automatically switch down to a lower frequency/voltage in the model specific registers. Once the CPU die temperature drops to a lower thermal value, the transition to the previous frequency/voltage level occurs. Some of Intel's other low power features include AutoHALT Powerdown, HALT/Grant Snoop, Sleep, Deep Sleep, and Deeper Sleep. Looking over our EIST results today using a Pentium M 750 mounted inside of a ThinkPad notebook, Intel's technology certainly has its advantages. Switching between the five power-modes, there was evident change in the CPU temperature as well as the power consumption. In addition, the frequency/voltages immediately scaled appropriately when switching between demanding and non-demanding environments. For those, however, seeking to over-ride the switching process at a low frequency/voltage, in order to conserve the battery life, heat output, and noise will find the performance to be severely hampered. As the last portion of our tests had shown, the compilation time had increased 212%, encoding time 218%, frame-rate had decreased 5FPS, and the Mflops had dropped over 57%.