Revisiting Voltage, Frequency and Power
Figure 4 – Frequency Scaling Characteristics of the CELL processor
Since the announcement of Playstation3’s technical specifications, some hardware enthusiasts have expressed their disappointment in the fact that the CELL processor in Playstation 3 will be limited to operate at 3.2 GHz as opposed to the 4 GHz discussed at the CELL processor’s introduction (never mind the 5.6 GHz discussed in the first installment of this series of CELL processor articles). However, Sony’s choice to reduce the operating frequency of the CELL processor to 3.2 GHz is easy to understand once few basic facts are taken into account. First, power consumption characteristics of modern processors are (to a first order approximation) linearly proportional to the frequency, and to the square of the voltage. Hence, a simple reduction of frequency from 4 GHz to 3.2 GHz at the same supply voltage can reduce power consumption by 20%. However, if the supply voltage can also be reduced, the power reduction can be far more dramatic than 20%.
Figure 4 illustrates the measured frequency and voltage scaling characteristic of one sample CELL processor. Figure 4 shows that with a supply voltage of 1.1V, the CELL processor that was tested has a maximum operating frequency that is well above 4 GHz. Conversely, Figure 4 also shows that if the maximum operating frequency of this CELL processor is limited to 3.2 GHz, then the supply voltage can be reduced to approximately 0.9V (not quite 0.9V, but with some process improvement, 3.2 GHz should be well reachable with a 0.9V supply voltage). Reducing the frequency of the CELL processor allows the system design engineers to also lower the processor’s supply voltage, and the combined effects of the lower frequency and lower supply voltage can lead to a 46% reduction in power consumption. Previously, power consumption of the CELL processor was estimated to be in the range of 50 to 80 W at 4 GHz and 1.1V. In subjecting the estimates to the 46% scaling factor, we further estimate that the power consumption of the CELL processor should be in the range of 27 to 43 W at 3.2 GHz and 0.9V, and 113 to 181W at 5.6 GHz and 1.4V. The presentation given by Takahashi-san from IBM confirms that at least one CELL processor was able to reach 5.6 GHz with a supply voltage of 1.4V. Although the power consumption for the CELL processor in this configuration was not released, it is intriguing to note that the 181W figure correlates with the previously unconfirmed report of 180W at 5.6 GHz and 1.4V.
More on the SPE Schmoo Plot
Figure 5 – Schmoo Plot for the SPE from ISSCC 2005
Figure 5 shows the schmoo plot for the SPE that was released at ISSCC in 2005. However, it is unclear how the power consumption and temperature measurements were taken, as the methodology was not fully revealed. Fortunately, the methodology was specifically addressed by Takahashi-san in his presentation at Cool Chips. Specifically, he wrote:
“The temperatures [are] measured by an on-chip thermal sensor. The power is measured at a supply by subtracting the power number of 1 SPE activated from that of 2 SPE activated at the same time. The power reflects a SPE running a single precision floating point intensive lighting and transformation workload using 16 KB portion of Local Store memory.”
The description of the methodology supplements the previous description of the schmoo plot given at ISSCC:
“. . . [The] voltage versus frequency schmoo shows SPE active power and die temperature while running a single precision intensive lighting and transformation workload that averages 1.4 IPC. This is a computationally intensive application that has been unrolled four times and software pipelined to eliminate most instruction dependencies and utilizes about 16 kb of LS.”