03.08.2013 Views

Copyright by William Lloyd Bircher 2010 - The Laboratory for ...

Copyright by William Lloyd Bircher 2010 - The Laboratory for ...

Copyright by William Lloyd Bircher 2010 - The Laboratory for ...

SHOW MORE
SHOW LESS

Create successful ePaper yourself

Turn your PDF publications into a flip-book with our unique Google optimized e-Paper software.

Watts<br />

5<br />

4<br />

3<br />

2<br />

1<br />

0<br />

Figure 5.7 GPU Power Model (Non-Gated Clocks) – 3DMark06-HDR1<br />

5.3.6 Memory<br />

clock gating<br />

Memory or DRAM power consumption is one of the more variable subsystems. Similar<br />

to the CPU, the application of various power management features yields a wide range of<br />

power consumption. For example consider the standard deviation of power consumption<br />

in SPECjbb of 1.096W compared to average its average of 1.71W. This variation is<br />

caused <strong>by</strong> the three modes of operation: self-refresh, precharge power down and active.<br />

Self-refresh represents the lowest power state in which DRAM contents are maintained<br />

<strong>by</strong> on-chip refresh logic. This mode has a high entry/exit latency and is only entered<br />

when the system is expected to be idle <strong>for</strong> a long period. <strong>The</strong> memory controller selects<br />

this mode as part of its hardware-controlled C1e idle [Bk09] state. Since this state is<br />

entered in conjunction with the hypertransport link disconnect, the power savings can be<br />

represented using the LinkActive% metric. Pre-charge power down is a higher-power,<br />

lower-latency alternative which provides power savings <strong>for</strong> short idle phases. This<br />

allows pre-charge power savings to be considered with normal DRAM activity power.<br />

96<br />

Measured<br />

Modeled<br />

Error<br />

0 20 40 60 80<br />

Seconds<br />

100%<br />

50%<br />

0%<br />

-50%<br />

Error(%)<br />

-100%

Hooray! Your file is uploaded and ready to be published.

Saved successfully!

Ooh no, something went wrong!