21.01.2013 Views

Lecture Notes in Computer Science 4917

Lecture Notes in Computer Science 4917

Lecture Notes in Computer Science 4917

SHOW MORE
SHOW LESS

You also want an ePaper? Increase the reach of your titles

YUMPU automatically turns print PDFs into web optimized ePapers that Google loves.

202 P. Raghavan et al.<br />

power consumption of the complete system can be analyzed and optimized (either by a<br />

change <strong>in</strong> architecture or a compiler optimization or transformation).<br />

To validate our framework, the energy consumption of an <strong>in</strong>-house processor<br />

SyncPro[13] 2 runn<strong>in</strong>g WLAN Synchronization was compared to same processor modeled<br />

<strong>in</strong> the proposed COFFEE flow. The processor was designed separately and there<br />

was no shar<strong>in</strong>g of tools/sources between this design and our flow. As a reference, detailed<br />

gate-level simulation (after synthesis, placement+rout<strong>in</strong>g and extraction) were<br />

performed on WLAN synchronization code. The power was then estimated us<strong>in</strong>g Prime-<br />

Power. For more details on the precise flow used for estimat<strong>in</strong>g energy consumption<br />

for SyncPro, the reader is refered to [13]. The same code was then run on the COFFEE<br />

framework to estimate the power consumption. The net error <strong>in</strong> the estimated energy<br />

us<strong>in</strong>g the COFFEE estimation and the energy estimate from PrimePower and gate level<br />

simulation is less than 13% for the complete platform (<strong>in</strong>clud<strong>in</strong>g the memories). More<br />

validation po<strong>in</strong>ts aga<strong>in</strong>st other processors are currently be<strong>in</strong>g added.<br />

5 Experimental Setup and Results<br />

In this section we demonstrate the COFFEE framework on a representative benchmark<br />

for two state of the art processors (TI’s C64 and ARM’s Cortex-A8). We further illustrate<br />

the potential of architectural design space exploration on a standard embedded<br />

VLIW.<br />

5.1 Benchmark Driver: WCDMA<br />

WCDMA is a Wideband Direct-Sequence Code Division Multiple Access (DS-CDMA)<br />

system, i.e. user <strong>in</strong>formation bits are spread over a wide bandwidth by multiply<strong>in</strong>g the<br />

user data with quasi-random bits derived from CDMA spread<strong>in</strong>g codes. WCDMA [22]<br />

is one of the dom<strong>in</strong>ant 3G cellular protocols for multimedia services, <strong>in</strong>clud<strong>in</strong>g video<br />

telephony on a wireless l<strong>in</strong>k.<br />

Frontend<br />

Tx Filter<br />

WCDMA Transmitter<br />

Scrambler Spreader Interleaver<br />

Channel<br />

Encoder<br />

Rx Filter Demodulator De<strong>in</strong>terleaver Channel Decoder<br />

WCDMA Receiver<br />

Fig. 5. Wideband CDMA as <strong>in</strong> Signal Process<strong>in</strong>g On Demand (SODA, [23])<br />

2<br />

[13] is a 5 issue SIMD VLIW and heterogenous distributed register file and a novel <strong>in</strong>terconnection<br />

network.<br />

To higher layers

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

Saved successfully!

Ooh no, something went wrong!