Next: Conclusions
Up: Experimental Results
Previous: Total power reduction
Without access to source codes of commercial tools, we could not
incorporate interconnect-awareness into them as we did to SCALP.
Moreover, most commercially available tools do not take directives
from RTL floorplanners to lay out a circuit. Therefore, physical
locality we try to maximize at RTL may be easily destroyed when
commercial tools put an IAPO RTL circuit through logic and
physical synthesis. Nevertheless, we were interested in whether a
RTL circuit synthesized using our techniques would be any
different if it was handed out to commercial tools for logic and
physical synthesis. We employed Synopsys Design Compiler to
synthesize both IUPO and IAPO RTL circuits, and employed Cadence
Silicon Ensemble for physical synthesis. A standard cell library
for TSMC 0.25m process developed by the VTVT group in
Virginia Tech. was used. Table IV shows the number
of semi-global and global wires (wires longer than 1m) in the
benchmarks synthesized by commercial tools. Surprisingly, 7 of the
8 benchmarks demonstrate fewer such wires for IAPO designs.
Although these commercial tools may destroy the circuit hierarchy
and do not take any floorplan hints from the RTL design, our
techniques are powerful enough to have positive impact in most of
the benchmarks. That is, a design with a better RTL interconnect
topology turns out to have a better interconnect topology after
final routing. (For the benchmark , the negative impact may
be contributed to its big area overhead of IAPO over IUPO, which
may overwhelm the interconnect topology improvement when going
through the RTL-hint-free commercial tools.) As many commercial
tools are going to take RTL directives for logic and physical
synthesis, we expect our techniques to be more beneficial when
working with commercial tools.
Table IV:
Number of semi-global and global wires in
commercial-tool synthesized IAPO and IUPO designs.
|
Chemical |
DCT_IJPEG |
DCT_Wang
|
Diffeq |
Elliptic |
IIR77
|
Jacobi |
WDF |
IAPO |
8720 |
15298 |
13037 |
2872 |
11520 |
13919 |
11264 |
7682 |
IUPO |
8770 |
17419 |
13999 |
2942 |
11726 |
18499 |
11584 |
6406 |
Next: Conclusions
Up: Experimental Results
Previous: Total power reduction
Lin Zhong
2003-10-11