I was just evaluating 9.1 (Webpack). I'm using small S3 and I have to wonder if Xilinx is just doing optimization(s) for high end parts. For my design, running in XC3S50, performance (both area and speed) degraded. I'm using a "make" driven flow and the settings (command line args and files) are the same for both 9.1/8.2 runs (actually I use a flow that I put togheter in 7.1 days or so).
Data from *syr:
8.2.03i - xst I.34: Number of Slices: 713 out of 768 92% Number of Slice Flip Flops: 428 out of 1536 27% Number of 4 input LUTs: 1306 out of 1536 85% Number used as logic: 1258 Number used as RAMs: 48
9.1i - xst J.30: Number of Slices: 696 out of 768 90% Number of Slice Flip Flops: 428 out of 1536 27% Number of 4 input LUTs: 1268 out of 1536 82% Number used as logic: 1220 Number used as RAMs: 48
Data from *par:
8.2.03i par I.34 Number of RAMB16s 4 out of 4 100% Number of Slices 699 out of 768 91% Number of SLICEMs 24 out of 384 6%
9.1i par J.30 Number of RAMB16s 4 out of 4 100% Number of Slices 753 out of 768 98% Number of SLICEMs 24 out of 384 6%
Data from *twr:
8.2.03i: Timing errors: 0 Score: 0
Constraints cover 100721 paths, 0 nets, and 5844 connections
Design statistics: Minimum period: 19.976ns (Maximum frequency: 50.060MHz)
9.1i: Timing errors: 1 Score: 325
Constraints cover 101771 paths, 0 nets, and 5659 connections
Design statistics: Minimum period: 20.650ns (Maximum frequency: 48.426MHz)
Run time for my flow:
8:2 420.14 real 343.23 user 6.20 sys
9.1 420.73 real 343.84 user 6.07 sys
I did not spend a lot of time in reading the new documentation for xst/ map/par to look for new switches/setting. If you are aware of eny please comment.
It seems the bit stream generated with 9.1 is running just fine in my applications (one data point, no temp/voltage/corners).