Report your hashrates. I have only tested it on the gtx960M (768 cuda cores)
980 TI -i 31 1641mhz I'm getting about 3285 which is +42ish vs Mod#7. -i 31.9 bumps it to 3292. edit: these are numbers reported by the miner.
Further testing....
980TI
Yiimp reports roughly the same hash rate (must run at least diff 2 which causes reported hash rate to vary depending on luck solving the job).
Suprnova poolside (diff 1) reports a MAX of around 2.65GH with an avg of roughly 2.45 at -i 31 (and I did let it run for over an hour and kept an eye on it the entire time) using Mod9 whereas Mod7 poolside shows 2.7-3.5 averaging around 3.1 (data taken today and from DAYS of monitoring). Mod9 -i 31.9 was averaging about 2.3GH (see offtopic discussion at bottom of post for more info).
750TI
-i 24.5 1339mhz (primary display) Mod7 614MH, Mod9 620 (Mod7 -i 24.2 reports about 602 with MUCH less screen lag)
I am getting better poolside reports with Mod9 580-720 vs 470-630 Mod7. I am also significantly more likely to solve a job with Mod9. This is just an eyeball estimate but solve rate seems to have gone from 60-75% to around 85% on my 750TI. It is uncommon on Mod9 to have a string of 3 or more jobs fail to find a solution whereas Mod7 often had runs of 5-7+. The 980 TI also reports a higher solve rate on Mod9 but it's not as profound.
Conclusion:
On my equipment, Mod9 appears to be a gain on compute 5.0 but something is broken on 5.2 cards, the poolside hash as reported on Suprnova is significantly lower.
Also, the --show-diff flag is broken in Mod9. It always shows (diff 0.000) for every solution. This flag was also broken in the 1.5% miner fee version from a few days ago.
Offtopic: For the 980TI I actually get better throughput using -i 31 than -i 31.9 (as reported poolside). The typical job time is about .8 seconds vs 1.1-1.2 for the higher intensity. These numbers I can see when I have a run of jobs for which it did not find a solution (don't use -q). I am a bit more likely to find a solution meeting the difficulty requirement of the job at the higher intensity but the slower processing time lowers throughput. This made me scratch my head more than a little bit...why the higher solve rate at the higher intensity? But I've watched the numbers stream past for hours. Tried -i 31.5 and it yields same results as 31.9 for throughput.