High Ram usage

What is your nginx config?

What makes you think that's high ram usage? What is the output of "free -m"? Are you swapping? And not just if you have some swap memory used, but are you swapping regularly, i.e. swap io. You WANT to "use" all your memory. You just want to use it all in the best way.

Maybe you want that memory as mysql cache or maybe as php child processes, or maybe not tied up in applications at all but rather free for the OS to use as buffer and file cache. Your ps output looks like a pretty good usage to me if you're using 5 or 6 active child processes most of the time. and not emptying your mysql qcache too often.

The number of fpm children should be the number of children that you need. As a starting point you want generally at least as many as CPUs as you have, so maybe 1 or 2 or 4 depending on your computer, plus 2 or 3 more for when a child is waiting on something like a database backend. But that is only a general rule. If your child processes are blocking for long periods of time for something, like your php script is retrieving something offsite, you might want more. With just Drupal accessing a database, you don't need that many extra.

You could install collectd and add a bit of configuration and then have a handy graph of how many child processes are active or idle over a period of time. If you find that the most you're ever using at once is 2 or 3 child processes at a time, and that 7-8 are idle 99% of the time, then maybe reduce to 4 or 5 minimum and 20 maximum. Maximum of course is dictated by available RAM like the formula that perusio stated. But if you're only ever using 2 or 3 at a time, then set the minimum to 5 and free up the rest of the ram for os file cache use. Or increase your qcache by 16MB if you need it. If you get a spike then fpm will spawn those new processes up to the maximum, but the other 99% of the time you'll be serving your static files a bit faster because more of them fit in the file OS cache. Or if you're constantly using 13 active children, then raise your minimum up to 16 if that's what you need.

Stat monitoring will also tell you your swap IO. And can also tell you your qcache efficiency. Or you can check qcache evictions with mysqltuner.pl and tuningprimer.sh. Maybe you want to kill off two of those php-fpm children and use that 25MB for a memcached instance for your variables table and views and ctools cache tables.

If this machine was devoted exclusively to php then you could just set children to use 90% of the memory and be done. But if you're doing mysql and static files from the same machine, you have to allocate between them.

The ondemand option is really for a different use case. If you were hosting 40 different clients and wanted each to have their own pool for security reasons, and their traffic was low enough, then you wouldn't want to spend your RAM by giving each even 1 minimum. You might want to give each 0 minimum. That is the use case for ondemand and it's just being implemented. Until ondemand, fpm hasn't been practical for shared hosts to replace suexec.

For a single fpm pool, it's hard to justify using ondemand. You'd have to have an unusual use case where you wouldn't want to spare the memory for at least 1 or 2 children available at all times.