Ps. Does SVP recalculate the limit for what it considers to be "detailed and bright" areas when only dark scenes are displayed (i.e. as a fraction of the maximum brightness of a block in the current scene) for a number of consecutive frames, or does it continue to consider the entire frame's blocks to be 'unimportant' or 'less visible'?

I only watch anime and from my observation, the default setting is better than setting described in this wiki

Frames Interpolation Mode: Adaptive (or "2m", depending on the severity of motion artifacts)
In this case, maybe you can tweak smooth.scene.limit options in override.js,

Target Frame Rate: Source multiplied by 3 (use "To screen refresh rate" for 60Hz monitors)
This certainly less artifact since you have exact integer multiple than the display with 60hz refresh rate

Nintendo Maniac 64

I think the most recommended is, the setting which give the smoothest video while keeping cpu usage downs as much as possible
for now it's the trade-off between the comfortable smoothness and cpu usage (I emphasized the comfortable smoothness )
higher setting which consumes cpu highly is not necessarily better

==========================================
Ok, back to topic,
Sometimes cpu is spiked when there's sudden change scene, it spikes really high but not very high to make the noticeable slowdown,
it's like normally ~40% to suddenly ~60-80%.
Since the TS, Mystery, keeps his setting about 80-100%, isn't just normal that he notices the slowdown?

Mystery
I tweaked my settings to keep the CPU between 80-100%

not a great idea

Just for reference sake, what would you recommend?

yep, I read. Thanks for detailed answer. This full post can be copy-pasted to wiki 

TL;DR: skip to the end for my settings.

Well first I'd like to give the reasoning behind my settings. Anime content poses some unique challenges; specifically, only panning scenes or computer-generated content actually have a solid mathematical relationship (where motion is concerned) between past and future frames. The rest of the content, however, is normally hand-drawn at some fraction of the video's frame rate (i.e. only drawing 6 or 12 changed frames per every 24 frames of motion).
Since SVP uses a block-matching algorithm, (and artists re-drawing a character multiple times to create the illusion of motion rarely draw an exact 3D projection every time) very few blocks will match up close to perfectly (unlike the case with 'normal' video). Furthermore, the stop-start nature of the motion (caused by drawing fewer moving frames than the frame rate requires) also makes it hard for SVP to calculate the true motion since as far as it is concerned, that is how the motion is supposed to be. SVP has no way of knowing when two equal frames are the result of a production being too cheap to even draw 24fps motion, rather than simply signaling that the moving object has come to a halt. This leaves us with greatly increased artifacts (mostly in the form of solid lines becoming 'wavey' as shown in The SVP Wiki) or with the motion just not being any smoother, even for slowly moving or rotating objects.
Yes, in theory one can write some sort of algorithm to check for and remove the superfluous frames (while leaving legitimate motion-free portions alone) and then use SVP-framedoubling to interpolate the resulting variable frame rate video, but that is way beyond the scope of dlr5668's question...

As for the SVP settings, we have one of two choices:
1) We can only smooth reliable, global panning motion (produces fewer artifacts but does almost nothing to moving objects or characters)
2) Live with the increased artifacts and try to tweak SVP's settings as best we can for each specific anime's drawing style.

I prefer option 2 (but you can easily adapt my settings for 1 by using a very large 'motion vectors grid' setting, disabling 'decrease grid step' and setting motion vectors precision to one or two pixels).
These are my settings for MOST anime content:

Profile Settings:
Processing Threads: 24 (use less for more stability but less performance or for quad-core CPUs)
GPU-acceleration: Pick my second GPU (One not used by MadVR)
Frames Interpolation Mode: Adaptive (or "2m", depending on the severity of motion artifacts)
SVP shader: 2. Sharp
Target Frame Rate: Source multiplied by 3 (use "To screen refresh rate" for 60Hz monitors)
Motion vectors grid: 14px. Average 1 (for 720p: 8px. Small 9)
Decrease grid step: By two with global refinement
Search radius: Large
Motion vectors precision: One pixel (for 720p: Half pixel)
Wide search: Strongest
Artifacts masking: Weak
Processing of scene changes: Repeat frame
Decrease frame size: Disabled (this setting makes very little sense to me, I suppose it is for compatibility with very slow PCs?)

The following settings are best left to the profile defaults unless you have read through The svpflow dcumentation and The SVP MVTools2 dcumentation and understand what they do (can decrease both quality and performance).

Hidden settings: (SVPMgr.ini, override.js, etc. Very content dependent)
SetMemoryMax(768)
analyse.main.search.distance = -16 or -32
analyse.main.search.satd    = true (better quality but requires much more performance)
analyse.main.search.coarse.width = 1080 (set this >= to your video width)
analyse.main.search.coarse.distance    = -16 or -32
analyse.main.search.coarse.trymany    = (depends, can improve or worsen performance)
analyse.main.search.coarse.bad.sad    = 1000 (750~1250)
analyse.main.search.coarse.bad.range = -32 or -64
analyse.main.penalty.lsad    = 4000 (4000~8000)
analyse.main.penalty.pnew = 40;
analyse.main.penalty.pglobal = 40
analyse.main.penalty.pzero = 75;
analyse.refine[0].thsad = 300;
analyse.refine[0].penalty.lsad = 6000;

not a great idea

With 3.1.6, I was still using updated libraries so upgrading to the latest package didn't give improvements in itself.

That actually explains it. If I change the libraries to the new ones then I also get identical performance between 3.1.6 and 3.1.7, which makes sense since I believe all performance inprovements should come from the dll libraries (the SVP manager should not use any substantial amount of resources). Also, if there were any architecture-specific improvements, I believe that Mystery's Ivy-Bridge based processor would also have benefitted from it as Ivy-Bridge also supports the AVX2 instruction set.

I also downloaded and tested that Youtube video after I disabled 4 CPU cores in the BIOS and backed off the "threads" setting to better aproximate more "mainstream" systems, and I still seem to get around a 10% performance improvement (lower total rendering time) with the new libraries.

I think that the reduced performance for some scenes can best be explained by SVP having to do more work to find suitable motion vectors when there are many different moving objects having large motion displacements, as compared to still or slowly moving frames.
If I change my SVP settings to those more suited to interpolating Anime content, SVP's performance and quality on that video also improves quite markedly.

Ps. You can tweak this scene-dependant adaptive behaviour by changeing the "badSAD" values and altering how far and with what algorithm SVP should search for replacement motion vectors; however the default values seem to be optimal for most mainstream 4-core machines and videos.

And yes, I have "only" a quad-core i7 3630QM, which is still near the top of this list after 2 years with a score of 7638
https://www.cpubenchmark.net/laptop.html

Intel 5960X @ 4.6GHz (I also have it in the "System Specs" area of my forum profile), why do you ask?

Uhhh... holy bananas, you literally have the fastest consumer CPU model in existence combined with a considerable overclock.  I'd like to point out that, if I recall correctly, Mystery "only" has a Nehalem i7 (though I obviously could be wrong).

Remember, SVP is very multi-thread friendly, so it's actually one of the cases where "moar cores!" actually gives quite the performance benefit, and an 8-core/16-thread Haswell E CPU is the current king of the crop when it comes to that kind of thing in the consumer marketplace.