Of course, backing off on the block overlapping, for instance, would make running both the CPU and IGP, at the same time, much more feasible. Then again, maybe he already has a 5GHz 6700k with some overclocked IGP in there to boot.

I would also recommend running MadVR on a second discrete GPU

SVP is light enough in its GPU usage that you can have SVP running on even Intel integrated graphics while MadVR runs on a discrete GPU.

Hey, you seem really well versed with this, what can you recommend to get the smoothest video? Cause for most videos I max everything out and have some cpu power to spare.

Thank you for your kind words. Nintendo Maniac 64 is right, if there are a lot of thin lines (or, rather, many high-amplitude and high-frequency spacial components), then complicated will 'smooth' them over if it cannot find a smooth progression from one frame to the next. Please see here for the general difference between 'smooth' and 'sharp' SVP settings (when it comes to difficult moving objects that are not very thin (the thin lines get blurred into the wavy image for the 'smooth' settings).
What I would recommend depends a lot on your system specs, but this is what I'd recommend for the smoothest real-time interpolation of standard high-quality (>5Mbps x264 encoded @ High10 & slower preset or more) 1080p23.976 series to 1080p60:

Use the same GUI settings that you posted, except for setting the 'SVP Shader' to "23. Complicated" and setting 'Artifacts masking' to "Weakest".
Then make use of the following values for override.js (a settings file in the main SVP install directory that can override the, normally hidden, settings that the developers deemed should not be altered (which is normally the case as these can do much more harm than good). Also, if you have the time, you may want to look over the documentation of these Advanced SVPFlow Options and the MVTools2 parameters, to better understand what the overrides do.

Remember to first create a backup of your original override.js file (so that you can easily restore your SVP configuration to the way that it was). Then try out my override.js recommendations by simply overwriting the default file in the SVP directory with the one that I attached here.

EDIT: I would also recommend running MadVR on a second discrete GPU (if possible) and to use a sharp bicubic scaler with the anti-ringing filter, as well as making use of the available sharpening post-processing options, to sharpen up the blurry SVP output (remember to check that this doesn't expose a lot more artifacts and doesn't make existing artifacts much more visible).

After I upgraded to my current 5960X-based system, I received a very noticable increase in image quality (still not what I would call 'good', but I don't think it is reasonable to expect any interpolation software to be able to generate a good picture from such a heavily aliased source as a 24fps recording).
After I overclocked said system to 4.7GHz, the quality did not increase, merely because SVP was not developed with such systems in mind.
However, redefining the values in the override.js file, did in fact significantly improve the quality of SVP's interpolated images, while also running at around 80~90% CPU load, on average. Simply put, my 5960X is not up to the task of providing enough performance to allow for a proper interpolation to be done (at 1080P, I don't even try 4k).

So while SVP3 (and certainly not SVP4) will not be able to make use of increased CPU or GPU power at default, editing the configuration files allows SVP to make full use of any CPU you can give it, with the accompanying massively improved image quality it produces.

I am in the process of thoroughly testing and documenting the effect that these 'hidden' settings have on the quality and CPU load of SVP's interpolated output (similarly to what has already been done to compare different image upscaling algorithms on other forums).
It is slow and tedious work (I am using the lossless SVT_MultiFormat sources which are 48MB per frame (more than 20GB for a 10s native 50fps video) and SVP is almost always running at less than 1fps), but I hope to be able to demonstrate, to the developers, the significant improvements in quality (with proper mathematical similarity metrics to back it up) that allowing these heavier settings has.

Hey, you seem really well versed with this, what can you recommend to get the smoothest video? Cause for most videos I max everything out and have some cpu power to spare. Here's what I'm currently rolling with.

http://cdn.overclock.net/2/26/260e926a_ … cried.jpeg

Not bad 

Test summary
-----------------------
  Date: 2015-08-29T17:04:45
  CPU:  Intel Core i7-5820K @3305 MHz [12 threads]
  GPU:  NVIDIA GeForce GTX 970 [ver.355.60]
  Mode: FHD + GPU [25 threads]

Overall scores
-----------------------
  Synthetic CPU:                  MC3609
  Synthetic GPU:                  MG6312
  Real-life:                      FG4567

Details: synthetic
-----------------------
  CPU: compose (single-threaded): 696
  CPU: compose (multi-threaded):  4214
  CPU: search (single-threaded):  518
  CPU: search (multi-threaded):   3156
  GPU: system -> GPU transfer:    1881
  GPU: GPU -> system transfer:    2067
  GPU: calculations:              4483
  GPU: total score:               10875

Details: real-life /FHD
-----------------------
  decode video:                   21.84x (524.2 fps)
  48 fps - vectors search:        2.42x (116.3 fps)
  60 fps - frame composition:     6.61x (396.8 fps)
  48 fps - [SVP] fastest:         9.44x (453.2 fps)
  48 fps - [SVP] simple 1:        7.50x (360.1 fps)
  60 fps - [SVP] good:            4.09x (245.1 fps)
  60 fps - [SVP] high:            3.24x (194.7 fps)
  60 fps - [SVP] highest:         1.45x (87.1 fps)
  72 fps - [SVP] simple 2:        6.66x (479.7 fps)

That's close to the other 4790k@4.8 in the thread. What's your single threaded score?

Sadly, no-one has posted a 6700k bench yet.

NOTE: SVP Mark doesn't read that my CPU is running @ 4.7Ghz, it just displays the stock non-turbo clock of the 4790K.

Test summary
-----------------------
  Date: 2015-08-27T13:01:40
  CPU:  Intel Core i7-4790K @4000 MHz [8 threads]
  GPU:  AMD/ATI Radeon R9 200 / HD 7900 [ver.1800.8]
  Mode: FHD + GPU [17 threads]

Overall scores
-----------------------
  Synthetic CPU:                  MC3109
  Synthetic GPU:                  MG5462
  Real-life:                      FG3935

Details: synthetic
-----------------------
  CPU: compose (single-threaded): 856
  CPU: compose (multi-threaded):  3611
  CPU: search (single-threaded):  657
  CPU: search (multi-threaded):   2731
  GPU: system -> GPU transfer:    1143
  GPU: GPU -> system transfer:    2044
  GPU: calculations:              3807
  GPU: total score:               9075

Details: real-life /FHD
-----------------------
  decode video:                   22.56x (541.3 fps)
  48 fps - vectors search:        2.00x (95.8 fps)
  60 fps - frame composition:     5.37x (322.3 fps)
  48 fps - [SVP] fastest:         9.44x (452.9 fps)
  48 fps - [SVP] simple 1:        6.26x (300.4 fps)
  60 fps - [SVP] good:            3.39x (203.2 fps)
  60 fps - [SVP] high:            2.61x (156.7 fps)
  60 fps - [SVP] highest:         1.19x (71.7 fps)
  72 fps - [SVP] simple 2:        5.95x (428.7 fps)

brucethemoose wrote:

Necro

Someone with a Skylake rig just ran SVPMark, looks like it's ~20% faster than Haswell clock-for-clock.

http://www.overclock.net/t/1568663/inte … t_24336762

4790K @ 4.7Ghz

Synthetic CPU: 3106
Synthetic GPU: 5470
Real-Life: 3933

I have Chrome and Word running in the background too while running the benchmark.

That's close to the other 4790k@4.8 in the thread. What's your single threaded score?

Sadly, no-one has posted a 6700k bench yet.

Necro

Someone with a Skylake rig just ran SVPMark, looks like it's ~20% faster than Haswell clock-for-clock.

http://www.overclock.net/t/1568663/inte … t_24336762

4790K @ 4.7Ghz

Synthetic CPU: 3106
Synthetic GPU: 5470
Real-Life: 3933

I have Chrome and Word running in the background too while running the benchmark.

Someone with a Skylake rig just ran SVPMark, looks like it's ~20% faster than Haswell clock-for-clock.

http://www.overclock.net/t/1568663/inte … t_24336762

Bong34
You said you were "getting a new setup"; If that entails also buying a new GPU, I would definitely recommend waiting for the new 14nm GPUs to be released. At least to see if they got Samsung (or maybe even GloFo...) to do a special 'High-Power' node for their products (all current 14nm processes, excluding Intel, are either 'very-low-performance' or 'low-performance' for small, mobile-device SOCs). They reason I suggest waiting (which I wouldn't normally do, since there is always 'something new' on the horizon), is because a GPU built on a dedicated 14nm HP node will be able to easily double the performance of even the fastest Titan X card, in the same 250W package. If, however, they have to make do with a LPE or LPP node, then we can expect (at best) a 50% performance increase in the same 250W package.

Either way, I doubt it will make any difference to SVP's performance (since SVP doesn't actually do all that much on the GPU), but thought it relevant since you mentioned gaming performance as another priority.

I'm not going to say I understand all the background work that SVP does, because I don't. But what I was telling the OP wasn't a negative statement about SVP (sorry if anyone misunderstood it that way), basically what I was trying to say from the start was that "don't expect that just because you are going to buy the 6700K you can "max-out" all the settings".

I didn't want to go into the detail of what every setting does or how it works differently for different types of media and the such; but I guess I should have written something along the lines of "don't expect to max out all the settings for SVP because that certainly isn't what you should be aiming to do; you're going to need different settings for different types of media and different resolutions. You're going to need to do things differently if you upscale or downscale. And lastly different people will like things differently; some use artifact-masking but lose a bit of fluidity and that's understandable for them (because they prefer it that way), others however hate anything that ruins perfect fluidity, so they prefer to see artifacts... At the end of the day, BUY the CPU that is within your budget, and work your settings around your hardware". The great thing with SVP is that it makes you always want "something better", but for now there's nothing that is "perfect" when it comes to motion-interpolation, and that is why I recommend you just get what is within your budget.

PS: can't wait for SVP 4

Thanks to all the developers and beta testers for all their hard work.

Sure, a 6700k is probably "enough" for most people's anime interpolation preferences, but that is only because it rarely contains any mathematically coherent motion to actually interpolate.

For all other use cases, no CPU & GPU combination currently exists that can deliver anything close to 'good enough' performance. Then again, some people like their interpolated frames to still contain aliased motion (aka. the whole 'film look' vs. 'soap opera effect' debacle) and SVP4 makes achieving that with very few other artifacts quite easy with very modest computational requirements.

This is actually a pretty complicated issue, but the simple answer is a combination of:

1) SVP not being very well threaded (or more accurately, the MVTools library running on avisynth has a badly hacked form of threading support (optimal number of threads for maximum SVP performance on a 4-core 4790k = 24) that does not scale very far (hard 32bit memory limits my 8-core 5960x to threads = 26)) and has a pretty high threading overhead) which leads to a fairly bad case of Amdahl's law

2) Skylake has a ~10% IPC advantage over haswell (properly vectorised AVX2 code sees a much larger increase (almost 50% over Haswell's AVX2 code execution) but no one writes code like that these days), and the 6700k overclocks (on average) another 5-10% further. Together with the faster caches and memory (assuming decent DDR4 memory since you have to buy new RAM anyway), I would put the 6700k's net per-core advantage at around 20%, which is very close to the real-world benefit of 50% more cores.
(Having an 8-core 5960x, I can testify that doubling the cores from the 'standard' i7 4700-series makes absolutely no difference in 99% of applications and games; if I had the choice between a 32-core @ 4Ghz CPU and one with the same architecture, but only 4 cores at 6Ghz I would jump on the latter in an instant).

However, if you already have a 5820k, getting a 6700k will not improve your performance in SVP, they should perform very similarly if both have been given a solid overclock. It probably will improve your performance in most other applications a bit, but don't expect a constant 20% improvement in all single-threaded apps either. That is the real reason for recommending the 6700k over the 5820k, not because the 6700k will be much faster in SVP.
Also, using the IGP in any way would be a very bad idea. Combining an otherwise equal CPU and GPU onto one die just serves to dramatically reduce the performance of both. If intel could replace all that wasted die space with some more cache and expanded speculative execution resources, then we would probably have an instant 50~60% jump in IPC, but they'd never do that... more useless cores markets way better.

A simple statement that keeps floating around that does not really explain the problem that well. 'maxing' the settings in SVP's GUI by simply pushing all the blue bars as far right as they will go does nothing but make most people's PCs drop frames (and display severely artifacted ones in many other cases).

As I (and indeed all the developers aswell) have said before, this is why there are so many reports of SVP4 looking so much better while requiring less CPU resources to do so (apart from most people simply preferring the 'sharp' look of temporal aliasing). If people don't educate themselves about the algorithm being implemented by actually reading the source code, seeing where those variables go and understanding how that mathematically alters the accuracy of the extracted motion vectors in different situations, then how are they supposed to set the 'best' values for their own system?

Well, I can give you an idea if you'd like. My system specs is in my profile (a 5960x @ 4.7GHz, SLI Titan X'es @ 1.3GHz and 128GB RAM) and my current setup is to encode my video and watch it later since I get about 0.8~1fps. Getting such a high speed requires splitting my video file into halves and encoding each with 4 cores and one of the Titan x'es in a separate virtualdub instance (this explicit parallelism almost doubles the processing framerate).
I believe these SVP settings that I am using for high quality Blu-Ray playback is about as far as the current SVP libraries can be pushed with regards to maximum SSIM (what I optimise for; aka. the subjective part) for good, noise and artifact free sources in a 32-bit process.
So, to answer your question, to 'max out' SSIM with the current SVP libraries in real-time (assuming 24->60fps) would require a system capable of around two orders of magnitude more processing power than what I have, without resorting to more cores or more GPUs (since the split file hack can't be used in real-time), which is equivalent to around four times! the performance increase we got (while clockspeeds were still being increased between 1985 and 2005).

Theoretically, this can be done within the next decade, but would require a complete re-write of the MVTools library with explicit multi-threading and advanced AVX-512 inline ASM optimisations for all data processing functions to efficiently offload the data to a CUDA (OpenCL could also work I guess) implementation of the actual MV-searching and determining functions which (if you have taken a look at the MVTools library) is a task that would probably cost a few million dollars in software development (although with these Russian coders you never know... they have this astonishing way of just getting stuff done with much less resources than most people thought would be required).