Here is a partial summary of GPUs for madVR. madVR includes several chroma / image up(down)scaling algorithms and GPU requirement depends on each algorithm. For the sake of convenience I introduced the notion of "levels", each of which consists of a chroma upscaling algorithm, an image upscaling algorithm, and an image downscale algorithm.
- Level 1: The most primitive algorithms (apart from Nearest neighboring).
- DXVA2: When image is scaled, the driver's chroma / image scaling algorithms are used; when image is not scaled, Bicubic 75 chroma upscaling algorithm is used.
- Level 2: MadVR's current default algorithms, that are supposed to work even with a weak GPU.
- Level 3: Anti-ringing flter is added to each algorithm of Level 2.
- Level 5: The best algorithms available today.
I think the above classification is pretty reasonable. There is a huge gap between Level 3 and Level 5. So I added an intermediate level, Level 4. But which is to be weaken from Level 5, chroma or image? Image scaling has much more visual impact than chroma upscaling. So I kept image scaling of Level 5 and lowered chroma upscaling algorithm a little bit (that will reduce the GPU requirement greatly when the source resolution is large).Test system
- Core i5-4670
- ASRock Z87 Extreme4
- DDR3-1600 2 x 4GB
- 1 x SSD (OS), 1 x HDD (data)
- Windows 8 Professional, AMD graphics driver 13.4
- MCP-HC 1.6.8, LAV Filters 0.57.0 (DXVA2 copy-back), madVR v0.86.4Test clips
There are two categories of video contents, film and video, in terms of frame rate. Film
is shot by a movie camera @24fps (or 25fps or 30fps). Anime
is included in this category. Video
is shot by a video camera @60fps (or 50fps or 30fps), up to deinterlacing. Roughly speaking video is 2.5 times harder
than film (or possibly more because of deinterlacing) to play back (precisely speaking, GPU load at "video" playback is roughly 2.5 times more than "film" playback) because madVR has to process each frame within a time interval of
- 41.67ms = 1/24fps for film
- 16.67ms = 1/60fps for video
(24 x 2.5 = 60.) The clips used:
- F1: SD (720x480i60): DVD and NTSC movies.
- F2: HD (1280x720p24): ATSC movies and movies found on the Internet
- F3: FHD (1920x1080p24): BD and ATSC movies
- F4: 4K UHD (4096x2304p24)
Video (i = interlaced, p = progressive)
- V1: SD i (720x480i60): DVD and NTSC videos
- V2: HD p (1280x720p60): Camcorder and ATSC videos
- V3: (3/4)FHD i (1440x1080i60): HDTV videos in several countries (if you live in US, you can ignore it.)
- V4: FHD i (1920x1080i60): Camcorder and ATSC videos
- V5: FHD p (1920x1080p60): Camcorder videos
- V6: 4K UHD p (4096x2160p60)
F4 and V6 were used only for the 4K UHD target because downscaling 4K to HD / FHD by Catmull-Rom is too taxing (for example, even HD 7970 can't play V6 in a FHD display at Level 5 without dropped frames). If you play 4K files in a HD/FHD display, you'd better use DXVA2+EVR (e.g. WMP).AMD dGPU for madVR
For a FHD (1920x1080) display:
If you are a film only person, HD 6570 is a good choice. Otherwise HD 7770 is minimum for Level 5.
For a HD (1366x768) (or 1280x720) display:
For a 4K UHD display:
If you are a film only person, HD 7770 is a good choice. Otherwise HD 7970 is the only choice for Level 5.
*: Not good for (3/4)FHD i.Edited by renethx - 6/29/13 at 7:26am