Quote:
Originally Posted by huhn
you don't have to test HEVC and and HDR doesn't help here anyway h264 should be good enough.
|
Unfortunately the Radeon 7870 doesn't even support 4K H264, so useless to test. The real limitations of copyback decoding only start to become a problem with 10-bit 4K, because it takes up 8 times the bandwidth of 8-bit FHD. With lower resolutions, using copyback or native doesn't have an impact on which madVR settings I am able to use. With 4K 10-bit it does.
So I've done some benchmarks on my HTPC...
Notes:
- 'GPU' and 'video' numbers are the frequency reported / usage % (e.g. 1290 MHz at 50% load = 'GPU 645')
- CPU and GPU usage counters of DXVA Checker are completely wrong, I don't know how it computes them. Maybe the GPU counter reports only shader usage so it could be right but not useful, but the CPU usage is always wrong. I used HWMonitor which gives the same values as other monitoring tools.
CPU @ 3500 MHz (FSB 333), RAM verified dual channel
1. 4K HEVC 10-bit. Best of 5 passes with DXVA Checker decode/playback, best of 3 runs with madVR. Playback at 1920x1080. madVR settings: scale chroma separately; no compromise on HDR quality; SSIM2D downscale; clip pre-measured for HDR; no black bars detection.
RAM @ 666:
Decode: 63,0 fps, CPU 65, GPU 1006, Bus 24
Playback: 34,9 fps, CPU 81, GPU 731, Bus 21
madVR: 439 dropped frames, avg 50,16 ms, max 78,17 ms, GPU 1772, CPU 95
RAM @ 800:
Decode: 66,8 fps, CPU 60, GPU 1017, Bus 25
Playback: 34,5 fps, CPU 84, GPU 656, Bus 21
madVR: 315 dropped frames, avg 45,68 ms, max 63,78 ms, GPU 1772, CPU 90
For reference, with Native:
Decode: 178,5 fps, CPU 46, GPU 1642, video 1467
Playback: 177,0 fps, CPU 54, GPU 1785, video 1467
madVR: 0 dropped frames, avg 34,38 ms, max 38,46 ms, GPU 1613, CPU 68
Difference of dropped frames and max render times under madVR just with 20% faster RAM is massive.
With DXVA Checker, CPU is not fully loaded with decode and only 6% faster decode with 20% faster RAM. Software/platform inefficiency?
2. Same test but with madVR 'light' settings: compromise on HDR quality checked; Bicubic downscaling instead of SSIM2D
copyback: avg 16,5 ms, max 25,04 ms, GPU 1136, CPU 78
native: avg 14,93 ms, max 17,78 ms, GPU 592, CPU 25
max render time is 40% better while GPU is two times less loaded, CPU three times less loaded. Massive performance impact.
I understand why CPU would be loaded if it has to wait for frames to be read/written from/to system RAM, but why more GPU load? Can't the GPU render a frame it has received from the renderer while the next queued frame from the decoder is transfered over the PCIe bus and back?
A single 4K P010 frame is 25 MB, at PCIe 2 x16 it should take 3,125 ms, 6,25 ms round-trip just for the time over the bus. If the rendering has stalls it could explain the difference of a few ms between copyback & native even with very high end GPUs.
3. A lighter test comparing Jellyfish clip at 1080p HEVC, same bitrate, in 8-bit and 10-bit:
decode 8-bit: 266,5 fps, CPU 54, GPU 1797, video 1430, bus 11
decode 10-bit: 210,7 fps, CPU 60, GPU 1797, video 996, bus 22
playback 8-bit: 240,2 fps, CPU 75, GPU 1797, video 1141, bus 15
playback 10-bit: 181,9 fps, CPU 75, GPU 1743, video 852, bus 20
We see 10-bit decode takes up exactly two times the bus bandwidth as 8-bit, as expected.
The 10-bit decode performance doesn't scale to 4x the speed of the 4K clip (would be 267 fps).
for reference, 10-bit native: 299,2 fps, CPU 18, GPU 1613, video 1415, bus 2
4. Just out of curiosity I underclocked the CPU to 2100 MHz (FSB 200), to be able to test more different RAM speeds:
(Jellyfish 10-bit DXVA Checker decode):
RAM @ 400: 131,6 fps (native 268,4), CPU 76, GPU 1589, video 989, bus 13
RAM @ 533: 139,6 fps (native 275,7), CPU 70, GPU 1642, video 909, bus 14
RAM @ 666: 148,8 fps (native 281,1), CPU 67, GPU 1642, video 798, bus 15
RAM @ 800: 146,0 fps (native 281,8), CPU 68, GPU 1428, video 766, bus 15
for reference, CPU @ 3500 & RAM @ 800: 210,7 fps, CPU 60, GPU 1797, video 996, bus 22
With same RAM speed but 66% faster CPU, 40-45% more fps.
With same (slow) CPU speed but 66% faster RAM, 13% more fps.