Performance

UltraGrid Performance

This site contains the results of performance tests of UltraGrid.

End-to-End Latency

The following table compares performance of individual cards. The results are measured as end-to-end frame-delay, i.e. the number of frames sent before the reciever outputs the original frame.

Linux

Setup:

• Testing machine: hd2
• Reference machine: hd4 with BlackMagic Decklink HD Extreme
• For the test, we used 1080i HD video at 29.97Fps.

Card	send	recieve
DVS Centaurus II	5	3
BlackMagic DeckLink HD Extreme	3.75	3.75
BlackMagic DeckLink 4K Extreme	3.5	3
BlackMagic Decklink Quad	4.5	4
BlackMagic Decklink Intensity PRO	4.5	4
BlackMagic Decklink Intensity	4.5	4.5
Deltacast 3G	4.5	3
OpenGL	-	2.5

macOS

Setup:

• Testing machine: hd7
• Reference machine: hd4 with BlackMagic Decklink HD Extreme
• For the test, we used 1080i HD video at 29.97Fps.

Card	send	recieve
AJA Kona 3G	4	3.5
DeckLink HD Pro (Quicktime)	4.5	5.5
DeckLink HD Pro (native API)	4.5	4
OpenGL (with VSync)	-	2.25
OpenGL (without VSync)	-	1.75

Windows

Setup:

• Testing machine: hd7
• Reference machine: hd4 with BlackMagic Decklink HD Extreme
• For the test, we used 1080i HD video at 29.97Fps.

Card	send	recieve
BlackMagic DeckLink HD Extreme	4.5	3.5
BlackMagic Decklink Quad	4.5	4
BlackMagic Decklink Intensity PRO	4.5	3.5
BlackMagic Decklink Intensity	4.5	4
Deltacast 3G	4	3

Compression

Performance

Here you can see the performance of individual compression modules (encoding). For the measurement purposes, we used machine hd2 running Ubuntu as receiver and hd7 as sender. We used Decklink HD Extreme as grabbing cards on both sides. We used a 4k video with increasing framerate to pin down the biggest achievable that still offers fluent video experience.

Encoder (setting)	res	FPS	Content	HW	Ver
cineform	2160p	65	NZ2500	i7-4960X	1.6
gpujpeg (90)	2160p	157	NZ2500	5960X+9	1.3
libx264	2160p	43	NZ	i7-4960X
libvpx	2160p	28	NZ	i7-4960X
mjpeg	2160p	65	NZ	i7-4960X
cineform - R10k	2160p	42	NZ2500	i7-4960X	1.6
cineform - R12L	2160p	25	NZ2500	i7-4960X	1.6
gpujpeg (90)	2160p	150	NZ2500	kypo	1.3
gpujpeg (90)	2160p	137	NZ2500	hd12	1.3
gpujpeg (90)	2160p	201	NZ2500	5960X+B	1.3
gpujpeg (90:8)	2160p	207	NZ2500	5960X+B	1.3
gpujpeg (90:16)	2160p	200	NZ2500	5960X+B	1.3
gpujpeg (90)	2160p	178+146	NZ2500	5960X+B9	1.3
gpujpeg (90)	2160p	92	NZ2500	bunny	1.3
gpujpeg (90)	2160p	96+96	NZ2500	bunnyX2	1.3
gpujpeg (90)	2160p	130	NZ2500	hdd1	1.3
libx264	2160p	32	NZ	i7-980X
libx264	1080p	4	NZ	VIA Nano	1.5d
mjpeg	1080p	10	NZ	VIA Nano	1.5d

Legend:

• kypo
  i7-4770S, NV GTX 980, 4x8G DDR3@1600 (kypowall0)
• hd12
  i7-4960X, 32 GB 1866 MHz DDR3, NV GTX 960
• 5960X+9
  i7-5960X, DDR4@2166, NV GTX 960
• 5960X+B
  i7-5960X, DDR4@2166, GeForce GTX Titan Black
• 5960X+B9
  i7-5960X, DDR4@2166, GeForce GTX Titan Black + GTX 960
• bunny
  2x Xeon E5-2660 v2, [email protected], GeForce GTX Titan
• bunnyX2
  2x Xeon E5-2660 v2, [email protected], 2 x GeForce GTX Titan
• hdd1
  i7-4930K, 780Ti,2x4GB DDR3@1333
• VIA Nano
  Via Nano U2250, 1 GB ram<
• 1.3
  v1.3-140-g08dba83
• 1.5d
  1.5 (rev 3fa1a0d7)
• NZ
  New Zealand UYVY
• NZ2500
  New Zealand frame 2500 UYVY

Latency

We also measured the latency added by the compression modules. For the tests 1080p@30fps was used, compressing at hd7 running Ubuntu and decompression done by hd2

module	end-to-end latency
uncompressed	3.75
cuda_dxt	3.75 (+0)
RTDXT:DXT1	6 (+2)
RTDXT:DXT5	5.5 (+1.75)
JPEG:90:0	4 (+0.25)
JPEG:97:0	4 (+0.25)
H.264	5 (+1.25)

Bandwidth

Here you can see the measured bandwidth including overhead with 9000B Ethernet frames. Uncompressed signal was 8-bit YUV422.

module	1080i@30	2k@30	4k (4096 × 2160)@25fps
uncompressed	980 Mbps	1504 Mbps	3489 Mbps
DXT1	245 Mbps	376 Mbps	870 Mbps
DXT5 YCoCg	489 Mbps	752 Mbps
JPEG:90	80 Mbps	85 Mbps	160 Mbps
H.264	22 Mbps	22 Mbps	60 Mbps
cineform:quality=4 (default)			580 Mbps
cineform:quality=1			300 Mbps