ReferenceExaminable
Numbers to know
Print it, fold it, put it in the pocket of the jacket you wear to work. There is a five minute verbal quiz at the top of every class.
Student reference card. Learn these. There is a five minute verbal quiz at the top of every session, and you may bring this page into the calculate station of the practical exam.
Print it. Fold it. Put it in the pocket of the jacket you wear to work.
Block 1: Data rates (learn before the production visit)#
| Thing | Number |
|---|---|
| Speed of sound | about 343 m/s, so 3 ms per metre |
| One audio channel, 48 kHz, 24 bit | 1.152 Mbit/s |
| 64 audio channels | about 74 Mbit/s of audio, call it 100 Mbit/s on the wire |
| One DMX universe | about 0.25 Mbit/s |
| Uncompressed 1080p60, 10 bit, 4:2:2 | about 2.5 Gbit/s |
| Uncompressed UHD 4K 60, 10 bit, 4:2:2 | about 10 Gbit/s |
| NDI, one HD stream | roughly 100 to 250 Mbit/s |
The ranking, and you should be able to say it in your sleep: clock < control < audio < video.
The arithmetic:
audio = sample rate × bit depth × channels
video = width × height × frame rate × bits per pixel
Block 2: The machine (learn before Class 3)#
| Thing | Number |
|---|---|
| Spinning hard disk, sustained read | 100 to 200 MB/s |
| SATA SSD, sustained read | about 550 MB/s |
| NVMe SSD, sustained read | 2,000 to 7,000 MB/s |
| Audio buffer, 128 samples at 48 kHz | 2.67 ms per buffer |
| Audio buffer, 256 samples at 48 kHz | 5.33 ms per buffer |
| Audio buffer, 512 samples at 48 kHz | 10.67 ms per buffer |
| Round trip latency | roughly twice the buffer, plus converter time |
| Dynamic range per bit | 6.02 dB, so 16 bit is about 96 dB and 24 bit about 144 dB |
| Nyquist | you can only capture frequencies below half the sample rate |
| One byte in hex | 00 to FF, which is 0 to 255 |
| 16 bit signed sample range | −32,768 to +32,767, silence at 0 |
| Luma weights | 0.299 R + 0.587 G + 0.114 B. Green dominates, blue barely counts. |
52 49 46 46 | RIFF, so a WAV file. The header is the truth, the extension is a hint. |
FF D8 FF | always a JPEG |
89 50 4E 47 | always a PNG |
| The lossy steps in a codec | chroma subsampling and quantisation. Quantisation is the quality slider. |
| What sets bitrate | how predictable the picture is, not the resolution alone |
| Light in glass fibre | about 5 µs per km, so 5 ms per 1,000 km, each way | | Sound in air | about 3 ms per metre | | A frame at 25 fps | 40 ms | | Camera to LED wall, typical | 3 to 7 frames, so 120 to 280 ms | | Precedence window | 5 to 35 ms: two arrivals fuse and localise to the first | | Delay tower setting | the distance from the PA, in ms, plus 10 to 15 |
Bits to bytes: divide by 8. A 2.5 Gbit/s video stream is about 312 MB/s of storage throughput. Getting this the wrong way round is the most common mistake in the exam.
Block 3: Addressing and the network (learn before Class 4)#
| Thing | Number |
|---|---|
| Cat cable maximum channel length | 100 m (90 m solid plus 10 m patch) |
| Link speeds you will meet | 100 Mbit, 1 Gbit, 10 Gbit, 25 Gbit |
255.255.255.0 | the same as /24. First three parts are the network. |
169.254.x.x | "I asked for an address and nobody answered." Recognise on sight. |
| Private ranges | 10.x.x.x, 172.16 to 172.31.x.x, 192.168.x.x |
| Art-Net legacy range | often 2.x.x.x or 10.x.x.x |
| PoE (802.3af) | up to 12.95 W at the device |
| PoE+ (802.3at) | up to 25.5 W at the device |
| PoE++ (802.3bt) | up to 51 W (Type 3) or 71 W (Type 4) at the device |
The rule: two devices can talk directly only if the network part of their addresses is identical under the mask.
Three symptoms to recognise instantly:
169.254.x.x→ DHCP was expected and there is no DHCP server.- Intermittent, confusing, both devices misbehaving → duplicate IP.
- Works fine, then stops several minutes later → multicast, snooping on, no querier.
The nine bit values (learn these first)#
| Bits set | Value | Bits set | Value |
|---|---|---|---|
| 0 | 0 | 5 | 248 |
| 1 | 128 | 6 | 252 |
| 2 | 192 | 7 | 254 |
| 3 | 224 | 8 | 255 |
| 4 | 240 |
The prefix table (learn these second)#
| Prefix | Mask | Block size | Usable hosts |
|---|---|---|---|
| /16 | 255.255.0.0 | 65,536 | 65,534 |
| /22 | 255.255.252.0 | 1,024 | 1,022 |
| /23 | 255.255.254.0 | 512 | 510 |
| /24 | 255.255.255.0 | 256 | 254 |
| /25 | 255.255.255.128 | 128 | 126 |
| /26 | 255.255.255.192 | 64 | 62 |
| /27 | 255.255.255.224 | 32 | 30 |
| /28 | 255.255.255.240 | 16 | 14 |
| /29 | 255.255.255.248 | 8 | 6 |
| /30 | 255.255.255.252 | 4 | 2 |
The subnet method, four steps#
1. Block size = 256 - (last non-zero mask octet)
2. Find which block the address falls in, counting in steps of the block size
3. Network = first address in the block. Broadcast = last address in the block.
Usable = everything between them.
4. Usable hosts = 2^(32 - prefix) - 2
The OSI layers#
| # | Layer | On a show |
|---|---|---|
| 7 | Application | sACN, Art-Net, Dante, NDI, OSC |
| 6 | Presentation | Codecs, sample formats, encryption |
| 5 | Session | A Dante subscription, an open control session |
| 4 | Transport | TCP and UDP, port 5568 and 6454 |
| 3 | Network | IP addresses, masks, routers, IGMP |
| 2 | Data link | MAC addresses, switches, VLAN tags |
| 1 | Physical | Cable, fibre, connectors, link lights, PoE |
Diagnose from the bottom up. Link light, then VLAN, then IP, then port, then software.
VLAN numbers#
| Thing | Number |
|---|---|
| VLAN ID range | 1 to 4094 |
| The 802.1Q tag adds | 4 bytes to the Ethernet frame |
| Access port carries | exactly 1 VLAN, untagged |
| Trunk port carries | many VLANs, tagged |
| Subnets from borrowed bits | 2^bits, so 2 bits gives 4 subnets |
Block 4: Protocols (learn before Class 5)#
| Thing | Number |
|---|---|
| DMX512 bit rate | 250 kbit/s |
| DMX512 slots per universe | 512, each one byte, 0 to 255 |
| DMX512 maximum refresh | about 44 Hz (roughly 23 ms per full frame) |
| DMX512 devices per segment | 32 unit loads, then a splitter |
| DMX512 termination | 120 ohms on the last device |
| RGB pixels per universe | 512 ÷ 3 = 170 |
| RGBW pixels per universe | 512 ÷ 4 = 128 |
| Art-Net | UDP port 6454 |
| sACN (E1.31) | UDP port 5568, multicast |
| Timecode frame rates | 24, 25, 29.97, 30 |
The DMX refresh calculation, learn to reproduce it:
512 slots × 11 bits = 5,632 bits
5,632 ÷ 250,000 = 22.5 ms
plus break and mark ≈ 23 ms per frame
1 ÷ 0.023 ≈ 44 frames per second
Block 5: Media over IP and perception (learn before the practical exam)#
| Thing | Number |
|---|---|
| Dante latency settings | 0.25, 0.5, 1, 2, 5 ms. 1 ms is the safe default. |
| Channels on a 1 Gbit link, 48 kHz 24 bit | several hundred each way. Bandwidth is rarely the limit. |
| HD-SDI | 1.485 Gbit/s |
| 3G-SDI | 2.97 Gbit/s, carries 1080p60 |
| 12G-SDI | 12 Gbit/s, carries UHD60 |
| ST 2110 network requirement | 10 Gbit minimum for HD, 25 Gbit for UHD |
| PTP accuracy | well under a microsecond |
| In ear monitoring feels wrong beyond | roughly 5 to 10 ms |
| Audio ahead of picture, detectable at about | +45 ms (ITU-R BT.1359) |
| Audio behind picture, detectable at about | −125 ms (ITU-R BT.1359) |
| Broadcast delivery tolerance commonly cited | +40 / −60 ms (EBU R37) |
Why the asymmetry: thunder always follows lightning, so we are built to accept sound arriving after picture. Sound arriving before picture is unnatural and we catch it roughly three times faster.
The two things you will be asked to do, in every assessment#
1. Data rate maths. Will this fit down this wire?
channels or pixels × rate × depth = bits per second
2. Latency budget. Add every stage from trigger to eardrum or retina. Compare it against a perceptual threshold. State pass or fail, and say which threshold you used.
The two definitions people get wrong#
Latency is delay. If it is constant, you can measure it and compensate for it. A delay tower is a latency compensation device and nobody calls it a fault.
Jitter is delay that varies. You cannot compensate for a number that will not sit still. Your only defence is a buffer, and a buffer costs latency.
Jitter is what breaks shows.