If video freezes on a camera network, the camera port is rarely to blame. The bottleneck is in the uplink, the backbone or the recording server’s network card. The method below starts from per-camera bitrate and works out separate totals for the access switch, backbone, NVR/VMS server and WAN; it was written for IT managers and network specialists sizing a camera project together with the network team.
The streams that make up camera traffic: recording, live viewing, substream
A common calculation mistake is stopping at “number of cameras × bitrate”. A camera system has at least five independent traffic items, and each passes through a different section of the network:
- Recording stream (main stream): The high-resolution stream flowing continuously or event-based from the camera to the recording server. It loads the access uplink, the backbone and the server NIC.
- Substream: A low-resolution second stream for mosaic viewing, mobile clients and some analytics engines. Some VMS products pull it continuously, others only on demand.
- Live viewing clients: Operator stations, video wall decoders. If the VMS relays the stream through the server (proxy), this traffic appears in the server’s outbound direction; if the client pulls directly from the camera, it passes through the camera uplink a second time.
- Backup and archive copy: Copying recordings to secondary storage, a disaster recovery site or an archive tier.
- Remote viewing: Streams pulled over the WAN/internet by branches, a central monitoring room or mobile users.
That’s why the calculation isn’t a single number but a separate total for each link. An access uplink carries only what the cameras on that switch send; the recording server’s NIC receives all cameras’ recording streams and sends the streams going to all clients. On a full-duplex link, the two directions are calculated separately. A single total is misleading.
How to determine per-camera bitrate
Bitrate varies with resolution, frame rate, codec (H.264/H.265 and the manufacturer’s smart codec mode), GOP length, scene motion and night noise. The “maximum bitrate” on the datasheet isn’t a design value. The real value is measured on site:
- Fix the target settings for each camera type (resolution, fps, codec, CBR/VBR cap) in the project specification. Don’t let anyone tweak them after installation to “bump it up a bit”.
- Pick a representative scene and record bitrate for 24 hours, by day, at night and at busy times, from VMS statistics or switch port counters.
- In the design, use the VBR cap or the measured peak, not the average. Noise under IR at night, rain, snow and trees swaying in the wind noticeably raise a VBR stream.
A design that measures by day and guesses the night blows up on the first winter night. For the effect of codec and bitrate control mode (CBR, VBR, smart codec) on the peak, the CBR, VBR and smart codec bitrate settings guide clarifies the input to this step.
Protocol overhead also goes into the calculation. Example calculation: for a 1,400-byte video payload carried over UDP/RTP via RTSP, the RTP header defined in RFC 3550 (12 bytes), UDP (8), IP (20) and the Ethernet header, FCS, preamble and inter-frame gap (38 bytes in total) are added; 78 bytes of overhead is about 5.6%. If the stream is carried via RTSP-over-TCP, the ratio is slightly higher because of the TCP header and the 4-byte interleave header. Increasing bitrate by about 6–8% to convert it to line rate is a safe assumption.
Calculating access switch and uplink capacity
At the access layer, the calculation is simple: add up all streams sent by every camera connected to the switch. On the camera port side, even 100 Mbit/s is enough for most cameras. The real question is the uplink.
Uplink traffic = Σ (main stream peak bitrate + continuously pulled substream + live streams pulled directly from the camera) × protocol overhead
Don’t skip two behaviors here. First, video comes in bursts with I-frames; a stream averaging 6 Mbit/s produces packet trains at port speed for brief moments. When many cameras empty into the same uplink, these micro-bursts overflow the output queue even if average utilization looks low. The effect is more pronounced on entry-level PoE switches with small buffers. The graph is calm, the recording is missing frames. Second, if the VMS takes live viewing directly from the camera, the same stream passes through the uplink as a separate unicast copy for each operator; the solution is multicast and IGMP snooping or relaying through the server.
When choosing the uplink, aim for calculated peak traffic not to exceed half the uplink capacity. This isn’t a standard; it’s design headroom left for micro-bursts and cameras added later. If the calculation for a 24-camera access switch stays at a few hundred Mbit/s, a 1G uplink is technically enough. But over the switch’s lifetime both camera count and resolution increase; choosing an access switch with 10G SFP+ uplinks in a new project is often cheap insurance. The PoE power budget is a separate sizing item, independent of this calculation.
Sizing the backbone and the NVR/VMS server side
In the backbone, the total of all access uplinks converges on the links to the recording servers. On the server side, three separate values are calculated:
- Ingest: Recording streams of all cameras assigned to the server + continuously pulled substreams.
- Egress: Streams going to live viewing and playback clients, the video wall and the archive/backup target.
- Disk write rate: Recording bitrate converted to bytes; 840 Mbit/s of recording means about 105 MB/s of continuous writes. This value is the input to storage design.
Bonding two 1G ports on the server with LACP increases total capacity, but LACP hashes each flow to a single physical link. Many camera streams usually spread evenly; a single large flow (for example an archive copy), however, is capped at the speed of one link. Two ports show in the console, the flow goes over one. In designs where server ingest approaches 1G, a 10G NIC is both simpler and more predictable. If there are several recording servers, also account for how cameras are distributed across servers and the extra load the standby server takes on during failover.
Remote viewing and WAN traffic
The WAN is the narrowest and most expensive link in a camera system. The basic rule: the main stream doesn’t go out to the WAN. Remote users view the substream or a low-bitrate stream transcoded by the VMS; the main stream is pulled only for incident review, and briefly. The calculation goes like this:
WAN traffic = concurrent remote users × views per user × bitrate per view
If branches record to the center, the picture changes: all of the branch’s recording traffic goes over the WAN. In this scenario, local recording (an NVR at the branch or an SD card on the camera) with only event clips or the substream sent to the center is usually more realistic. This architectural decision is explained in detail in the multi-site camera management article.
Example calculation: a 120-camera facility
The values below are an example calculation; bitrates are assumptions, so replace them with peak values measured on site in your own project. Assumptions: 100 4 MP cameras (H.265, 6 Mbit/s peak), 20 8 MP cameras (12 Mbit/s peak), a 0.75 Mbit/s substream pulled continuously from all cameras to the VMS, live viewing relayed through the server, 5 access switches (24 cameras each).
| Item | Calculation | Result |
|---|---|---|
| Recording (main stream) | 100×6 + 20×12 | 840 Mbit/s |
| Substream (continuous) | 120×0,75 | 90 Mbit/s |
| Server ingest | (840+90)×1,07 | ~995 Mbit/s |
| Operator mosaics | 4 stations×16×0.75 | 48 Mbit/s |
| Video wall | 12 windows×6 | 72 Mbit/s |
| Archive copy | 840×24÷8-hour window | 2,520 Mbit/s |
| Remote viewing (WAN) | 3 users×4×0.75 | 9 Mbit/s |
Decisions from the table:
- Access uplink: Even if all the 8 MP cameras are connected to the busiest switch (20×12 + 4×6 + 24×0.75 ≈ 282 Mbit/s), a 1G uplink is enough; a 10G uplink is recommended for growth.
- Server NIC: About 1 Gbit/s of ingest saturates a single 1G port; 2×1G LACP is also a borderline solution with no headroom. A 10G NIC should be preferred; the roughly 120 Mbit/s of viewing traffic in the outbound direction rides on the same card.
- Archive copy: Copying the day’s recordings in an 8-hour night window requires 2.5 Gbit/s. Either continuous replication (840 Mbit/s) is chosen or the copy is moved to a separate storage network. This is the item that gets overlooked.
- Disk writes: 840 Mbit/s ≈ 105 MB/s of continuous writes; for storage and retention, use the NVR recording time and storage calculation guide.
Measurement and verification: port counters and oversubscription
A calculation on paper isn’t finished until it’s verified on site. On Cisco IOS/IOS XE switches, the default 5-minute average hides peak traffic; shorten the averaging period on the interfaces to be measured:
interface TenGigabitEthernet1/1/1
load-interval 30
!
show interfaces TenGigabitEthernet1/1/1 | include rate|drops
show interfaces counters errors
The first place to look in the console is the “Total output drops” counter. If this counter increases while average utilization is low, the queue is overflowing because of micro-bursts. On the server side on Linux, instantaneous in/out rates are monitored with sar -n DEV 1 and driver-level drops with ethtool -S <interface>. Read frame loss and recording gaps in the VMS’s own statistics side by side with network counters; if both spike at the same time, the problem is in the network.
The oversubscription ratio is the ratio of total port capacity to uplink capacity: for a 48-port 1G switch with 2×10G uplinks, 48:20 = 2.4:1. In an office network, this ratio is a reasonable design parameter. In a camera network, what matters isn’t port capacity but the real sum of streams; since cameras produce continuous traffic in the same direction, track oversubscription not by port count but by the ratio of measured peak traffic to the uplink. For queue prioritization in networks that share infrastructure with voice and other sensitive traffic, see the QoS and DSCP design in a corporate LAN article.
If you want to address uplink and backbone design together with camera placement, you can see the scope on the PoE switch and camera network infrastructure and fiber optic camera backbone design service pages.
Frequently asked questions
How much bandwidth does an IP camera use?
There’s no fixed value; it varies with resolution, frame rate, codec, scene motion and night noise. Use not the maximum value on the datasheet but the peak bitrate measured on site over 24 hours with the target settings, and add protocol overhead on top.
Is a 1G uplink enough on a camera switch?
If the peak sum of the connected cameras’ main streams, continuous substreams and directly pulled live streams is below half the uplink capacity, it’s technically enough. For micro-bursts and growth headroom, a 10G uplink is usually the safer choice in new projects.
Why does an NVR or VMS server need a 10G network card?
While receiving all cameras’ recording streams, the server simultaneously sends streams to clients and the archive target. In the 120-camera facility in the example calculation, ingest rises to about 1 Gbit/s; since LACP limits a single flow to a single link, a 10G NIC gives more predictable results.
Why are there dropped frames when average utilization is low?
Video streams produce bursty traffic at I-frames; when many cameras empty into the same uplink, the switch’s output queue overflows briefly. This doesn’t show in a 5-minute average; it’s caught with the output drops counter and a shortened load-interval.