Camera capacity
Aperture estimates how much of your machine its cameras use, from a hardware profile you confirm and from what each camera makes Aperture do. The dashboard shows the result as a capacity meter, and the camera editor warns before a new camera would take the machine past what it runs well. This page explains how to set the profile, how to read the numbers and where the estimate stops being reliable.
Capacity is a soft limit. Nothing is blocked at 100%: past that point recording, live view and playback may stutter, and there is less headroom for exports.
Set the hardware profile
Section titled “Set the hardware profile”Without a profile, the dashboard shows “Set the hardware Aperture runs on to see how many cameras it handles.” instead of a meter. An administrator sets it once.
- Open Settings › Hardware.
- Detected on this server shows what the machine reports about itself: the processor, the CPU threads, a CPU limit on Aperture’s service if there is one, the memory and any accelerators found. Choose Use detected values to fill the form from it.
- Processor: search the catalogue and pick your CPU. It covers more than 850 processors: desktop, laptop and mini PC, NAS, server, single-board and Apple. In a virtual machine the detected CPU is often a generic one such as “QEMU Virtual CPU”; pick the host’s real processor.
- CPU threads for Aperture: all of the processor’s threads on bare metal; in a VM or container, the vCPUs it has, or the service’s CPU limit if that is lower.
- Memory: the memory of the machine or VM, in GB.
- Network: the link the cameras and any network storage share (100 Mb, 1 GbE, 2.5 GbE, 5 GbE, 10 GbE or 25 GbE).
- AI accelerator: switch it on and pick the device if object recognition runs on an accelerator in this machine. Aperture’s own recognition can use an NVIDIA GPU; other chips (a Coral, a Hailo and so on) help only an external detection server running on this machine. Recording and motion detection stay on the CPU either way. A re-encoded camera runs on an NVIDIA GPU when there is one (see Recording), and the meter counts it at almost no CPU.
- Choose Save hardware.
The Camera capacity section below the form previews the result with the values you entered, before you save. The installed service unit limits Aperture’s CPU and memory; Detected on this server shows those limits, and the threads field suggests the lower of the two.
The server only stores the profile. The calculation runs in your browser, from the profile, the cameras’ settings and their resolutions.
Read the meter
Section titled “Read the meter”The capacity meter is in the dashboard’s right column, under storage and system status, labelled Camera Capacity.
- The dot and the percentage. “12% used”, or “Over capacity · 115%” past 100%. The dot is green up to 85%, amber above 85% and red above 100%.
- The number next to the label is how many cameras the dashboard shows.
- The bar has one segment per camera. A segment’s width is that camera’s share of whichever resource limits the machine, and its colour is the camera’s resolution tier. A faded segment is a camera whose resolution is not known yet, counted as 4K. Over 100%, the bar is scaled to fit and ends in red. Hovering a segment names the camera, its tier, its share and any notes, such as “Motion detection decodes the full stream (no sub stream)”.
- The tier chips count the cameras per resolution tier: HD, 2K, 4K, and 4K+ when there is one.
- Hovering the percentage shows what limits the machine and how many plain cameras it runs, for example “Limited by the processor · runs about 32 × 1080p, 21 × 1440p, 10 × 4K”.
Administrators can click the bar to open Settings › Hardware, which shows the full breakdown: CPU, memory and network (and the accelerator) each as used of available, a table of every camera’s share with its notes, and a line such as “This hardware runs about 32 × 1080p · 21 × 1440p · 10 × 4K cameras, recording with motion detection on a sub stream”.
Resolution tiers
Section titled “Resolution tiers”| Tier | Main stream size | Typical cameras |
|---|---|---|
| HD | up to 2.5 megapixels | 720p, 1080p |
| 2K | up to 5.1 megapixels | 1440p, 4 and 5 MP |
| 4K | up to 9.5 megapixels | 4K / 8 MP |
| 4K+ | above 9.5 megapixels | 12 MP, dual-lens panoramas |
Aperture reads each camera’s resolution from a snapshot of its main stream. A camera whose snapshot cannot be loaded (offline, or not tested yet) counts as 4K until it can. Only enabled cameras count.
How it is calculated
Section titled “How it is calculated”The model measures work in core-equivalents: one Intel Skylake core at 4 GHz. The code is web/src/lib/capacity.ts, the processor catalogue is web/src/lib/hardware/cpus.ts and the per-architecture speeds are in arch.ts.
The machine’s budget
Section titled “The machine’s budget”- CPU. Every catalogue entry is described by its core clusters: how many cores, which microarchitecture, the top clock, and whether it has SMT (Hyper-Threading). Each core counts as its architecture’s speed relative to Skylake, times its clock derated to a sustained all-core speed (85% of boost on Intel and AMD, 95% on others), divided by 4 GHz. Threads beyond the physical cores add a quarter of a core each. The total is limited to the threads you entered, and a fifth of it is kept free for the system, the web app, playback and exports.
- Memory. 85% of the memory you entered. Aperture itself is counted as 0.5 GB before any camera.
- Network. 70% of the link speed.
- AI accelerator. How many cameras’ object recognition the device keeps up with, from the catalogue (shown as “object detection for about N cameras” when you pick it).
What a camera costs
Section titled “What a camera costs”Costs scale with the camera’s size relative to 1080p, called w below: a 1080p camera is 1, a 4K camera is 4. Unknown cameras count as 4.
| Work | CPU (core-equivalents) | Memory |
|---|---|---|
| Ingest and restreaming, snapshots | 0.02 × w | 0.01 GB + 0.012 GB × w |
| A share of live viewing and playback | 0.015 × w | |
| Recording, stream copy | 0.012 × w | 0.03 GB |
| Motion detection on a sub stream | 0.02 | 0.065 GB |
| Motion detection on the main stream (no sub stream) | 0.1 × w | 0.065 GB + 0.05 GB × w |
| Re-encoded recording | 0.1 × w + 0.25 × the output size (at least 0.5) | 0.25 GB + 0.07 GB × w, instead of the 0.03 GB |
| Object recognition on the CPU | 0.15 | 0.05 GB |
| Object recognition on an accelerator | 0.02, plus one camera of the accelerator’s budget | 0.05 GB |
For the network, the model assumes a camera bitrate of 3 × w^0.8 Mbit/s: about 3 Mbit/s for 1080p and 9 Mbit/s for 4K. A camera that records or mirrors to an SMB or SFTP destination counts that stream a second time, because the footage leaves the machine again (for a re-encoded camera, about 60% of the smaller output’s estimate).
Aperture works out these inputs from each camera’s settings:
| Input | Counted when |
|---|---|
| Recording | Recording mode is not Off |
| Motion detection | Motion detection is on, or the mode is Motion only |
| Sub stream | the camera has a Sub stream address |
| Re-encoding | any Recording quality field is not Original or Automatic |
| Object recognition | recognition runs on this server for the camera, or an external detector at a localhost address does |
| Network storage | the camera writes or mirrors to an SMB or SFTP destination |
The percentage is the fullest of the budgets: CPU, memory, network and, with an accelerator, the accelerator. That budget is named as the limit (“limited by the processor”).
A worked example
Section titled “A worked example”An Intel N100 has four Gracemont cores at up to 3.4 GHz and no Hyper-Threading. Its budget is 4 × 0.95 × (3.4 × 0.85 / 4) ≈ 2.75 core-equivalents, of which 80%, about 2.2, is for cameras.
- A plain 1080p camera (recording, motion on a sub stream) costs 0.02 + 0.015 + 0.012 + 0.02 = 0.067. About 32 fit.
- A plain 4K camera costs 0.047 × 4 + 0.02 ≈ 0.21. About 10 fit.
- The same 4K camera without a sub stream pays 0.4 instead of 0.02 for motion detection, because it has to decode its main stream: 0.59 in total, more than a quarter of the machine.
- A 4K camera re-encoded to 1080p adds 0.1 × 4 + 0.25 ≈ 0.65 on top of its plain cost: about 0.86, nearly 40% of this machine for one camera.
With 16 GB of memory and a 1 GbE link, memory and network allow far more cameras than that, so this machine is limited by its processor.
Where the figures come from
Section titled “Where the figures come from”The costs are calibrated on measurements of Aperture itself. Stream-copy recording is nearly free; decoding video, for motion detection without a sub stream or for a re-encode, is what fills a CPU. Memory figures come from the unique memory (PSS) of the processes involved: about 65 MB for a motion decoder on a sub stream, about 30 MB for a stream-copy recorder, and 0.5 to 0.8 GB for a re-encoder. The architecture speeds are rounded ratios from published benchmarks, applied to every processor by the same rule rather than a number per model.
Plan more cameras
Section titled “Plan more cameras”When you add a camera that would take the machine past 100%, the camera editor shows a warning at the top (“This camera takes Aperture to 112% of its capacity”) with what limits it. Saving asks Add camera anyway?, and Add anyway adds it. Editing a camera shows the same warning when your changes add load past 100%, without the extra question. A new camera counts as 4K until Test connection shows its real resolution, so run the test before you judge the number.
To make room for more cameras:
- Give every camera a sub stream. Motion detection on a sub stream costs 0.02; on a 4K main stream it costs 0.4, twenty times as much.
- Avoid re-encoding unless you need the space. To record smaller files for every camera at no CPU cost, use Settings › General › Recording quality › Prefer low quality — sub stream (saves space) instead (see Recording).
- Move object recognition to an accelerator if you run it on many cameras.
- Leave headroom. Staying under about 85%, where the dot turns amber, leaves room for several people watching at once, multi-camera playback and exports.
Limits of the estimate
Section titled “Limits of the estimate”The meter is a model, not a measurement. It does not watch the machine’s real CPU or memory use.
- Bitrate and codec are assumed. The network figure uses an estimated bitrate from the resolution, not what the camera actually sends, and the model does not distinguish H.264 from H.265.
- The global recording-quality choice is not modelled. Recording to the sub stream costs less than the model counts, though recording is a small part of the total.
- Viewing is an average share. Many simultaneous viewers, multi-camera playback, timelapses and exports come out of the 20% kept free, and heavy use can exceed it.
- Only listed hardware is rated. If your processor is not in the catalogue, pick the closest one: the same architecture, core count and clock. An accelerator is rated by how many cameras’ recognition it keeps up with, from its published throughput.
- Storage is not part of it. Disk space and disk speed are not modelled; see Storage for space.
- Unknown resolutions count as 4K, which overstates the load of small cameras until their snapshot loads.
Related: Requirements · Recording modes and schedules · Object recognition · Storage