Video Analytics GPU Sizing
How many GPUs a camera estate's analytics needs — stream density by model, resolution and frame rate, capped to what a chassis actually holds, with the power the deployment draws.
Enter the camera count, the analytics you are running, resolution and frame rate. The calculator returns the streams a GPU carries, the cards you need, how they split across servers, and what the deployment draws — with the whole derivation shown.
What it works out
Analytics sizing is a stream-density problem. How many camera feeds one GPU sustains while running your model decides everything else, and that figure moves six-fold between a light detector and a transformer one.
- Streams per GPU for the analytics model you actually run
- The effect of resolution, frame rate and the tracker, separately
- GPUs needed, capped to what the chassis holds
- Servers, when the requirement exceeds one box
- Typical draw, PSU rating, wall draw and heat load
Where the numbers come from
Two of the density figures are measured, from NVIDIA's own VSS RT-CV benchmark table: a ResNet50-class detector at 29 concurrent 1080p30 streams on an RTX PRO 6000, and a Grounding-DINO-class transformer at five. The rest are placed by model weight relative to those, and the page says which is which.
How video analytics sizing actually works
Stream density is the whole question
Analytics sizing comes down to one number: how many camera streams a single GPU sustains while running your model. Everything else is arithmetic on top of it. That number is not a property of the card alone — it moves several times over depending on which model you run, at what resolution, at what frame rate, and whether a tracker is doing part of the work.
NVIDIA's own published benchmarks make the range plain. On the same card, a ResNet50-class detector sustains around 29 concurrent 1080p streams at 30fps, while an open-vocabulary transformer detector manages five. A sizing that quotes one figure for 'analytics' is hiding a six-fold difference.
- Model weight is the largest factor — heavier detectors cost several times more per stream
- Frame rate scales almost linearly: 15fps roughly doubles what a card carries against 30fps
- Resolution scales sub-linearly — 4K is about twice the cost of 1080p, not four times
The tracker is not a detail
Production analytics rarely runs detection on every frame. It detects every second or third frame and uses a tracker to carry the boxes between, which roughly doubles the streams a card handles for no meaningful loss in what the system sees. Turning it off is sometimes correct — very fast motion, or forensic work where every frame matters — and it costs you half your density.
This is the single largest lever available after model choice, and it is worth deciding deliberately rather than inheriting from a default.
Why the answer is servers, not more GPUs
A compact chassis takes two GPUs. A tower or rack takes four. Past that, the answer is another machine — not because the arithmetic says so, but because that is what a chassis physically holds and what its power and cooling support.
That constraint is usually a benefit. A large estate split across several servers has more than one failure domain, distributes naturally across sites or buildings, and can be grown a server at a time rather than replaced. The sizing here caps at the chassis limit and adds machines, because a recommendation you cannot buy is not a recommendation.
What this does not size
Recording. Analytics compute and storage are separate problems with separate answers, and a complete surveillance build needs both — capacity for the retention period, and drives that sustain the write load without dropping frames. The NVR storage calculator covers that half.
Common questions
- How many cameras can one GPU handle for video analytics?
- Between roughly five and fifty 1080p streams, depending almost entirely on the model. NVIDIA's published benchmarks put a ResNet50-class detector at about 29 concurrent streams at 30fps on a high-end card and a transformer detector at five. Frame rate, resolution and whether a tracker is used move it further.
- Does frame rate or resolution matter more for analytics sizing?
- Frame rate, usually. Halving frame rate roughly doubles the streams a card carries, because it halves the frames analysed. Resolution costs less than people expect — 4K is around twice 1080p rather than four times, since detectors commonly resize to a fixed input.
- How many GPUs fit in one server?
- Two in a compact chassis, four in a tower or rack. Past that the deployment scales by adding servers, which is both what the hardware allows and what large estates do in practice — it gives you separate failure domains and a natural split across sites.
- Do I still need storage if I am running analytics?
- Almost always. Analytics tells you what happened; recording is what you produce when someone asks for the footage. They are sized separately — analytics on GPU throughput, recording on bitrate, retention and drive write capacity.
- Are these stream densities guaranteed?
- No, and any tool that says otherwise is not being straight with you. They are published benchmark figures for specific models on specific cards, and real density depends on your VMS, decode path, batch size and scene complexity. They are a defensible starting point, and we refine them against the deployment before quoting.