GPU server for sewer inspection video: AI defect coding to EN 13508-2 on your own hardware

What a GPU server does with sewer inspection video

A GPU server for sewer inspection video runs an AI model over the CCTV footage of a pipe inspection and proposes the defect codes an inspector would enter: cracks, displaced joints, roots, deposits, infiltration, each with its position along the pipe. The inspector reviews the proposals in the inspection software, confirms or corrects them and signs off the survey. The video stays on the operator's or contractor's own hardware, where it can be processed as soon as the vehicle returns, and the coded results go into the asset-management system as before.

This guide is for the managers of municipal sewer operators, water utilities and inspection contractors who decide whether to code footage with AI in-house, and for the IT staff who will run the server. The first sections explain the task, the published accuracy and the speed in plain terms. The configurations follow in one table, and the technical notes at the end are for IT. Where the article touches EU law it gives general information as of October 2026; your legal counsel confirms what applies to you.

Defect coding under EN 13508-2 and where AI fits

The European standard EN 13508-2, "Investigation and assessment of drain and sewer systems outside buildings - Part 2: Visual inspection coding system", is current in its version EN 13508-2:2003+A1:2011, according to the standards body DIN Media (German edition of August 2011). Its scope "specifies a coding system for the description of the internal condition of drains, sewers" and related structures, as identified by visual inspection, and says it does not in general specify how inspections are carried out. The codes describe what the inspector sees; the assessment of what to repair comes afterwards.

Coding is the slow part of an inspection. Someone watches the footage, stops at each observation, enters the code with its details and the distance from the start manhole, and moves on. An AI model can take the first pass: it watches every frame, marks the frames where something is visible and proposes a code. The inspector then jumps from proposal to proposal instead of watching the whole run in real time, and still sees every frame the model flagged.

Two points shape any project. First, the standard's code list is the target vocabulary, so the model is trained on footage that your inspectors have already coded to the same version and the same national rules. Second, the standard is a licensed document: putting its text into an AI assistant is a separate question from training a vision model on coded footage, which our guide to standards in your company AI explains.

What published research shows about accuracy

The largest open benchmark is Sewer-ML, published in March 2021 (arXiv 2103.10895). It contains 1.3 million images from 75,618 inspection videos of three Danish water utilities, recorded between 2011 and 2019 and annotated by professional sewer inspectors under a common Danish standard. Models predict 17 defect classes, and images were resized to 224 × 224 pixels for training. The paper's best method reached an F2 score weighted by class importance of 55.11 per cent on the test split, and an F1 score of 90.94 per cent for telling normal pipe from pipe with defects.

For a director, the reading is simple. Telling a clean pipe section from a section with something in it works well, which is what saves the most viewing time. Assigning the exact defect class is much harder, so the codes the model proposes are suggestions that a qualified inspector confirms. A model trained on your own coded footage, with your cameras and your pipe materials, can do better than a general one, and the only way to know by how much is a test on runs you have already coded.

The same dataset shows where training data comes from. A utility or contractor that has coded inspections for years already holds a labelled archive: every confirmed code is a frame with a label. Before using Sewer-ML itself for a commercial model, check its terms on the project page; the paper does not state a licence.

Why process the video in-house

Inspection video is large. As an example with values to replace by your own, one hour of 1080p footage at 8 Mbit/s is 3.6 GB, and 40 hours a week is 144 GB a week before backups. Processing on a server in the depot or the data centre avoids uploading that volume, lets the coding start when the vehicle returns and keeps footage of public infrastructure under the operator's control.

Model licences are the second reason to look closely. Ultralytics, which publishes the YOLO detection models, states on its licence page that its software is under AGPL-3.0 and that "An Enterprise License is required if you want to use Ultralytics YOLO without open-sourcing your entire project", also for internal use, and that "All Ultralytics YOLO trained models fall under the AGPL-3.0 License by default". Detectors under permissive licences exist: RT-DETR v2 with a ResNet-50 backbone, 43 million parameters, is on Hugging Face under Apache 2.0, trained on the COCO dataset of everyday objects and therefore fine-tuned on sewer footage before use. Our guide to the Ultralytics YOLO licence covers the choice in detail.

How fast one card processes footage

Two parts of the card do the work. The hardware video decoders, NVDEC, turn the compressed video back into frames, and the GPU cores run the model on the frames you choose to analyse. Inspection footage is recorded, not live, so the question is how many hours of footage one card processes per hour.

Decoder counts and power from NVIDIA's product pages (read 25 October 2026); ceiling = engines × NVIDIA's indicative 903 (Ada) or 2,172 (Blackwell) frames per second per engine at 1920 × 1080 from the NVDEC Application Note of Video Codec SDK 13.1, measured at the highest video clock, an upper bound by our arithmetic; frames per second from NVIDIA's DeepStream 9.2 performance page (RT-DETR at 3 × 640 × 640 FP16 without tracker; C-RADIO-B at 3 × 224 × 224 FP16), end to end on a 1080p H.265 sample file with rendering off, read 25 October 2026.

As a worked example by our estimate, take 1080p footage at 25 frames per second (an example value) and a detector that analyses five frames per second of video. One hour of footage is then 90,000 frames to decode and 18,000 to analyse. Planned to 70 per cent of NVIDIA's figures, an L40S decodes the hour in about 47 seconds and analyses it in about 40, so the decoder sets the pace at roughly 75 hours of footage per hour of processing. On an RTX PRO 4500 Server Edition the decoders are faster and the model slower: about 20 and 68 seconds, roughly 53 hours of footage per hour. A classifier at 224 × 224, like the Sewer-ML benchmark models, needs less computing than the detector.

Those numbers change how the server is sized. A contractor with ten vehicles that each record four hours a day, example values, brings in 40 hours of footage, which one L40S works through in well under an hour by this estimate. The card count is set by the other jobs: a backlog of archive footage, the training of your own model, and a language model that drafts the survey text. A backlog of 50,000 hours, for example, takes about 670 hours on one L40S and a quarter of that on four. These are estimates from NVIDIA's sample models; your own model and video settings decide, and a test run with a week of your footage gives the real figure.

We supply the NVIDIA L40S and the RTX PRO Server Edition cards and build GPU servers for inspection footage to order; tell us your hours of footage per week, the codec and the size of your archive, and our team replies with a configuration.

Configurations for operators and contractors

The table translates the estimates into servers. All of them belong in a server room or a cooled rack: the cards are passive and depend on the server's fans. Power is our estimate for the whole server with two AMD EPYC 9355 processors (cards, processors and 15 per cent for the rest), to be confirmed by a measurement on the delivered system; one 230 V 16 A feed carries 2.94 kW at 80 per cent.

Footage per hour by our estimate from the worked example above (75 hours per L40S, 53 per RTX PRO 4500 Server Edition, five analysed frames per second of 1080p25 footage, 70 per cent of NVIDIA's figures); power by our estimate as above (2 × L40S: 700 W of cards plus 560 W of processors, plus 15 per cent), the other rows as on our L40S and RTX PRO 4500 Server Edition server pages.

Two L40S. The NVIDIA L40S has 48 GB of GDDR6 with ECC at 864 GB/s, PCIe 4.0 and 350 W, without MIG or NVLink. Two of them process the footage of a typical fleet on one card and leave the second for training your own model, or both run footage at night. Our page on buying an L40S server lists the configurations.

Four RTX PRO 4500 Server Edition cards. The Blackwell card has 32 GB of GDDR7 at 800 GB/s, three decoders and 165 W in a single slot. Four of them decode more video than four L40S, run the detector more slowly and fit a 2U server at about the power of two L40S. Our comparison of the L40S and the RTX PRO 4500 Server Edition weighs the two.

Four or eight L40S. These suit a utility that wants to code an archive of years, retrain models regularly and run a language model for the survey text on the same server. Our page on the eight-card L40S server covers power, airflow and the models it holds.

Training your own sewer defect model

The most useful model is trained on your own footage. The work is a vision task on frames with labels from your coded inspections: a classifier for frame-level defect classes, as in Sewer-ML, or a detector that also draws a box around the defect. Both are small next to language models. RT-DETR v2 has 43 million parameters, so the 48 GB of an L40S leave room for large batches, and a two-card server can train on one card while the other processes the day's footage. Our guide to a GPU server for training vision models explains memory, batch size and storage for the training set, and visual inspection AI without data scientists shows how a team without machine-learning staff gets started.

Hold back a set of runs that the model never sees in training, coded by your best inspectors, and measure every new version against it per defect class before it replaces the old one.

Report text and the AI Act

A language model can turn the confirmed codes of a run into the plain-language survey summary for the client. Qwen3.8-27B in Qwen's FP8 version fits one L40S and leaves room for seven conversations of 32,768 tokens by our users-per-card rule, as our page on Qwen3.8-27B on the L40S works out. The text starts from codes a person has confirmed, and the inspector approves it.

Point 2 of Annex III of the AI Act lists as high-risk "AI systems intended to be used as safety components in the management and operation of critical digital infrastructure, road traffic, or in the supply of water, gas, heating or electricity". By our reading, a model that proposes inspection codes for an inspector to confirm is not a safety component in the operation of a network, and waste water is not named in that point. Your legal counsel confirms the classification for your use. Our guide to the AI Act for companies using LLMs covers the general duties.

Technical notes for the IT team

1. Pipeline. NVIDIA DeepStream decodes on NVDEC and runs TensorRT engines; its inference plug-in can skip frames, so the model analyses five of 25 while every frame is decoded. Our guide to serving vision models with Triton, ONNX and TensorRT covers the deployment options.
2. Codecs. The NVDEC Application Note lists MPEG-2, MPEG-4, VC-1, H.264, HEVC, VP8, VP9 and AV1 on Ada and Blackwell, so older archive formats decode on the GPU too. Decode rates scale with the video clock that nvidia-smi reports on the card.
3. Position along the pipe. Each proposal needs the distance from the start manhole. Take it from the inspection software's metadata where it exists; where it is only burnt into the picture as text, read it from the frame and check it against the run length.
4. Integration. Write proposals back into the inspection software as unconfirmed observations, so only confirmed codes reach the asset-management export. Keep the original footage unchanged.
5. Security. Keep the server on an internal network, apply NVIDIA's fixed driver releases and harden the management controller as in our guide to securing a GPU server.

What we supply

You can buy this system from us: our team sizes the configuration to your workload and budget and replies within one business day. We supply the NVIDIA L40S, the RTX PRO 4500 Server Edition and the RTX PRO 6000 Server Edition as cards for servers you already run or in GPU servers built to order, assembled and tested before delivery, delivered anywhere in the EU on one EU contract and invoice, with manufacturer warranty on every component handled through us; quantities and lead time are confirmed in writing with the quote. Operating system, drivers, CUDA and a container runtime are installed on request; send us a short description of your fleet and archive and we propose a server for a test with your own footage.

FAQ