Skip to main content

Capabilities · Edge AI & Vision

From the camera to the decision, on the silicon your product can carry.

Camera pipelines, model optimisation and deployment on AMD adaptive SoCs, NVIDIA Jetson and Arm Cortex-M with Ethos NPUs, for industrial vision, automotive, drones and smart-city products built in India.

Send us the briefSensor to inference
  1. 01 Sensor to inference
  2. 02 Three silicon stacks
  3. 04 How to start
01 Sensor to inference

Six disciplines between a camera and a decision.

Each card lights its stage of the pipeline, from the sensor through the silicon to the model update that comes back from the field.

An edge AI pipeline from camera to model to three silicon targets, with the six GSAS edge AI services numbered as on the cards FP32 → INT8 FPGA SoC GPU module M55 NPU MCU + NPU Inference Field data · OTA 1. Camera and ISP: Vision pipeline 1 2. Trained model: Edge model optimisation 2 3. FPGA SoC: FPGA-accelerated inference 3 4. GPU module: NVIDIA Jetson engineering 4 5. MCU with NPU: TinyML on Cortex-M and Ethos-U 5 6. Safety monitor: Functional safety for AI 6 7. Fleet and OTA: MLOps 7
  • 1. Camera and ISP: Vision pipeline
  • 2. Trained model: Edge model optimisation
  • 3. FPGA SoC: FPGA-accelerated inference
  • 4. GPU module: NVIDIA Jetson engineering
  • 5. MCU with NPU: TinyML on Cortex-M and Ethos-U
  • 6. Safety monitor: Functional safety for AI
  • 7. Fleet and OTA: MLOps
Hover, tap or focus a number or a card to follow the pipeline.

  1. Vision pipeline and MLOps

    From the sensor to a model you can update in the field.

    • Camera bring-up, ISP tuning, MIPI CSI-2
    • Multi-camera sync and preprocessing
    • On-device data capture, drift detection, OTA model updates

    Pairs with embedded software and security

  2. Edge model optimisation

    A trained model, made to fit the silicon you actually picked.

    • INT8 quantisation
    • Pruning and distillation
    • Operator fusion for the target runtime

    Silicon-first, not vendor-first

  3. FPGA-accelerated inference

    Deterministic inference on AMD adaptive SoCs.

    • Vitis AI on Zynq UltraScale+, Versal AI Edge, Kria SOM
    • DPU integration and PetaLinux
    • HLS accelerator IP for non-DPU operators

    Pairs with FPGA design

  4. NVIDIA Jetson engineering

    GPU inference on a module, carried through to a product.

    • JetPack / L4T bring-up, carrier-board device tree
    • TensorRT optimisation, DeepStream pipelines
    • CUDA tuning and thermal envelope work

    Pairs with hardware design for carriers

  5. TinyML on Cortex-M and Ethos-U

    Always-on inference on a microcontroller budget.

    • Cortex-M55 with Ethos-U microNPUs
    • Vela tuning, CMSIS-NN, TensorFlow Lite Micro
    • Keil MDK underneath

    Arm Keil MDK supplied by GSAS

  6. Functional safety for AI

    ML in the loop, for teams working to ISO 26262 or IEC 61508.

    • Deterministic inference paths
    • Redundant-channel architecture
    • MISRA C checked glue code and static analysis

    Pairs with functional safety

02 Fitting the silicon

Make the model fit, then pick where it runs.

A model trained in the cloud rarely fits the device as it is. Optimisation shrinks it to the target; the target is chosen for latency, power and cost.

Target memory FP32 weights, as trained INT8 weights, on the device
Quantisation stores each weight in 8 bits instead of 32, so the model fits the target. Accuracy is checked after every step.

Edge model optimisation

Quantise, prune, distil and fuse, then measure accuracy again on the real target.

  • INT8 quantisation
  • Pruning
  • Distillation
  • Operator fusion

  1. A

    AMD adaptive SoCs

    Families
    Zynq UltraScale+, Versal AI Edge, Kria K26/K24 SOM
    Tool chain
    Vitis AI, Vitis HLS, Vivado, PetaLinux
    Reach for it when
    latency must be deterministic, or the pipeline has custom operators

    Through the customer’s own AMD entitlement

  2. B

    NVIDIA Jetson

    Families
    Orin Nano, Orin NX, Orin AGX, Xavier NX
    Tool chain
    JetPack / L4T, TensorRT, DeepStream, CUDA, cuDNN
    Reach for it when
    models are large, or many camera streams run at once

    Through the customer’s own NVIDIA entitlement

  3. C

    Arm with Ethos NPUs

    Families
    Cortex-M55 with Ethos-U55/U65/U85, Cortex-A with Ethos-N78
    Tool chain
    Keil MDK, Vela, CMSIS-NN, TF Lite Micro, Arm NN
    Reach for it when
    power is tight and the device runs on a battery

    Arm tool chain supplied by GSAS

03 Where it lands

Where edge vision shows up in products built in India.

The bench GSAS supplies

The commercial pieces of an edge-AI bench, supplied in India. AMD and NVIDIA tools come through the customer’s own entitlement.

Edge AI rarely ships alone

04 Working together

Three places teams usually get stuck.

Before the silicon is chosen, when the model will not fit, and on the road from a demo to a product.

  1. 01

    Choosing the silicon

    Before the board is designed

    Bring the model and the constraints: frame rate, power, cost, safety. GSAS helps evaluate the candidate platforms against them and recommends a target before you commit hardware.

    • Platform evaluation against your constraints
    • Camera and interface plan
    • A recommendation, with the reasons
    Compare platforms

  2. 02

    A model that will not fit

    Trained, not yet shippable

    A model that runs in the cloud but misses its budget on the device. Quantisation, pruning, distillation and operator fusion bring it inside memory and compute, checked against the accuracy you need.

    • INT8 quantisation
    • Pruning and distillation
    • Accuracy checked at each step
    Send the model brief

  3. 03

    Taking it to production

    From demo to product

    The work around the model that a product needs: camera pipeline, carrier board, safety architecture for ML in the loop, signed OTA updates and drift detection in the field.

    • Pipeline and carrier hardware
    • Safety and static analysis
    • OTA updates and MLOps
    Discuss a product
05 Contact

Building an edge AI or vision product?

Send the brief: the sensor, the model, the target and the constraints. GSAS comes back with the engineering conversation, not a brochure.

Send an RFQ All capabilities

sales@gsasindia.com +91 80 6590 1783

Questions

Edge AI and vision engineering with GSAS in India

What edge AI and vision engineering does GSAS do?

Vision pipelines (camera bring-up, ISP tuning, MIPI CSI-2, multi-camera sync), edge model optimisation (quantisation, pruning, distillation), FPGA-accelerated inference with Vitis AI, NVIDIA Jetson bring-up and TensorRT work, TinyML on Arm Cortex-M with Ethos-U, safety architecture for ML in the loop, and MLOps for models in the field.

Which edge AI platforms does GSAS work on?

AMD Zynq UltraScale+, Versal AI Edge and Kria SOMs; NVIDIA Jetson Orin and Xavier modules; and Arm Cortex-M55 with Ethos-U microNPUs and Cortex-A with Ethos-N. The platform is picked by the silicon target the product needs, not by vendor.

How do we choose between an FPGA, a Jetson and a microcontroller with an NPU?

Roughly: an AMD adaptive SoC when latency must be deterministic or the pipeline has custom operators, a Jetson module when models are large or many camera streams run at once, and a Cortex-M with Ethos-U when the device is power-constrained or battery-powered. GSAS helps make that comparison against your constraints before you commit.

Can GSAS make our trained model run on smaller hardware?

Yes. Quantisation to INT8, pruning, distillation and operator fusion take a model trained in the cloud and fit it to the target’s memory, compute and runtime, checked against the accuracy your application needs.

Does GSAS supply the tools as well?

GSAS supplies Arm Keil MDK, Digilent FPGA boards, Perforce Klocwork and Razorcat TESSY in India. AMD and NVIDIA tools come through the customer’s own entitlement.

Can GSAS help with functional safety for an AI feature?

GSAS works with teams following ISO 26262 or IEC 61508 on deterministic inference, redundant-channel architecture, and MISRA C checked glue code with static analysis and unit testing.

Blog

Edge AI & Vision Insights

Horizontal stacked bar showing where an ADAS test vehicle's bandwidth budget is spent, split into cameras, lidar, radar and bus traffic, with the logger uplink limit drawn as a vertical rule crossing the bar, from GSAS Micro Systems India
Automotive Ethernet Automotive & Mobility

ADAS Sensor Data Logging: Bandwidth Budgets That Add Up

Every page that tells you an ADAS test vehicle produces terabytes a day states the headline and skips the arithmetic, so you cannot redo it for your own sensor set. This article publishes the arithmetic instead: one formula, every table row derived on the page, a worked eight-hour drive that chains those rows into a sustained write rate, a media count and an offload window, and the five places bandwidth budgets go wrong. Written by the GSAS Micro Systems engineering team in India.

29 Aug 2026 · 15 min read
Three-panel workflow diagram of a SerDes camera bench: camera to deserialiser to recorder, then recorder to serialiser into the ECU for replay, then the same replay path with a fault injection block inserted before the ECU, from GSAS Micro Systems India
Automotive Ethernet Automotive & Mobility

SerDes Camera Links: GMSL, FPD-Link, ASA-ML and A-PHY

Four families carry automotive camera video, and only two of them have public documentation written for buyers. GMSL and FPD-Link are silicon-vendor interface families documented for board designers; ASA-ML and MIPI A-PHY are open specifications with published rate and reach figures and almost no practical buyer-facing writing. This page puts all four side by side with every rate class traced to the page it came from, then covers the part nobody writes: how you record, replay and inject a camera stream on a bench, and why none of it shows up in Wireshark. Written by the GSAS Micro Systems engineering team in India.

29 Aug 2026 · 14 min read
Two Cortex-M embedded engineers in a Bengaluru lab joining an Arm Keil MDK live webinar, with a CI/CD pipeline and an edge-AI model on screen
Keil MDK Arm

Two Arm Keil MDK Webinars This Summer: DevOps and Edge AI for Indian Cortex-M Teams

Arm is running two free live Keil MDK webinars this summer: DevOps with Keil MDK on 23 June and edge AI with ModelNova Fusion Studio on 21 July. What Cortex-M teams in Bengaluru, Hyderabad, Pune, Chennai, Mumbai, and Delhi NCR should know, shared by GSAS, Arm's authorized partner in India for Arm Development Tools.

17 Jun 2026 · 4 min read