FloodWatch Joins the NVIDIA Inception Program

We're pleased to announce that FloodWatch has been accepted onto the NVIDIA Inception Program.

What is NVIDIA Inception?

The NVIDIA Inception Program is a virtual accelerator supporting AI startups with
go-to-market resources, technical expertise, NVIDIA Deep Learning Institute training,
and preferred access to GPU compute and cloud infrastructure. It's a global network
of over 19,000 companies building at the frontier of AI.

Our Current NVIDIA Stack

FloodWatch is already training on an NVIDIA stack. We're developing LSTM and GRU
models for river gauge level forecasting across hundreds of monitoring stations
nationwide, using CUDA and PyTorch with Automatic Mixed Precision. Training runs are
parallelised across multi-GPU nodes; our most recent model run executed on an
8x NVIDIA H100 SXM instance via Lambda, with 80GB HBM3 VRAM per GPU, 208 vCPUs,
1.8TB RAM, and 22TB of NVMe SSD, with GPU-to-GPU communication running at 900GB/s
via fourth-generation NVLink.

Alongside this, RAPIDS cuGraph is being integrated for river network graph analytics
as we move deeper into Graph Neural Network research.

Lambda: GPU Compute Built for Developers

Lambda — an NVIDIA Inception partner — has been a natural fit for our
infrastructure. Their developer-first GPU cloud offers instant access to on-demand
H100 nodes without the friction of traditional cloud provisioning, making it
straightforward to spin up large training runs and scale back down when done. For a
startup training at this scale, that flexibility matters.

Accelerating Our Research

The NVIDIA Inception Program will help accelerate our work across a number of
research areas:

Multi-horizon river level forecasting
1D and 2D hydrological modelling
Precipitation nowcasting
Flood extent modelling
Satellite-based flood analysis using SAR and optical imagery
DTM and DSM terrain processing at national scale

Our Vision: A Hydrological Digital Twin

The longer-term vision is a hydrological digital twin for the UK and Europe — and
building it is an exercise in assembling the right technical stack in the right order.

Data ingestion and normalisation from hundreds of live gauge networks feeds into model
training pipelines running on multi-GPU nodes. Trained models are deployed via
Triton Inference Server with TensorRT optimisation, supporting multiple
concurrent models and time horizons through a single inference API. Geospatial
acceleration via cuSpatial and cuCIM handles terrain and satellite imagery
processing at scale. NVIDIA Modulus brings physics-informed modelling into the
mix for 1D and 2D hydraulic simulation. And NVIDIA Omniverse provides the
visualisation and digital twin layer — the environment in which all of these
components come together as a living, continuously updated model of how water moves
through the landscape.

"I'm super excited to be on the NVIDIA Inception Program — we're already making the
most of it, training our latest model on multi-GPU nodes whilst running R&D locally
to build the full stack for a nation-scale hydrological digital twin."

— Wayne Gibbins, Founder, FloodWatch

Looking Ahead

We look forward to engaging with the NVIDIA ecosystem and sharing our results with
the community as our models mature.

© 2025 NVIDIA, the NVIDIA logo, CUDA, cuCIM, cuGraph, cuSpatial, Deep Learning Institute, Earth-2, H100, Inception, Modulus, NVLink, Omniverse, RAPIDS, TensorRT, and Triton Inference Server are trademarks and/or registered trademarks of NVIDIA Corporation in the U.S. and other countries. PyTorch is a trademark of The Linux Foundation. Lambda is a trademark of Lambda, Inc.