Tractable

In Property & Casualty insurance, claims cycle time is a primary driver of customer satisfaction and loss ratio performance. Tractable's computer vision platform automates accident and disaster damage assessment for insurers, but the model must be continuously retrained to account for regional vehicle fleets, local repair labor costs, and evolving damage patterns across each new market. With expansion into multiple geographies, each requiring its own model variant, Tractable's research team was bottlenecked by GPU-based training infrastructure that limited how many experiments could be run per day. Slower iteration meant delayed model improvements, longer time-to-market for new insurance clients, and reduced ability to stay competitive in a rapidly advancing AI landscape.

Tractable migrated their AI training workloads to Graphcore's IPU-POD system — a processor architecture purpose-built for the irregular, fine-grained parallelism characteristic of deep learning research. Unlike GPUs, the IPU's bulk synchronous parallel model gave Tractable's researchers greater architectural freedom to explore novel computer vision model designs that would have been impractical or prohibitively slow on standard GPU clusters. The deployment was facilitated through cloud provider iomart and hardware partner Boston Limited, enabling Tractable to access IPU capacity without overhauling their internal infrastructure. Computer vision models for damage assessment — which must classify damage severity, part identification, and repair cost from photographic inputs — are compute-intensive during training, making the IPU's parallel execution particularly well-suited to this workload.

Tractable achieved a 5X improvement in AI training speed compared to their previous GPU infrastructure, and researchers were able to run up to 5X more experiments within the same timeframe. The compounding effect of faster iteration was significant: unsuccessful model architectures could be discarded sooner, and promising approaches reached production-ready quality in a fraction of the prior development time. Key outcomes include:

• 5X faster model training cycles vs. prior GPU setup
• 5X increase in research experiments per development cycle
• Faster deployment of improved models to insurance carrier clients
• Expanded architectural exploration — researchers could test model designs previously ruled out due to GPU compute constraints

• Compute infrastructure is a strategic R&D asset, not just an IT cost — a 5X training speedup compounds across hundreds of model iterations, yielding outsized quality gains over time.
• Purpose-built AI hardware can expand the model design space, unlocking architectures that GPU memory and parallelism constraints would otherwise discourage.
• For AI companies expanding across markets, faster retraining cycles directly shorten geographic go-to-market timelines — infrastructure investment has a product velocity payoff.
• Evaluate AI accelerator hardware not just on raw throughput, but on researcher iteration rate — the number of experiments per unit time is often the binding constraint in applied ML development.