AMD Silo AI™ and Bayer Reach Key Milestone in Scaling Foundation Models for Pharmaceutical Research 

Co-development project deploys Bayer's large-scale foundation models for histopathology on AMD Instinct™ GPU accelerators and delivers improved training throughput at scale.

Pharmaceutical research generates vast amounts of high-resolution image data yet turning that data into actionable insight depends on training foundation models that can learn from billions of tissue patches. Bayer is addressing this challenge with the development of its own large vision model for digital histopathology trained in a self-supervised way. To bring this work to production scale, training needs to run reliably on hundreds of GPUs, with predictable performance, efficient hardware utilization, and the flexibility to evolve as new accelerators come to market. In collaboration with AMD Silo AI™, Bayer has reached a key milestone: a fully deployed, tuned training pipeline running on AMD Instinct GPUs with improved recipes that increase throughput compared to Bayer's existing HPC environment.

Deploying a production histopathology model on AMD Instinct GPUs

Bayer's training pipeline was deployed on AMD Instinct GPUs end-to-end. The build environment worked out-of-the-box on the AMD ROCm™ software stack without any code modifications. This seamless deployment is a tangible signal of the maturity of the ROCm software stack and minimizes the risk and costs associated with moving large training workloads onto the new infrastructure.

With the environment in place, the collaboration shifted quickly from enablement to optimization, tackling the system-level questions that determine how a real pharmaceutical R&D workload behaves in production rather than in a vendor benchmark.

“Scaling foundation model training for digital histopathology is as much a systems problem as a modeling one. Working side by side with the AMD Silo AI team, we were able to deploy our DINOv2-based training pipeline on AMD Instinct GPUs and measure performance on the publicly available subset of our training data. Reaching this milestone gives us confidence that AMD hardware is a credible option for the next generation of our research infrastructure,” says Adrian Wolny, Project Lead & Research Scientist at Bayer.

Co-development is at the core of this cooperation. More than providing infrastructure, the AMD Silo AI engineers worked side by side with Bayer to deploy the digital histopathology training stack on AMD Instinct GPUs, instrument the full pipeline with profiling so that performance behavior could be measured rather than assumed, and translate those measurements into concrete training improvements for Bayer's team.

Turning a co-development project into a benchmarked milestone

Bayer and AMD Silo AI focused the milestone on three priorities: deploying Bayer's DinoV2 training on AMD Instinct GPUs end to end, producing trustworthy multi-GPU and multi-node performance data, and feeding the lessons learned back into Bayer's own development workflow. Baselines were established on AMD Instinct MI300X and then extended to MI355X GPUs as the newer generation became available.

Building on this foundation, AMD Silo AI delivered FSDP-enabled training recipes for AMD Instinct MI300X and MI355X GPUs that considerably increased throughput compared to Bayer’s existing HPC environment. Key enablers included memory-efficient attention via xFormers, which worked out of the box on AMD GPUs, and evaluation of different parallelism strategies to identify the best fit for Bayer’s model size and cluster topology.

For Bayer this means a validated, higher-throughput path to train and iterate on its digital histopathology foundation model on AMD Instinct GPUs, expert support proven against the team's real workloads, and a system co-designed for the scale of pharmaceutical research rather than generic benchmarks run in isolation. For AMD Silo AI, it demonstrates the company's competence in taking a customer's existing model code, deploying it cleanly on AMD hardware, and delivering measurable performance improvements in a domain where reliability and reproducibility are non-negotiable.

The collaboration continues, with next steps focused on data scaling experiments that expand training across larger portions of Bayer's digital histopathology dataset, further multi-node scaling on AMD Instinct GPUs, and continued co-development of recipes that support Bayer's longer-term training roadmap.

About Bayer

Bayer is a global enterprise with core competencies in the life science fields of health care and nutrition. Within its Pharmaceuticals division, Bayer invests in advanced AI and foundation models for digital histopathology and drug discovery, with the aim of accelerating research and improving outcomes for patients.