Absci accelerates AI-driven antibody design on AMD Instinct™ GPUs

AMD and Absci are building on their partnership to accelerate Absci's Origin antibody-design platform on AMD Instinct™ accelerators, running on the open AMD ROCm™ software stack. The AMD-Absci team achieved a 3x performance improvement compared to the production version of Absci’s AbsciBind model, a co-folding-based antibody binding filter, leveraging AlphaFold2-Multimer with optimized Triton kernels.

Origin, Absci’s end-to-end AI platform for target discovery and de novo antibody design, combines high-throughput wet-lab data with deep-learning models in a rapid feedback loop. As Absci scales, training next-generation structural-biology models and serving high-throughput inference, adding AMD Instinct™ accelerators brings a larger-scale platform while unlocking cost savings and avoiding vendor lock-in.

From Joint Benchmarks to 3x Faster Inference 

Scientific applications highly prioritize accurate and consistent results, so teams are cautious about integrating a new accelerator platform. Absci set up an AMD Instinct™ environment to closely match its production stack in Oracle Cloud Infrastructure (OCI). With shared access, engineers across both companies could act as a single team, validating the same workloads. The team then ran Absci's own benchmarking protocols to confirm output quality and computational performance against the existing GPU fleet. Quality metrics were consistently equal in the initial runs, and the team quickly addressed any performance gaps.

"Enabling this model to run on the ROCm stack was straightforward. It could run out-of-the-box after we correctly set up the ROCm environment. The kernel optimizations came as a result of a deeper collaborative analysis. Once AMD provided the kernels, we were able to add them to the model fairly quickly and substantially increase the throughput of our antibody design pipeline. Scoring antibody designs using co-folding is one of the most computationally expensive tasks in our entire workflow”, says Jovan Cejovic, Principal Software Engineer, AI Platform Lead at Absci. 

A significant portion of Absci’s inference time is spent in AbsciBind, which relies on co-folding: inferring a 3D structure of an antibody-antigen complex based on the input sequences. This model uses two performance-critical low-level functions: triangle-attention and triangle-multiplicative-update. As part of AMD’s contributions to OpenFold3 (both Absci and AMD are part of the OpenFold Consortium), they have optimized these triangle operations through a set of Triton kernels. By integrating AMD’s Triton kernels into AbsciBind, the team achieved a 3x speedup over Absci’s production model, while maintaining output quality. The same kernels were also applicable to Absci’s other models using the same triangle operations, such as AbsciDiff.

Impact and Future Collaboration 

Based on the strength of these results, Absci has acquired a new cluster of AMD Instinct™ MI300X GPUs from OCI, giving the team a scaled platform to develop Origin-3, its next-generation structural-biology and antibody-design platform. AMD and Absci continue their partnership, including a planned upgrade to the next-generation AMD Instinct™ accelerators, while extending kernel and other optimizations across the broader deep learning stack and deepening the shared engineering expertise. 

“Working closely with Absci shows us how our GPUs perform in production scientific applications, where numerical tolerances are stricter than in most ML workloads, and their team brings the domain expertise and established validation protocols to test that rigorously. The kernels behind Absci's speedup are already open source and carry over to any model based on the same operations, so what we learned taking them through a full production integration benefits the next team doing it on their own", says Geoffrey Martin-Noble, Principal Member of Technical Staff at AMD.

About Absci

Absci is a data-first AI drug creation company designing differentiated therapeutics using generative AI. Their Integrated Drug Creation platform powers cutting-edge de novo AI models and AI lead optimization models aimed at designing better biologics against difficult-to-drug targets.

_{All performance and/or cost savings claims are provided by the 3rd party organization featured herein and have not been independently verified by AMD. Performance and cost benefits are impacted by a variety of variables. Results herein are specific to such 3rd party organization and may not be typical. GD-181a.}