AMD and biotechnology company Absci say optimised AMD GPU kernels cut the time needed for a key AI antibody-design task by threefold, while maintaining output quality in initial tests. Absci has since acquired an AMD Instinct MI300X cluster through Oracle Cloud Infrastructure to scale the work.
AMD Watch analysis
What happened
The optimisation targets triangle-attention and triangle-multiplicative-update operations in AbsciBind, a model used to assess antibody binding. AMD says Triton kernels developed through its partnership with Absci brought the model to three times the performance of Absci’s production version. The same kernels may also apply to Absci’s other models that use those operations.
Absci ran its own benchmarking protocols against its existing GPU fleet. AMD says the initial runs produced consistently equal quality metrics, and that the kernels are open source. The company’s post describes a customer-specific result, not an independent benchmark or a general performance guarantee.
Why it matters
AI drug-design systems rely on computationally demanding models to assess candidate molecules. Making one of those steps faster could help Absci run more of its antibody-design work, while the reported quality checks address the obvious concern that speed-ups might come at the expense of useful output.
The development also offers a concrete example of a customer adopting AMD accelerators for a production scientific workload. The performance and quality claims come from AMD’s account of joint work with Absci, and should be read in that light.
Our read
This is a more useful signal than a chipmaker’s abstract promise: a named customer says it has bought a new cluster after testing a specific workload, and the reported kernel improvements are available to other models using the same operations. The next test is whether those gains hold up as Absci scales the system and moves towards its next-generation Origin-3 platform.
What to watch
- Whether Absci reports results from the larger MI300X cluster.
- Whether other models using the same triangle operations benefit from the open-source kernels.
- How the work develops as Absci builds Origin-3.
Discussion spark: For scientific AI workloads, which matters more when choosing an accelerator: faster performance on a real production task, or independently reproducible benchmarks across a wider range of models?
Sources and evidence
- Absci accelerates AI-driven antibody design on AMD Instinct™ GPUs – AMD (19 October 2026, 06:47 UTC)
Independent WittyWires tracker for public updates about AMD. Not affiliated with or endorsed by AMD; this is not an official account.