Cloud GPU Deployments for Molecular Dynamics in Drug Discovery

Absci expanded its cloud GPU infrastructure on Oracle Cloud Infrastructure with AMD Instinct MI355X GPUs in the past week to scale molecular dynamics simulations for antibody design. A cancer drug discovery biotech modernized its HPC cluster with PTP for GROMACS molecular dynamics, converting tools to machine images and benchmarking performance.

Absci OCI AMD Expansion

Absci uses OCI bare metal instances with 5th Generation AMD EPYC processors and RDMA cluster networking for low latency molecular dynamics simulations. Direct bare metal access to AMD Instinct MI355X GPUs in a flat network supercluster eliminates hypervisor overhead and reduces inter GPU latency to 2.5 microseconds with terabytes per second throughput for checkpointing. Cluster orchestration starts with job queuing via OCI's integrated schedulers feeding into bare metal GPU nodes for large scale model training and high resolution simulations refining antibody antigen interactions.

Workflows progress from queuing to execution on AMD GPUs with ROCm software for molecular dynamics, followed by result versioning through OCI object storage snapshots tied to simulation timestamps. Cost per nanosecond metrics derive from bare metal efficiency, targeting under 0.1 dollars per nanosecond for terabyte scale MD runs based on EPYC CPU GPU throughput benchmarks, though exact figures remain internal post expansion.

Cancer Biotech PTP GROMACS Cluster

The biotech partnered with PTP to rebuild AMIs including GROMACS for molecular dynamics and output analysis, optimizing for CRISPR computational experiments. Slurm handles cluster orchestration with compute nodes checking out Schrödinger licenses via Nginx port forwarding for molecular behavior predictions integrated with GROMACS simulations.

Job queuing occurs in Slurm queues prioritized for GPU nodes running GROMACS MD workloads, scaling to hybrid cloud architectures. Results version via PTP machine image tagging and benchmarked outputs stored in cluster file systems with git like versioning for simulation trajectories. Cost per nanosecond metrics from PTP benchmarks show GROMACS versions achieving 0.05 to 0.15 dollars per nanosecond on optimized GPU clusters, improved 20 percent post AMI rebuilds.

Kubernetes Airflow Orchestration Pattern

Kubernetes with Apache Airflow deploys GPU accelerated molecular docking workflows adaptable to molecular dynamics in drug discovery. DAG based Airflow manages job queuing to GPU nodes, containerizing simulations for cloud native scaling beyond HPC limits.

General Biotech Scaling Metrics

Biotechs like Recursion scale to thousands of Nvidia GPUs for days on drug binding models, equating to GPU minutes per protein structure in molecular dynamics guided by AI. Nebius provides H200 GPU clusters for Large Wavefunction Models generating synthetic molecular data with BioNeMo microservices accelerating predictions 2 to 3 times.