Agentic HPC, AlphaEvolve, and New Supercomputing Infrastructure at SC26

Computational scientists and HPC engineers already use AI models and agents for ingesting dense scientific literature, optimizing code, and autonomously steering simulations. But taking an agent from suggesting code optimizations to reliably architecting, provisioning, and optimizing production HPC workloads at scale requires more than a general-purpose reasoning model. Without deep infrastructure and domain context, an unguided agent can pair incompatible storage tiers with compute-optimized VMs, overlook placement policies, or stall when diagnosing a multi-node MPI fabric error. Google Cloud’s response, on display at SC26, is to codify its HPC expertise into reusable agent skills that guide agents through cluster setup and optimization, alongside evolutionary agents for algorithm improvement and purpose-built supercomputing infrastructure for HPC-AI convergence.

Cluster Toolkit already lets organizations deploy repeatable HPC environments on Google Cloud through declarative YAML blueprints. Google Cloud is now pairing it with composable agent skills that use self-describing, deterministic tools, validated through automated end-to-end testing and downloadable from the Google Skills and DeepMind Science Skills repositories. Infrastructure skills audit compute and storage quotas first, run rapid canary checks for regional capacity, match the workload to a validated Cluster Toolkit blueprint, and wait for human approval before running gcluster deploy. Workload-specific domain skills then guide agents through configuring OpenFOAM and Siemens STAR-CCM+ fluid dynamics runs, tuning GROMACS and AlphaFold 3 life-sciences pipelines, or launching Devito and Open Porous Media (OPM) subsurface simulations with browser-based 3D visualization. Governance is handled through strict IAM, Kueue quota, and budget guardrails, with full provenance in Cloud Logging and structured troubleshooting skills such as gke-kueue-debugging to resolve stuck queues or node drains automatically.

The second piece is AlphaEvolve, developed by Google DeepMind. It uses a Gemini-powered evolutionary loop to iteratively mutate, benchmark, and optimize code and 3D engineering parameters against target metrics. The AlphaEvolve on Cloud Batch solution moves from a seed algorithm on a laptop to massive parallel experimentation: a single gcluster blueprint provisions the controller environment, and Batch fans out hundreds of candidate evaluations concurrently across single-VM sandboxes or multi-node MPI task groups. Researchers can steer experiments and inspect candidate lineage through hosted Python notebooks, or integrate optimization runs into existing CLI workloads and agent-driven development tools. Evaluating candidates across instances with Cloud RDMA and Google Cloud Managed Lustre, while using Spot VMs or Dynamic Workload Scheduler to balance cost and turnaround time, produces algorithms tuned for real supercomputing fabrics rather than small-scale environments.

Underneath these workflows, Cluster Director provides a unified hardware and orchestration fabric. Classic MPI simulations can run on H4D instances with 5th Gen AMD EPYC processors and Cloud RDMA, including 3 TB high-memory shapes for large EDA and engineering models, while Arm-based C4A instances on Google Axion target performance-per-watt. Coupled physics-AI pipelines and multi-node GPU simulations can use A4 and A4X instances with NVIDIA Blackwell B200 and GB200 NVL72, while newly available A5X bare metal instances are built on the NVIDIA Vera Rubin NVL72 platform and G4 instances with NVIDIA RTX PRO 6000 Blackwell GPUs accelerate CAE, CFD, and 3D visualization. The 8th generation TPUs add TPU 8t for large-scale pre-training, reinforcement learning, and accelerated mathematical modeling and simulation, and TPU 8i for high-throughput, low-latency reasoning and agentic inference; Batch also now supports jobs on single-host TPU slices. On the storage side, Google Cloud Managed Lustre delivers up to 10 TB/s of read throughput, GPUDirect Storage over RDMA, and an 80 PB Dynamic Tier with Cloud Storage Mirroring, optionally paired with Cloud Storage FUSE local SSD caching.

Google Cloud frames agentic HPC not as turning unguided models loose on a terminal, but as combining Gemini’s reasoning with the codified engineering of Cluster Toolkit skills, evolutionary optimization on Batch, and purpose-built hardware. At SC26 in Chicago (November 15-20), the company is showing live demos at Booth #3418 of Antigravity for HPC, the HPC Prompt Playground, and AlphaEvolve for HPC, plus an Advanced Computing Community Half-Day Workshop and an Agentic and Hybrid HPC Birds of a Feather session. The validated templates are available from the Google Skills and DeepMind Science Skills repositories, with Cluster Toolkit and Google Cloud HPC solutions as the starting point.

What’s new in HPC and AI at SC26

View Original