1 min read
Add as a preferred source on Google

AMD and RGS Build Kubernetes Architecture for HPC and AI Workloads

The design combines AMD Instinct GPUs and ROCm software with RGS’ Carbide Platform and RKE2 Government distribution.

Two metal carriers converge on an open GPU compute chassis / TokenPost.ai
Two metal carriers converge on an open GPU compute chassis / TokenPost.ai

AMD and Rancher Government Solutions (RGS) introduced a Kubernetes reference architecture for research and national-security organizations to run high-performance computing (HPC) jobs alongside AI inference services on shared GPU clusters. The design combines Slurm job management with Kubernetes workload scheduling.

AMD announced the architecture Monday, Nov. 16. Its components include AMD Instinct GPUs, ROCm software and the ROCm GPU Operator, alongside RGS’ Carbide Platform and RKE2 Government distribution. RKE2 Government is a hardened Kubernetes distribution, while Carbide automates deployment from GPU drivers through job scheduling.

The design links two scheduling systems. Slurm manages HPC jobs, and Slinky brings Slurm capabilities into Kubernetes. Kueue manages job queues and resource quotas, including when queued workloads can start.

That setup is intended to let batch research jobs share a cluster with long-running AI services, including model inference workloads using vLLM. AMD describes the architecture as production-ready. AMD validated core components for GPU operation and matching workloads with accelerators.

“As workloads grow, organizations need to share GPU resources efficiently while meeting high security standards,” said Chris Reynolds, RGS’ vice president of solution architects.

The announcement includes no measured GPU utilization gains or inference benchmarks.

Simon Yoon

Reporter

Simon Yoon reports on blockchain technology for TokenPost. Send corrections or tips to info@tokenpost.com.

Loading…