Real-time monitoring of compute grid workloads at AMD

Advanced Micro Devices (AMD) is a global semiconductor company that designs high-performance CPUs, GPUs, and AI accelerators. Its internal compute grids support large-scale Electronic Design Automation (EDA) and engineering workloads that require efficient resource scheduling and real-time operational visibility.

AMD operates large internal compute grids supporting EDA workloads across multiple regions, requiring real-time visibility into job execution and resource utilization.

Limited visibility into grid resource usage

AMD’s grid infrastructure runs thousands of jobs across internal systems. Capturing resource usage data such as CPU, DRAM, job timing, and developer activity required a data platform capable of ingesting and querying metrics in real time.

High-volume operational metrics ingestion

The AMD HPC data center runs over 20 million jobs daily, with more than 1 million running concurrently at peak. Each job generates telemetry data used to track resource utilization and scheduling behavior. The monitoring platform needed to continuously ingest high volumes of data while maintaining low-latency access for analysis and troubleshooting.

Scalable infrastructure across multiple regions

As AMD expanded compute workloads for EDA, the monitoring system must scale across multiple data centers and regions. The platform needed to support future growth while maintaining reliable performance and operational simplicity.

AMD deployed Aerospike as the data platform for its grid monitoring application, ingesting and storing job-level telemetry for real-time analysis across its compute infrastructure.

Database for grid monitoring platform

Aerospike serves as the data layer for AMD’s grid computing monitoring application, storing metrics such as CPU utilization, DRAM usage, job start time, and user information associated with compute workloads.

Real-time ingestion of operational metrics

The monitoring system continuously collects and stores telemetry from the grid environment. Aerospike enables rapid ingestion and retrieval of these metrics, allowing infrastructure teams to analyze grid activity and resource usage.

Scalable deployment for expanding grids

The initial deployment runs in a single AMD data center supporting internal workloads. The architecture is designed to scale as AMD expands monitoring capabilities to additional regions and compute environments.

Built on AMD EPYC and Google Cloud infrastructure

Aerospike's multi-threaded, NUMA-aware Hybrid Memory Architecture is designed to fully utilize available CPU and DRAM resources. Running on Google Cloud C4D VMs powered by 5th Gen AMD EPYC processors, Aerospike delivers up to 80% lower infrastructure costs than legacy databases while sustaining the throughput required for high-demand AI and HPC workloads.

“We selected Aerospike because it can handle large volumes of operational data with predictable performance. That reliability allows us to monitor compute jobs across the grid while maintaining the efficiency required for high-demand modern workloads. This also allowed us to offload the query capabilities from the HPC grid scheduler so that the scheduler does the most important job of scheduling and not handling queries.”

Rajdeep Sengupta

AMDApplication & Systems Engineering

Aerospike provides AMD with a scalable data platform for monitoring compute grid activity in real time, improving visibility into resource utilization while supporting expansion across its infrastructure.

Real-time visibility into grid workloads

Infrastructure teams can track operational metrics for compute jobs as they run, helping identify inefficiencies and monitor resource usage across EDA workloads.

Operational insight for job scheduling

The monitoring system provides insight into job execution patterns, enabling teams to analyze how resources are consumed and ensure compute jobs are scheduled efficiently.

Foundation for expanded monitoring infrastructure

With Aerospike deployed in production, AMD plans to extend the monitoring system to additional data centers and regions, supporting larger EDA workloads across its compute grid.

Enabling agentic AI workflows at production speed

The Aerospike and AMD EPYC stack that powers AMD's grid monitoring also underpins broader agentic AI use cases. In a fraud detection workflow demonstrated with Google Gemini and the Agent Development Kit, the combined stack cut investigation time by 90%, from payment event through risk scoring to human analyst decision.