CPU computing
General-purpose processors support applications, virtual machines, databases and many inference tasks. Core count alone does not determine useful application performance.
INFRASTRUCTURE
Understand the layers.
Plan the system.
Start with the workload. A useful architecture connects performance requirements with data, operations and facility constraints.
01 / COMPUTE
CPU, accelerated and high-performance systems
Each workload has a different balance of concurrency, memory, latency and throughput.
Accelerated computing
GPUs and other accelerators suit particular parallel workloads. Software compatibility, memory capacity and data movement can matter as much as peak arithmetic capability.
High-performance computing
Simulation and scientific modelling may need fast interconnects, parallel storage and batch scheduling. Benchmark your own workload before committing capacity.
02 / THE COMPLETE SYSTEM
The supporting layers determine the experience.
A server is one part of an operating environment.
Applications & workload
Service-level objectives, software support, model size, data sensitivity.
Scheduling & operations
Orchestration, monitoring, access controls, patching and recovery.
Compute · memory · storage · network
Balance throughput with memory bandwidth, I/O, latency and resilience.
Power · cooling · facility
Validate electrical headroom, rack density, heat removal and location constraints.
03 / DEPLOYMENT
Choose an operating model around constraints.
Location is a business decision as well as a technical choice.
Public cloud
Useful for variable demand and rapid experimentation. Validate service availability, data-transfer charges, quotas and contractual controls.
Private infrastructure
Offers direct control over hardware and operating processes. Requires procurement, capacity management, facilities and lifecycle planning.
Hybrid & edge
Place selected workloads near data or users while connecting to central services. Account for distributed management, synchronisation and recovery.