GPUs and data management is important in an AI push, but don't ignore your legacy kit
THE REGISTER EXPLAINER One of the biggest problems facing enterprise AI initiatives is inadequate infrastructure. After buying GPUs and defining data strategies, companies often falter because their existing server infrastructure can't keep pace.
That's because AI workloads demand fundamentally different compute characteristics than virtualized applications. Traditional virtualization is optimized for steady-state workloads, while AI training is bursty and highly resource-intensive. Most enterprises built their server estates between 2015-2020, before AI became a production requirement.
Legacy servers create three problems:
- They're data bottlenecks: Legacy architecture often can't feed GPUs fast enough.
- They're vulnerable to attack: Incumbent security protections were not designed for distributed AI training.
- They're not readily observable: IT teams have no visibility into performance constraints.
Running old hardware looks cheaper until you calculate the opportunity cost of delayed AI projects. Higher operational costs from inefficient power consumption compound the problem, alongside growing security exposure.
Why do multi-generation environments amplify the problem?
Most enterprises run three or four server generations simultaneously, creating no consistent performance baseline. AI workloads get scheduled wherever capacity exists, not where they'll run well. Training jobs that should take hours stretch into days because older servers become bottlenecks.
What does modern AI infrastructure actually need?
Memory bandwidth: This matters more than most realize, especially for inference at scale. Inference workloads are generally memory-bound, not compute-bound. Legacy servers starve GPUs of data. Gen11 and Gen12 platforms engineered for AI-era performance address this gap.
AI-ready security: Distributed AI training means models and data move across infrastructure, expanding the attack surface. Hardening IT architecture for this means banking security into the silicon. HPE Integrated Lights-Out provides silicon-rooted security embedded into the server life cycle.
End-to-end visibility: Operational intelligence across the compute estate is a crucial requirement. IT teams need visibility into what's constraining performance. HPE Compute Ops Management delivers unified, AI-driven management across distributed, multi-vendor server estates.
Where do I begin?
HPE's approach centers on compute platforms designed for AI density alongside traditional workloads. ProLiant Gen11 and Gen12 servers power everything from virtualization to AI and edge computing. Morpheus VM Essentials modernizes virtualized environments while reducing licensing exposure.
More than a mere hardware refresh, this is about gaining the management layer that shows where AI workloads are constrained. Without operational intelligence, teams guess at bottlenecks. Tech Care Services help customers modernize faster and optimize operations.
Once you've chosen your equipment, begin with your highest-value, lowest-effort AI use cases for quick wins, then expand. Identify which AI projects are infrastructure-bottlenecked for immediate ROI. Phased modernization reduces risk and proves value before full estate refresh.
With a solid infrastructure foundation for your AI efforts, you will be able to tackle these projects with confidence.
Sponsored by HPE.