012
Checkpointing is the write burst that sizes your file system
A cluster that reads steadily and writes rarely still needs to absorb its entire memory footprint in a few minutes. That burst, not the average, decides the design.
AI infrastructure
GPUs, fabrics, power, and the parts nobody sizes.
011
GPUDirect Storage: what it buys you, and when it does not
Moving data from storage into accelerator memory without a detour through the host. Real gains in specific conditions, and a long list of prerequisites.
010
Inference is not training, and your sizing model probably assumes it is
Two workloads with opposite characteristics, sized with the same reference architecture. The result is a cluster that is wrong in both directions at once.
009
What a GPU node costs before it produces a single token
The accelerators are the headline. Power, cooling, fabric, storage and the eighteen weeks of waiting are the budget.
008
Nobody sizes the fabric, then blames the GPUs
East-west traffic in an AI cluster behaves nothing like the enterprise traffic your network was designed for. The symptom looks like slow accelerators.
007
5G and NFV infrastructure: telco workloads on general-purpose kit
The network function became software running on servers you recognise. The requirements it brings with it are not the ones enterprise virtualization taught you.
006
Seismic interpretation belongs in the data centre, not under a desk
Oil and gas workstations are the strongest remaining case for GPU-backed virtual desktops. The data is already in the data centre and the workstation keeps travelling to meet it.
005
How much accelerator memory a model actually needs
Parameters times bytes per parameter is the start. What catches people is everything else in the memory budget, and it is not small.
004
The GPU utilisation metric is lying to you
A device reporting 100% utilisation may be doing almost nothing useful. The number means something narrower than its name suggests.
003
Sharing accelerators: time-slicing, partitioning, and what each one costs
One device, several consumers. Three mechanisms with genuinely different isolation properties, and choosing by convenience produces surprises.
002
Air, rear door, or direct liquid: the cooling decision is a building decision
Rack density crossed the point where air alone stops working. Choosing what replaces it commits your facility for a decade, and IT is not the deciding party.
001
Sovereign AI infrastructure, minus the press release
The phrase covers four different requirements with four different technical answers. Deciding which one you actually have is most of the work.