005
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.
004
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.
003
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.
002
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.
001
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.