Why TOPS alone cannot tell you which AI board to buy
Precision, memory and model support can change the meaning of a headline compute figure.
A publication about the hardware behind local AI.
Essays and guides, organised by their coverage period.
Retrospective essays are grouped by the month they cover. They were written and added on November 2, 2026; that is the publication date shown on each article.
Precision, memory and model support can change the meaning of a headline compute figure.
The board is only one line in the bill of materials. Plan the system around the workload.
An add-on accelerator and an integrated developer kit begin from different assumptions.
A short evidence sequence can separate process failure, resource contention and an application that is alive but stuck.
A device without dependable network access needs installation and rollback assets available before the update starts.
Measurement endpoints must follow useful output across host work, device execution and synchronization.
A local assistant needs rules for what enters a request and what happens when it no longer fits.
Treat supply, cable, peripherals and operating mode as separate suspects in a restart investigation.
Useful diagnostics need bounded retention and enough free space for the next recovery operation.
Keep the target environment and conversion recipe attached to an optimized model file.
Resource limits should preserve the processes that keep an edge system observable and recoverable.
Treat static quantization as a data-dependent conversion step with its own acceptance set.
A fair energy comparison needs a defined task, a measurement boundary and a rule for idle time.