Systems for Machine Learning

3 credits · 60h required · after AIE-204

The hardware reality underneath everything above, and the answer to a question the rest of the programme keeps raising without settling: why do modern architectures look the way they do. They look that way because of memory bandwidth. Attention is not slow because of its arithmetic — the arithmetic is cheap — it is slow because it moves more data than it computes on, and once that is a number you have measured rather than a claim you have read, half the design decisions in a modern model stop being arbitrary.

The lab. Free-tier hosted GPU, and honesty about the constraint. There is no GPU on this laptop and the droplet is a 1 vCPU box that is not a laboratory for anything. The profiling items run on a free Colab or Kaggle session; the arithmetic items — intensity, roofline, communication volume — need no accelerator at all and are done on paper and checked in numpy. Where an item genuinely needs hardware that is not free, it says so and gives a smaller substitute rather than pretending.

0/17 assessed items demonstrated · 0/11 read or watched. The second number is how you close a gap; only the first is evidence that it closed.