Modules D04, D06 · concept player
An order of magnitude, every step down
Every byte a GPU touches travels through one of a handful of tiers, each roughly ten times slower than the one above it. Slide the message size and watch all eight tiers move at once — the gaps between them are the entire subject.
Time to move this message, by tier (log scale)
Midpoints of D04's own order-of-magnitude ranges, not new specs. D04's table gives each tier as a range ("10s of GB/s", "100s of GB/s", and so on) precisely because a single number would overclaim precision it does not have; this page uses one representative point from each range so the ladder can be drawn. The shape is the citable part — D04 section 2.1: bandwidth drops roughly two orders of magnitude from HBM to the network while latency rises roughly four, which is why small messages suffer so much more across the network than a bandwidth ratio alone suggests.
