Cogs and Cache

How to read it

Each row is a milestone: a robot computer put on sale, a module announced, a paper, or a dataset released. The page holds 13 milestones from 15 public sources. A row sits at the date its own source prints, to the month at least; a paper on arXiv is dated by its submission month and is a preprintA preprint is a research paper posted in public, often on arXiv, before a journal or conference has reviewed it, or without any review at all., not reviewed by a journal.

The chart follows the computers that robot makers build on, NVIDIA’s JetsonJetson is NVIDIA’s family of small computer modules for robots and other machines that run SI on board, each with a GPU, a CPU and memory on one module. modules, because they are the line whose makers print memory and bandwidth for every generation. Memory is the size of the module the source names, and bandwidthMemory bandwidth is how fast a processor can read and write its memory, in bytes per second. For SI inference it often sets the speed more than raw compute does. is the peak the maker prints. Both use a log scale, where each gridline step multiplies the value. Where a module came in two sizes, the row shows the larger and says so in its note.

Datasets are charted in trajectoriesA trajectory, in robot learning, is one recorded run of a robot doing a task: its camera frames, joint positions and commands from start to finish., the recorded demonstrations or episodes each paper counts, also on a log scale. A dataset can set a record from the month its paper was released; one that pools earlier datasets counts them all, as its paper does. Sources are the makers, companies and journals themselves; news sites and trade press can point to a milestone, but the row cites the primary page.

What it will not tell you

This is a record of milestones with a dated public source, not a census of robots. A robot whose maker never printed its memory is not here, and the chart follows one family of robot computers, so it shows how far the memory on them has climbed, not what every robot carries. A company’s outlook for future robots is a forecast, not a milestone, so it is left off.

A paper describes the robots, models and flash it tested, not every machine of the same kind. For how flash and memory work underneath, see Flash Foundations and Beyond NAND. For optics for memory, see Photon Fabric; for memory and flash in space and in SI datacenters, Signals; and for the makers’ results each quarter, the Earnings Desk.