liangsu9988/fast-kernels-are-not-fast-pipelines

Redraw the concurrency film with live rate readouts

Make the rate readouts live, the way the robot panes already were

Link the project name in the lead

Rewrite the prose, and give the contents one level per idea

Split the two language films into their chapters, and name the host

Split the engines film into its two engines

Split the concurrency film into its four batch levels

Neutral wording on the vendor comparison

Rewrite the lead around the actual claim, and cut the padding

Drop the scope section

Remove the private-repo link, and stop over-declaring the scope

Drop the kernel-provenance section and the 'built on' framing

Say plainly where the kernels come from, and move the boundary section

Restore the table figures, and lead the guide with the demos

Lead with the demos, and put each one's benchmark under it

A real player: replay, a scrub bar, and 1x / 2x / 4x / 8x

Play the films in real time, with a 2x control

Put the films first, and give each one its own section

Serve the films as VP9 as well, and verify one of them actually decodes

Watch it run: eleven films in the article, as video

figure 3: a before/after table under the bars

side-by-side boundary comparison, aligned engine bars, and the kernels/structures split

every figure opens resolved; captions can no longer lift a bar; drop the stray rules

figure 2 opens resolved; align every bar on one baseline

open on the full picture; tighten every figure for phones

opening figure: the four-step cascade, click by click

boundary figure: row-count selector, and name the eager comparison

interactive article: fast kernels are not fast pipelines

initial commit