In that 16 years, we got new (and mostly unnecessary) layers of abstraction on what existed previously. Whole single-page-application crowd (eg. Sencha, React...) is busy reinventing things that already exists three layers of abstraction down from the level on which they code.

DirectX was interesting in that age because of exactly the thing in article: you can allocate sound buffer and one surface and view that two COM objects as hardware and not care much about performance implications of the abstraction in between. Once you start drawing to (to borrow X11 terms, as I'm not too familiar with GDI/DirectX) IndexColor surface that is shown on TrueColor display, you get slowdowns in the abstraction layer on the order of "2005 highend Windows NT workstation renders at 2FPS what random office computer from '94 on DOS rendered at 60FPS" (see OpenTTD for example of exactly this performance regression).

For last 25 years hardware acceleration for graphics is readily available for all performance tiers but still, you need to have quite low-level understanding of how the hardware works to get reasonable performance (and as for working-around hardware and driver bugs, we are currently in mostly same situation as 25 years ago and it was better in the meantime).