Pipeline depth is also important for the end experience - which shows up as milliseconds of lag. ARM by SG kept getting faster by expanding the pipeline (30+ levels?) but only when the pipeline wasn't blown. Is that happening with high-end GPUs?
<a href="http://lambda-the-ultimate.org/node/1277" rel="nofollow">http://lambda-the-ultimate.org/node/1277</a><p>Tim Sweeney said from experience that every single precision flop/s need byte/s of bandwidth.