Forum Overview :: Rants :: time and progress turn compute problems into I/O problems

[quote name="Ray of Light"]Dividing line for game graphics is ably illustrated by <a href="http://www.anandtech.com/printarticle.aspx?i=3520">AnandTech's recent multi-GPU shootout</a>, wherein <a href=http://images.anandtech.com/graphs/multigpu4wayfeb09_022709070723/18434.png>the very first FPS graph</a> shows a 3-GPU system outperforming a quad. Note also that "doubling time" is for aggregate FLOPS, and lots of things add to that number without improving Quake performance at all: multi-core (>2) dies, super-dense blade enclosures, dirt cheap gigabit ethernet. Some workloads can be smoothly distributed throughout a mountain of processors strung together via onboard ethernet; neural simulation appears to be one, and Quake not. To answer the question you did ask (but oughtn't): My first Q3 system was a P3-550 bought in 1999; Today's QL system is a 2.2 GHz Athlon X2 from <a href="http://www.caltrops.com/pointy.php?action=viewPost&pid=53005">2006</a>. 7 years, 4x improvement in clock speed, 5x improvement in <a href="http://freespace.virgin.net/roy.longbottom/linpack%20results.htm">naive LINPACK</a>, and (if I'm reading this thing right!) a 35x LINPACK improvement after optimizing for presence of SSE. Meanwhile, it scores only 90.6fps on <a href="http://www.quakelive.com/forum/showthread.php?p=99530">some QL timedemo</a> @ 1280x800, compared to <a href="http://episteme.arstechnica.com/eve/forums/a/tpc/f/67909965/m/568099011">about half that</a> on similar Q3 demos @ 1024x768, a 3x improvement after correcting for display size. [/quote]