There were questions on CSA math function size / latency. The following estimate came from some of our math library team members. Our expectation at this point is we will likely start out sharing instances of math functions as a single instance, only replicating ("inlining") if the overall graph is small enough and that actually helps performance. (e.g. if you had a loop body with 2 calls to sin(), there would likely be one body by default, with picks/switches as appropriate for the call/return.) Operation Operation est Latency est sqrt64 14 42 sin64 30 15*3 cos64 30 15*3 log64 24 12*3 exp64 14 10*3 sqrt64 11 11*3 invsqrt64 11 11*3 Floating div 64 15 15*3 Floating inverse 13 13*3 floor64 3 (1 with ROUND instruction) 3*3 mod64 5 (2 with ROUND instruction) 5*3 ceil64 3 (1 with ROUND instruction) 3*3 round64 5 5*3 cvt single int 1 2 cvt double int 1 2 cvt single<=>double 1 2 Kent