That is not true; compare kr7's comment [1]: > On Skylake [...] 32/64-bit floating point multiplication has a 4 cycle latency with a throughput of 2 instructio

That is not true; compare kr7's comment [1]: > On Skylake [...] 32/64-bit floating point multiplication has a 4 cycle latency with a throughput of 2 instructions / cycle. Of course, there are some operations that are very expensive (trigonometric functions, for example), but they're not necessary here, and they're also very expensive on the GPU. [1] https://news.ycombinator.com/item?id=13693749 https://news.ycombinator.com/item?id=13693749