The newest Nvidia gadgets ("Turing" architecture) are, according to friends who keep up with this sort of thing, pretty much a whole bunch of 4x4 matrix multipliers ("tensor cores") along with some stuff for traversing bounding volume heirarchies to accelerate ray tracing. They claim 16.3 TFlops (single precision) for their top end TITAN RTX, or "T-Rex" card. That looks like about 1 trillion 4x4 matrix-vector multiplies per second. If you're doing 3D graphics, that's 9,000 point transforms per pixel at 1920x1080 resolution and 30 Hz. Only $2500. Holy geez. (Only 24GB memory, which is pretty thin for something that fast.) On Thu, Jan 17, 2019 at 2:23 PM Henry Baker <hbaker1@pipeline.com> wrote:

The new CPU from "Tachyum" (www.tachyum.com) claims to do full 64bit floating point 4x4 matrix multiply- accumulate operations; i.e.,

(R4x4) = (M4x4).(N4x4)+(P4x4)

So we can now do full *quaternion* arithmetic as easily as real & complex arithmetic.

Also, makes block matrix ops a lot easier.

_______________________________________________ math-fun mailing list math-fun@mailman.xmission.com https://mailman.xmission.com/cgi-bin/mailman/listinfo/math-fun