Why is half-precision slower than double-precision?

While executing same calculations in MATLAB, speed of double-precision variables is faster than half-precision.

Is this normal? Is there any way to speed up?

Answer 1

[MathWorks Support Team]

This is true.

n = 100;
t1= zeros(1,n);
t2= zeros(1,n);
for i = 1:n

a = ones(10,10,100,100);
b = zeros(10,10,100,100);
a2 = half(a);
b2 = half(b);

tic
temp = plus(a,b);
c = sum(temp(:));
t1(i) = toc;

tic
temp2 = plus(a2,b2);
c2 = sum(temp2(:));
t2(i) = toc;

end

sum(t1)
sum(t2)

ans = 
0.1456

ans=
0.2737

Computations with half-precision data types is slower than those with double precision data types. This is because, unlike double, half-precision is not a native data type in MATLAB and hence, requires additional tweak to do the computations.

Answer 2

[Krishna Bindumadhavan]

To accelerate half precision floating point math, from 20a onwards, you can now generate CUDA code from your MATLAB Algorithm for NVIDIA GPU's with Compute Capability > 5.3 which have native half precision floating point support: https://www.mathworks.com/help/gpucoder/ug/sobel-edge-detection-in-half-precision.html.