% Tiled f32 matrix multiply (the GPU has no f64): 30 products of
% 2048 x 2048, about half a teraflop of work.
% For an apples-to-apples MATLAB run, also try single precision there:
%   A = rand(n, n, 'single'); B = rand(n, n, 'single');
n = 2048;
reps = 30;
A = rand(n, n);
B = rand(n, n);
C = A * B;   % warm-up
tic;
for k = 1:reps
  C = A * B;
end
t = toc;
fprintf('%.1f GFLOP/s over %d multiplies of %dx%d\n', 2*n^3*reps/1e9/t, reps, n, n);
fprintf('checksum %.2f\n', sum(C(:))/n^2);