This does not produce the exact result you desire, but it may be instructive. Suppose we compute the p-value for different F values. This is the upper tail of the F distribution, so we subtract the cdf from 1. We can do this okay for F=2, but we get zero for F=20 because the cdf is too close to 1 and the difference is less than eps:
>> 1-fcdf(2,6,145584)
ans =
0.0620
>> 1-fcdf(20,6,145584)
ans =
0
We can instead compute the p-value as the lower tail of 1/F and swapping the degrees of freedom. That allows us to achieve values less than eps:
>> fcdf(1/2,145584,6)
ans =
0.0620
>> fcdf(1/20,145584,6)
ans =
1.6677e-23
So if we use this technique to investigate values closer to 272, we find that we can get a non-zero value for F=200 but not F=272:
>> fcdf(1/200,145584,6)
ans =
5.5078e-255
>> fcdf(1/272,145584,6)
ans =
0
Maybe the result is less than realmin which is approximately 2e-308. If I plot the function on a semi-log plot, it looks like the result is close to 1e-340:
semilogy(200:272,fcdf(1./(200:272),145584,6))