I recently have been developing a project for my students to explore some functionality of the Wolfram Language and have requested that they investigate the timing for several different methods of list generation. Here are the techniques i've asked them to explore:

p1 = 
    Table[First[Table[Prime[n], {n, 1, k}] // AbsoluteTiming], {k, 1, 200000, 1000}], 
    PlotStyle -> Black];
p2 = 
    Table[First[Prime[Range[k]] // AbsoluteTiming], {k, 1, 200000, 1000}], 
    PlotStyle -> Red];
p3 = 
    Table[First[Reap[Do[Sow[Prime[n]], {n, k}]] // AbsoluteTiming], {k, 1, 200000, 1000}], 
    PlotStyle -> Green];
p4 = 
      First[Reap[For[n = 1, n <= k, n++, Sow[Prime[n]]]] // AbsoluteTiming], 
      {k, 1, 200000, 1000}], 
    PlotStyle -> Blue];
p5 = 
    Table[First[Array[Prime[#] &, k] // AbsoluteTiming], {k, 1, 200000,1000}], 
    PlotStyle -> Orange];

Show[{p1, p2, p3, p4, p5}]

The result is the following graph:

Plot of timing as a function of iterations

Is there something special happening around $n=80000$? Looks like the rate of change of cost for creating a list of prime numbers is abruptly changing near this value.

  • 4
    $\begingroup$ Using Prime as test is somewhat atypical. It would provide much more insight using a function that is also vectorized such as f = Sin[N[#]] &;. I guess that Prime shifts from a very cheap algorithm for small primes (maybe a lookup table) to a more general, but also more expensive algorithm. The fact that the performance curves look continuous tells me that this switching is either well-balanced or that simply an additional effort has to be made for primes after the 80k-th one. Btw., Array[Prime, k] will speed up p5 a little. $\endgroup$ Commented Sep 23, 2018 at 21:19
  • 2
    $\begingroup$ @HenrikSchumacher I replaced Prime with Sin and no change in slope is visible. $\endgroup$
    – mikado
    Commented Sep 23, 2018 at 21:29
  • $\begingroup$ @mikado Right. That was my point. Prime does not show how Mathematica works in general. I am not sure whether any real computations are done at all with Prime in this range of numbers; maybe only two different file formats for the storage of primes are used: one with low compression for the first 80000 primes and one with higher compression (and thus slower access) for the next 120000 primes. $\endgroup$ Commented Sep 23, 2018 at 21:38
  • $\begingroup$ Can this be the point where compilation no longer works because the numbers involved become too large to be represent by machine integers? $\endgroup$ Commented Sep 23, 2018 at 21:46
  • $\begingroup$ @SjoerdSmit I don't think so: Mathematica's integers are 64 bit long. N[Developer`$MaxMachineInteger/Prime[80000]] returns almost 10^13. So the machine integer border is far, far away. $\endgroup$ Commented Sep 23, 2018 at 21:57

1 Answer 1


Regarding this statement

Looks like the rate of change of cost for creating a list is abruptly changing near this value.

I have a different opinion. We can use RepeatedTiming to get a somewhat consistent measurement and check the runtime for one single Prime call taking the list creation out of the game

data1 = Table[{i, First[RepeatedTiming[Prime[i]]]}, {i, 1, 100000, 

Mathematica graphics

So we see the same jump, and I believe it is a switch from one algorithm to the next in calculating the primes. On closer inspection, you will find that the position of the jump appears to be precisely at n=78499. Why? I guess the reason is that we reached one million:

(* 999983 *)
(* 1000003 *)

In the documentation we find

Prime and PrimePi use sparse caching and sieving. For large n, the Lagarias–Miller–Odlyzko algorithm for PrimePi is used, based on asymptotic estimates of the density of primes, and is inverted to give Prime.

So let's assume for n<78499 Mathematica uses caching. Then getting a result, mostly depends on how fast the lookup is. Of course, you need to take into account the real smart users, who, when using Prime[n] definitely will request Prime[n+1] (or 2,3,...).

Let us make an experiment. Instead of using RepeatedTiming, we will simply access dn primes and measure the time. So we will touch each prime but we measure in chunks:

With[{dn = 100},
  dataDoForward = 
   Table[{i, First[AbsoluteTiming[Do[Prime[j], {j, i, i + dn}]]]}, {i,
      1, 150000, dn}]
ListLinePlot[dataDoForward, PlotRange -> All]

Mathematica graphics

No surprise here. The same jump. However, what happens if we access the primes in the inner Do in reverse order?

With[{dn = 100},
  dataDoReverse = 
     First[AbsoluteTiming[Do[Prime[j], {j, i + dn, i, -1}]]]}, {i, 1, 
     150000, dn}]
ListLinePlot[dataDoReverse, PlotRange -> All]

Mathematica graphics

It seems the Wolfram Developers are counting on that we are somewhat reasonable people that don't do these stunts. Please pay particular attention to the order of the difference in time. For 100 primes around n=75000 we need under 0.1 microseconds, while the reverse access is almost 1000x slower.

So my conclusion without official information is that primes under 1 million are heavily cached and accessing them loads some larger primes so that the next access is faster. For lager primes, an algorithm is used that really calculates the result. This seems stable in timing, no matter if you access chunks of primes forward or backwards

ListLinePlot[{dataDoForward, dataDoReverse}, 
 PlotRange -> {Automatic, {0, 0.0005}},
 PlotLegends -> {"Forward Access", "Backward Access"}

Mathematica graphics

I cannot explain, why the forward access method is a bit slower at the beginning for primes larger 1 million, but it is possible that some caches need to be cleared. This however is a topic for someone who actually knows what Wolfram is doing under the hood.

  • $\begingroup$ I see the cost of computing Prime is the suspect. I have edited this into the last sentence. $\endgroup$
    – JEM
    Commented Sep 25, 2018 at 19:09

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.