Baby Monster

[quaz0r]

good work sir.  its cool to see you guys really getting into it.   

[Dinkydau]

If mandel machine will update to infinite zoom, will you zoom it further? That's would be crazy!

I could try it.

[quaz0r]

have you guys made a collection of the deepest locations people have been to or anything like that?  that would be cool both for testing and exploring.

[Botond Kósa]

The first render at 25600×19200 took slightly longer than a week on my old computer, and this one (16000×12000) about one day on my new computer with i7-5820k.

Congrats for your new computer, with 6 HT cores it seems 2-3 times faster than your old one with 32 cores, right? More and more people are having Haswell CPUs supporting AVX2, I guess I should soon implement the FMA routines.

[3dickulus]

The last frontier? it's an infinite frontier for sure. with each push to deeper and deeper depths entire new realms are available to explore.

Cheers to the trail blazers 

[Dinkydau]

Congrats for your new computer, with 6 HT cores it seems 2-3 times faster than your old one with 32 cores, right? More and more people are having Haswell CPUs supporting AVX2, I guess I should soon implement the FMA routines.

Reference points are calculated 3,1 times faster. I don't have something to compare how good it is at computing pixels without full precision, but I'm pretty sure this one CPU beats the two opteron 6272s that I had. The speed-ups are at the cost of almost half the performance in fractal extreme.

I think AVX2 can be used to speed-up reference calculations. Is that correct?

[Botond Kósa]

Reference points are calculated 3,1 times faster. I don't have something to compare how good it is at computing pixels without full precision, but I'm pretty sure this one CPU beats the two opteron 6272s that I had. The speed-ups are at the cost of almost half the performance in fractal extreme.

That seems reasonable. The i7 CPU offers 3.1x single threaded performance on 6 cores instead of 32, resulting in 3.1 x 6 / 32 = 0.58x multithreaded performance compared to the Opterons. But this is only true for integer code. (References in MM and full precision calculations in MM and Fractal Extreme are calculated using integer numbers.) Per-thread floating point performance on the i7 should be about 4x faster than the Opterons, resulting in more than 2x multithreaded performance.

Could you make a comparison of the two systems by calculating a deep and large image? Floating point throughput is displayed in the Statistics box under GFLOPS.

[Botond Kósa]

I think AVX2 can be used to speed-up reference calculations. Is that correct?

Not really. AVX2 contains the MULX instruction that may simplify the integer multiplications used in reference calculations (details can be found in this article), but I only expect a speed improvement below 5%. But the FMA instructions could speed up floating point calculations by 50-100%.

[Dinkydau]

That's less than I had imagined. On wikipedia AVX2 is described as an "integer version of AVX".

[Botond Kósa]

That's less than I had imagined. On wikipedia AVX2 is described as an "integer version of AVX".

That's right, AVX2 extends all integer SSE instructions to 256-bit wide. But the full precision reference calculations in MM use no SSE instructions, they are based on the 64-bit MUL instruction that multiplies two 64-bit integers and produces a 128-bit result. Unfortunately there is no SSE version of the 64-bit MUL.