New Multi-Threaded Pi Program - Faster than SuperPi and PiFast

[poke349]

Heres my results for a moderately overclocked AMD 1090t (I opted for cool and quiet on this machine instead of pushing it to the max ):

25,000,000 digits:
6.543 - v0.5.4.9148 (fix 1) x64 SSE4.1 - Sam3 - AMD Phenom II X6 1090T @ 4.06 GHz - 4 GB DDR3 @ 1800 MHz


50,000,000 digits:
13.849 - v0.5.4.9148 (fix 1) x64 SSE4.1 - Sam3 - AMD Phenom II X6 1090T @ 4.06 GHz - 4 GB DDR3 @ 1800 MHz


100,000,000 digits:
29.577 - v0.5.4.9148 (fix 1) x64 SSE4.1 - Sam3 - AMD Phenom II X6 1090T @ 4.06 GHz - 4 GB DDR3 @ 1800 MHz


250,000,000 digits:
83.152 - v0.5.4.9148 (fix 1) x64 SSE4.1 - Sam3 - AMD Phenom II X6 1090T @ 4.06 GHz - 4 GB DDR3 @ 1800 MHz


500,000,000 digits:
178.526 - v0.5.4.9148 (fix 1) x64 SSE4.1 - Sam3 - AMD Phenom II X6 1090T @ 4.06 GHz - 4 GB DDR3 @ 1800 MHz

Updated
Cool and quiet at 4 GHz is pretty impressive. I can't get my i7 rig to do that past 3.5 GHz...



btw. Although I don't plan on making any major improvements to the program for a while, I do feel the need to overhaul the entire digit viewer/writer.

So after the 5 trillion digits of Pi computation, Shigeru Kondo sent me 3 hard drives containing all the digits.
That was actually a few weeks ago, but I didn't get the time until now to try and "process" all those digits.

By "process", I wanted to:
1. Split them up into 5000 files of 1 billion digits each.
2. Re-compress them.
3. And launch the mother of all torrents* - seeded using my university's connection.

*I don't expect anyone to complete the download - but I do want people to be able to get small portions of the digits.


Then I realized how horrifically slow the digit viewer is at processing digits...
It decompresses at like 10 million digits/sec per thread on my workstation... unacceptable... lol

So for v0.5.5, I'm gonna completely rewrite the entire digit viewer as well as all the digit-processing code.
This time I'm gonna keep the code separate from y-cruncher so that I can completely open source the digit viewer.

I noticed that on the many-core machines, the program spends a disproportionate amount of time formatting the digits and writing them to disk - all of which is single-threaded.
Guess I'm gonna try to fix that.