LMH status

[chalsall]

Originally Posted by ET_

POV-Ray? I used id a lot in the last 10 years to render scenes and animations...
It was PVM-aware. I guess how could be translated into CUDA...

Close...

Luxrender. Already enabled with distributed rendering and GPU support.

The image you see is a scale (and crop) of a 1920 by 1200 pixel image of a scene generated using Python and Blender, and then rendered by three machines at three different locations, all cooperating via the Internet....

[ET_]

Originally Posted by chalsall

Close...

Luxrender. Already enabled with distributed rendering and GPU support.

The image you see is a scale (and crop) of a 1920 by 1200 pixel image of a scene generated using Python and Blender, and then rendered by three machines at three different locations, all cooperating via the Internet....

Nice work!

So, Blender is GPU-aware... gotta look into it! Thanks.

Luigi

[chalsall]

Originally Posted by ET_

Nice work!

So, Blender is GPU-aware... gotta look into it! Thanks.

Luigi

Nice image.

To be clear, Blender is not yet GPU-aware. But "proof of concept" versions of Luxrender are (using only OpenCL, rather than CUDA, sadly). Be sure to look into Blender 2.5, and the new Luxrender exporter.

[ET_]

Originally Posted by chalsall

Nice image.

To be clear, Blender is not yet GPU-aware. But "proof of concept" versions of Luxrender are (using only OpenCL, rather than CUDA, sadly). Be sure to look into Blender 2.5, and the new Luxrender exporter.

Yep, I come from Blender.org, and noticed v2.54 beta...

I'll dig more into that when I have some more free time ;-)

Luigi

[TheJudger]

Hi,

Originally Posted by mdettweiler

Okay, so using a figure of 62.823648 Ghz-days/day, that means that it will take:

6924 GTX275-days to complete all exponnets to 64 bits
4775 GTX275-days for those exponents below 1M
2499 GTX275-days for those below 100K
1003 GTX275-days for those below 10K
44 GTX275-days for M1061

Anyone with a GPU up to a 44-day M1061 job?

Oh, and we can get all exponents up to 64 bits in a year with 18 GPU crunchers working full time on it.

Originally Posted by ET_

The good news: GTX 460 and higher are nearly 3 times faster than my GTX 275.
The bad news: mfaktc only works with exponents higher than 1,000,000

Luigi

Just a few comments:

I'm pretty sure that the Ghz-days/day ratio is not a constant. It depends on the exponent size, too. Lower exponent might give lower ratios.
Reason(s):
- more pressure on the sieve because the calculation time per candidate on GPU is lower (less iterations needed) => lower SievePrimes to keep the pace with the GPU
- the code needs to be modified for very low exponents (e.g. less precomputation for small exponents)

The lower limit of exponents is not a hard limit. But I didn't check the code for such low exponents. Early versions of mfaktc (proof-of-concept stage, before 0.01) actually failed on exponents below 1000. I think current versions will fail on those, too, but I didn't even tried them.

If someone really want to do those low exponents he/she has to make some changes in the code (removing the check for the exponent size is not enough!).

Oliver

P.S. too bad that the two new GPU-kernels of mfaktc 0.12 (not released yet) only work for factors above 2^64 (up to 2^79 and 2^92). Both are a little bit faster than the 75bit kernel on a sm_2x GPU.

[ET_]

Originally Posted by TheJudger

P.S. too bad that the two new GPU-kernels of mfaktc 0.12 (not released yet) only work for factors above 2^64 (up to 2^79 and 2^92). Both are a little bit faster than the 75bit kernel on a sm_2x GPU.

Not that bad, as long as we can still use version 0.11 on factors lower than 2^64...

Luigi

[TheJudger]

Hi Luigi,

Originally Posted by ET_

Not that bad, as long as we can still use version 0.11 on factors lower than 2^64...

Luigi

you can use 0.12 for factors lower than 2^64, too.
Each kernel has it's own limits, the correct kernel is choosen automatically.

kernels for 0.12
kernel - factor size limitation
71bit_mul24 - FC < 2^71
75bit_mul32 - FC < 2^75
95bit_mul32 - FC < 2^95
barrett79_mul32 - 2<64 < FC < 2^79, single bit level at once
barrett92_mul32 - 2<64 < FC < 2^92, single bit level at once
(FC = Factor Candidate)

Oliver

[petrw1]

Almost exactly 8 months ago at the end of Oct 2010 LMH got it's first 64-65 bit assignments in the 100M range.

If I make a big assumption that the LMH participation rate continues --- and that Monst continues to clear out 700M; Lindee/d et al clear out 900M and GIMPS Visualization continues in other ranges then "you heard it here first" ... LMH will:
- Reach 300M by mid September 2011
- Reach 400M by early October 2011
- Reach 500M by mid October 2011
- Reach 600M by early November 2011
- 700M will be complete to 65 bits by then by Monst
- Reach 800M by early December 2011
- 900M will be completed to 65 bits by then by Lindee/d et al
- Complete all exponents (above 100M) to 65 bits by year end 2011

[Uncwilly]

Originally Posted by petrw1

If I make a big assumption that the LMH participation rate continues --- and that Monst continues to clear out 700M; Lindee/d et al clear out 900M and GIMPS Visualization continues in other ranges then "you heard it here first" ... LMH will:
- Reach 300M by mid September 2011
- Reach 400M by early October 2011

All of the exponents from 330M to 399M are already at 65 bits or above.

[petrw1]

Originally Posted by Uncwilly

All of the exponents from 330M to 399M are already at 65 bits or above.

That's why I only gave 300-400M a couple weeks. Over the 8 months in the 100M range each 10M block took 3-4 weeks.

Mind you, all these predictions become unpredictable because I never know where GIMPS Vis is going next.

After 200M every other range has 30-70% progress and still moving.

[chalsall]

Originally Posted by petrw1

- Complete all exponents (above 100M) to 65 bits by year end 2011

I think you're being a little over-optimistic.

Currently approximately 13,000 exponents are being taken from 64 to 65 "bits" by "official" (read: coordinated by PrimeNet) and unofficial (true "Lone factorers") workers per day.

Even taking into account that factoring becomes faster the larger the exponent, I don't think all exponents above 100M (and below 1B) will be TFed to 65 before mid 2012.

But we will see....