Topic: Re: [AXIOM] AxiomMemHash, Schnorr Sigs Implemented, APOS 3.0, AXH 2.0 Proposed - page 105. (Read 204903 times)

grin.  I was looking at sp_'s code and seeing if I could fix his bug for him and decided it would be better to first play with the (working) CPU miner before I touched the GPU code.  Hence this speedup.

hehehe ...

maybe i should put the farm up with this also ...

ummm ... no - i dont think so ...

sp is willing to work on the algo in his fork of ccminer - if anyone is willing to donate ( and obviously if they are interested ) ...

#crysx

Well, heck - while all of the crypto mining devs are having fun offering toys for sale, how about:

2 BTC for a 15-20% speed boosted version (on my very large test machine, goes from 312 hash/s to 380 hash/sec).  x86 specific.

Exclusive;  it goes to only one person.  You get any speedups I make in the next day or so while I keep poking at it, though I don't guarantee results past what I've already got running.

PM me first, since I'll only sell this to one person.

AXIOM developers pinged me to give an opinion on the GPU speedup of the AXIOM PoW function.

I wrote the draft of the paper about RandMemoHash back in 2013, after playing with RandMemoHash a bit. I tested it with a SINGLE GPU (that was even older), and could not get any speedup. My paper does not show any proof, not any real world comparison over several GPU models, so it is definitively unfinished. The suggested parameters are not proven to be valid for all existent GPU models (and less for future models). People should never take for granted what an non-peer-reviewed privately-published draft paper says.

I would have liked that Axiom developers would have pinged me to give an update on it before implementing it.

Having said this, I find that a 5x GPU speedup quite good resistance, considering that a GPU is generally better than a CPU in almost every task. CPUs are good at executing large programs, doing paging, protecting memory, doing I/O. Nothing that a PoW can make use of.

GPU memory system is designed for throughput, and it has generally about 6 times the bandwidth available to a CPU. RandMemoHash bottleneck is the DDR/GDDR memory bus, so that speedup can be expected.

Also you must consider power-usage in the equation: that may make it better or worse for GPU-mining.

I don't know if tweaking the parameters would improve it. Maybe using 8 MB instead of 2 MB can prevent the use of some L2 caches in some boards. But the more space you require, the slower it is to verify the PoW. And this is a cat-and-mouse race, since future models will probably increase the size of the cache.

The Axiom developers & community could collect some real world numbers and try to find an improved parameter set (if it is really needed and if there is a better one).

I wish AXIOM users the best luck.

Sergio.

Nice clean consolidation. I like it.

Please dump some AXIOM, need some more to stake properly.