Quote from: AxiomCryptocurrency on July 29, 2015, 05:55:30 PM

Quote from: Sergio_Demian_Lerner on July 29, 2015, 05:44:59 PM

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.

Sergio,

Thank you so very much for making a presence in our ann. While we were conducting our research on GPU resistance algorithms the RandMemoHash seemed like the best option available. We're currently assessing the impact of GPU's on our current distribution model. We're warming up to them because they are not nearly as efficient as people initial suspected. The electricity cost is also a major factor in their efficiency as well. Currently, they do not have much of an advantage if at all. Thank you so much Sergio, we look forward to reading more of your draft papers and we look forward to doing great things as a collaborative community utilizing the research that is available to us to improve our crypto currency ecosystem.

Axiom Team

There have been too much work put into this for it to be a scam... but I might be wrong.. Grin