Make Parallella suitable for Litecoin mining

[toterpinguin]

I heard that 1024-core Parallellas would have high development costs and because of them not being suitable for mainstream, it is questionable whether they would be profitable. Now I had the idea that you could tweak your cores to optimize their performance for scrypt-based cryptocoin (Litecoin) mining. Bitcoin mining is not really worth optimizing because of low memory requirements which makes specialized bitcoin mining chips, which already start hitting the market, much faster and more efficient than any general-purpose processor like CPU,GPU or Parallella. But the scrypt algorithm has high (128kb I think) mem requirements which would make it quite suitable to the Epiphany chip.
While not as widespread as bitcoin, people mining Litecoin would certainly buy your devices in huge quantities if they provide better price/performance than existing GPUs. Don't forget to provide software for easy miner setup. Of course, the Epiphany always has to stay general-purpose, but stuff like a rotate instruction and maybe more memory per core would benefit other algorithms too. And miners just are a much bigger user group than the parallel algorithm geeks this was originally meant for.
I think it wouldn't be so hard to tweak your chips into becoming the most efficient for litecoin mining on the market. This would ensure development pays off and therefore maybe help shaping the future of parallel computing!

[9600]

[Hoernchen]

[toterpinguin]

I didn't mean litecoin should be the only thing the Epiphany chip should be optimized for. I just thought you could evaluate how to make a really efficient mining software for the chip and then maybe add some instruction if it doesn't take much die space and gives huge mining performance gains.
By making future Parallella boards the best choice for litecoin miners (I think they easily could be) it would be much easier to pay off your research cost. ASICs for litecoin are not expected anytime soon, probably because their performance gain would not be high compared to current GPUs: most of their core die space is memory, and most of the die space on a specialized ASIC litecoin miner would also be memory. Compare that to bitcoin with its small memory footprint that makes all that memory die space on GPUs wasted.

[solardiz]

[Hoernchen]

[solardiz]

[jar]

I missed this thread until now. Litecoin mining was one of my primary software development goals for this architecture since it (and future designs) fit the scrypt algorithm fairly well. I agree with most of the above analysis. A rotate instruction would cut three operations to one, and multiple 32-bit integer operations per clock would be very beneficial. But this is a RISC architecture, by definition, and adding instructions costs power and die size.

Each Epiphany III core has 32 KB of instruction/scratchpad memory for the scrypt lookup table (LUT). This might be sufficient for a lookup gap of 5 or 6 as you can see in the table below. This would result in a compute performance hit of 3-3.5x. I'm not certain about the Epiphany IV, but if a future core had 64 KB of instruction/scratch, one may be able to implement scrypt with a lookup gap of 3 and a compute cost of only 2x.

Gap | LUT (KB) | Compute cost slowdown
1 | 128 | 1
2 | 64 | 1.5
3 | 42.75 | 2
4 | 32 | 2.5
5 | 25.625 | 3
6 | 21.375 | 3.5
7 | 18.375 | 4
8 | 16 | 4.5

Consider that when dedicated Bitcoin mining hardware was announced by BFL in June-ish 2012, the Bitcoin market capitalization was around $50M. I don't know how many orders BFL has taken to date, but it's probably at that scale.

Currently, the Litecoin market capitalization is around $60M (though this may change significantly in either direction with little notice). One doesn't need the full 1024-core, 128KB Epiphany ASIC to make a Litecoin mining product. Multiple smaller ASICs could be used with no glue logic.

A demonstration of Litecoin mining performance and power efficiency on Parallella and a couple additional tweaks on the cores could go a long way to selling Adapteva cores to people that like the magic internet money and have plenty of USD burning a hole in their pocket. Litecoin could be the killer app for this company even if they don't care about it.

[9600]

[solardiz]