Performance of SHA-256 hashing algorithm on Epiphany

[SkeeterSkeeter]

Below is an email I sent to the epiphany they directed me to post here, I do not want to reword it all so I will post verbatim.


Hello! Just recently signed up to receive a parallela ASAP, I have a few questions.

Issue/Question 1:
What I want to try to do is "bitcoin mining" with this board if I can procure one. Bitcoin mining involves doing the SHA-256 hashing algorithm as fast as possible with as low power consumption as possible. I think the parallela could beat out existing technologies.

I am not sure as to the knowledge of the reader, but bitcoin mining started with using CPU cycles, then GPU cycles (much faster to mine on a GPU), then people looked into specific hardware. Since people have developed FPGA solutions, and newly ASIC's (Application Specific Integrated Circuits).

Speed comparison between mining on each hardware, CPU~100 Megahashes per second, GPU 300-900 MHash/s, FPGA 300-1000 MHash/s, and ASIC 4,500-1,500,000 MHash/s. ASIC's being 50x more efficient than FPGA's, about 25x more efficient over GPU's, and about a whole lot >100x more efficient than CPU's.

But these CPU's are pushing top 100 Mhash/s at ~4Ghz and 95W TDP! The epiphany can do 45Ghz at 2W TDP correct? So what types of speed and efficiency could I expect assuming I can write a kernel to do the SHA-256?

If we scale by GHz, (45/4)*100 = 1,125Mhash/s, at 45x the efficiency as before at half the cost (core i7 3570K = $200; terrible estimate).

Do you think this is a good estimate or way of thinking? What might've I missed in my logic?



Issue/Question 2:
So this chip support OpenCL applications. So I figure this might actually be plug and play and I might not have to write a kernel for it. There is software written to take advantage of OpenCL to do the SHA-256 algorithm.

So the board ships with Ubuntu, I figure I would download the software I use and start it up and I hope/figure it would see the epiphany chip as a OpenCL CPU just as any other OpenCL enabled CPU and it would start doing the work the program asks it to do.

Does that sound correct?


If this is correct you might be looking to put a few ASIC developers out of business in the next few years...

[tnt]

I recently considered that as well, however :

As a _very_ rough estimate of the bare minimum: it would take no less than 37 operations per SHA256 round, you need 64 rounds for sha256, so that's a bare minimum or roughly 2500 operation for one hash. You need two hash to check a bitcoin block. That's 5000 operation. Assuming you fill the pipeline 100% of both datapath that's 2500 cycles. Meaning that 16 cores at 600 MHz could do a maximum of 3.84 MHash/s, whish is ridiculously low by today's standard ...

Did I miss something there ?

[timpart]

I'd be a bit more pessimistic than you tnt. Unfortunately the processor doesn't have a rotate instruction, so would probably have to resort to shifts. One processing path is limited in this context as it can't do eor or shift. The adding it has got would be helpful. Not sure if a signed multiply-add by a power of two could be pressed into service as a left shift and combine with previously right shifted half!

Tim

[tnt]

Yes, all my estimations were purposefully optimistic to show that even in that case the hash rate was not good.

[viisual]

What about litecoins and the scrypt algorithm?

[viisual]

I'm not even sure if this will work, but if anyone has a parallella they can try to run this JSFiddle to get an easy answer to this question:

http://jsfiddle.net/pooler/XB5Fg/

[Hoernchen]

...You did not just propose a browser-based javascript benchmark to get a performance number for the cpu in the zynq that has nothing to do with the actual epiphany chip to estimate the performance of a key derivation function on the epiphany?

[viisual]

[ianpatt]

[TigPT]

anyway, it would be interesting to do a simple benchmark