Don't use less than 100 iterations for a benchmark, results from 3 iterations are not reliable.
On 7970 without OC you should get about 500col/m. 
Generally the more memory the better should be the result. 

Apparently the RX series cards support buffers bigger than 512MB. That makes the two buffer hack obsolete (ie. one buffer 1GB, or even 2GB version should be a few percents faster). 

Hi. This is an opencl miner. More specifically, it's a miner without network code. 
To compile, you'll need: Visual Studio 2010, AMD APP SDK, Boost, OpenSSL (or an alternative implementation of SHA512). The actual collision pair are computed to a vector on line 120 in momentum_gpu.cpp, if you know c++ you can take it from there. 
I was mining on this code by a dll hack - I compiled a dll with "mine" function and added LoadLibrary/GetProcAddr to first beer's miner. I'm not releasing that code though because it's an ugly hack.

https://github.com/iruu/protoshares_momentum_opencl

Licensed under GPL 2.0. It's the same code that was in the benchmark I gave to some people on irc.

Initially (in November) I was planning and creating a pool for this, but something completely unrelated came up and I wasn't able to finish this. Sorry if you were waiting, especially if you helped me benchmark the code. 

There's an exe in Release folder, you can use to benchmark. 

Why release now? Because someone released CUDA miner. A closed, proprietary language! CUDA should die.

EDIT: When benchmarking use at least 100 iterations for reliable results.

-

If you find it useful (or derived miner, which is very easy to do and probably will appear fast), need to exchange btc or like to trade btc on margin and you don't already have an account on bitfinex (cheapest btc exchange), during registration use my bitfinex referral code IiDoeUOa3i. A sort of a tip which doesn't cost you anything.

Sure thing you can get great results with 128-bit registers and vector rotates and shifts in one cycle. But GPU has just so much more ALUs it kind of compensates. Still, registers are 32-bit and there is no SHLD/SHRD to implement 64-bit ops nicely, so they are pretty much a showstopper.

I would say it differently - cpu compensates (VERY well) for less parallelism with vector instructions. Do you agree?
32 bit registers don't make it automatically slower because they're 32 bit, there's just not enough free registers left for all possible waves. Do you have 100% kernel occupancy on all your kernels? I don't think so. SHLD/SHRD isn't going to change it.

@reorder
http://www.intel.com/content/www/us/en/intelligent-systems/intel-technology/fast-sha512-implementations-ia-processors-paper.html

8.59cycles/bytes for sha512 on large data buffer. Thanks to AVX. It's ridiculous.

Taking into account cost of capital, expected roi and other things, gpus obviously win, but that's quite another point.
(nobody sane is going to buy any hardware for mining this coin)

most you could achieve through your attempted attack is to double spend your own PTS

The "most"? Two-three double spends with 10k pts are going to net upwards of $100k, if someone at the exchange doesn't stop depositing manually... it's enough to destroy the coin by destroying confidence. I think btc38 standing bids can support such volume. They would definitely delist pts after such incident.

In the most optimistic (for the attacker) case, you could wipe all bids using just a few thousand pts by a dozen large double spends. I think btc38 only waits for one confirmation which makes it easier. Tight operation, the way heists are supposed to be

Right now a 51% attack on the network would cost about $30,000 per day.

So even by your calculation, it's only $1250 per hour, long enough for anything.

Even that is a cloud price, not-that-optimized gpu miner for 7970 gets 530cpm, which takes about 200W - $0.02 per hour in some places. Suddenly the number gets much smaller...

The vesting period does not reduce the number of miners

Because you say so... all restrictions are going to reduce the number of miners.

Original client worked this way:
pblock->nNonce=pblock->nNonce+1;
testHash=pblock->CalculateBestBirthdayHash(pindexPrev, outdated);
nHashesDone++;
nHashesDone/time = original meter

So cpm is inflation, because as you said, there's less than one collision pair per work.

Not that it really matters now

Is that actually HPM, as in hashes per minute, or collisions per minute (which coyote miner prints as hpm for some reason)?

Coyote: collisions per minute reported as hpm (at least on github which is 12 days old! - the binary probably multiplies this by 2 also, otherwise coyote miner would be two times slower)
Beeer and ypool: collisions per minute * 2 reported as col/m

Which one is it?

edit:
http://bitsharestalk.org/index.php?topic=965.0
Ok, so 520HPM is impossible on amazon instance. So it's actually col/m*2 for all public miners.

So unless you specifically modified a miner to report true hpm you're getting 500 col/m. Did you?