Myself and the other members of our team over at BTCFPGA have worked very hard as well over the last couple of days to try and make the ModMiner quad compatible with your Bitstream and we did have success doing this.
Well, this is the first I've heard of this. Is TheSeven the one doing this? He's who I've been emailing back and forth with. As of yesterday he wasn't quite there yet, so I assume this has just happened in the last 24 hours.
It would have been nice if you had sent me a simple email before posting this "open letter" to the forums. I'm quite happy to discuss this stuff in a reasonable way. You and I seem to have lost contact two or three weeks ago although I've been corresponding regularly with TheSeven since then.
We were able to create a driver that would allow a MMQ to connect to your server, receive your Bitstream and we have had them able to mine at over 1Gh/s.
Great news! Your customers should appreciate the 25% performance boost.
with our own testing we are seeing very high amounts of stales from your signing servers,
Well, first off, the signing servers don't send you stales. They send you signed work. Your pool or bitcoind is what declares submitted shares to be stale.
I get under 1% stales, so it can't be the signcryption. If you enable the telnet status monitor (use -Dstatus_port=X and telnet to localhost:X) it will show you the breakdown of how much hashpower is wasted waiting for responses from the signcryption server vs. waiting for the pool vs. work rejected without a longpoll ("unavoidable stales"). You can tune the amount of pipelining used; the default is to always have at least one job pre-signed and ready to load. With this setting I see 0.05% of my hashpower wasted on signcryption delays.
The image below shows a sample screenshot of the status monitor. Each row is a board, each group of three cyan numbers is a chip, each of those three numbers is a ring, and the number is the clock frequency at which the ring is running. If you hit "enter" it will cycle through clock-rate/error-rate/accept-rate on a per-ring basis. The bottom line always displays the cluster-wide statistics. The screenshot below is from my testing cluster which consists of the 12 poorest-performing chips out of the batch (one died):
FWIW choice of pool does matter. I have done 95% of my mining with Eligius, but recently (last four weeks or so) I've found that with EclipseMC and BTCGuild I get under 0.25% stales; I'm not entirely sure why this is. X-Roll-Nonce helps a lot (which is why I favor EclipseMC over BTCGuild).
Also keep in mind that the TML host software will always submit a share, even if it knows in advance that the share is stale (due to a longpoll, a reject of another share with the same prevblock, or because the miner was working on stale work while waiting for the pool to return a new job). So you should expect to see a surge of stales every time a block is found. The "accept" number is what matters when comparing against other implementations. The "stales" number in the screenshot above will fluctuate a LOT over a 5 minute window; for more accurate measurement of stales, use something like -Drecent_history_len=3600000 which will use a 1-hour window.
The mining software is very primitive
Ur…. I really need to resist the urge to get into an argument here. In the aspects that affect performance it is more sophisticated than the other getwork-management (aka "mining host software") systems out there. It carefully tracks each millisecond that a board spends without work to do (or with stale work) and "charges" those milliseconds to the appropriate cause (getwork from pool, signcryption, resetting board due to error, etc) which is where those detailed statistics above come from. The clock calibration is at a 1mhz granularity and the software finds the optimal frequency using logarithmic regression, something no other mining software does. And the mining software produces
nifty charts which are essential for troubleshooting cooling problems (or overvolting, if you're into that, but I don't advise it).
Maybe you're having problems because you've been using it for less than a day? It also might have something to do with the fact that one of the most popular and quite excellent miners "mpbm" happens to be written by the same guy who wrote the TML's ModMiner driver for you.
There have been many reports of excellent resuts on other boards; see the "confirmed" links on
this page.
I will personally guarantee a Donation of 200 BTC from my own personal funds, plus another 200 BTC from the profits of ModMiner Quad sales to increase the profitability for the customers who have purchased them.
I appreciate your generous offer, but unfortunately I must decline.
The TML is already available, today, for your customers to use. There are no commissions until at least 07-Oct.
