Topic: [CLOSED] Bitmine CoinCraft A1 28nm chip distribution / DIY support - page 5. (Read 81287 times)

I added code to trim my supply and the good news is that it looks like the board is stable at 35GH/s at 0.975V. To get any faster than that I need to get over the max 1.050V my supply is able to put out (to do that I need to disassemble the cooling and change a sense resistor - not terrible but it is a hassle).

A single chip (the board has four) almost works at 1.050V/40GH/s. If I run all four then the supply under load drops to about 1.030V which doesn't work too well.

Need to validate this on more than one board of course

Yeah the Hex8A1 is in the 1.5-1.6W/GH range when running at 8x33GH=264GH and runs a hell of a lot noisier fans than 6x Hex16B do (1800 v 4900rpm).
I opt to run them at 880MHz(4x220)*/910mV = 27GH chip = 1.25W/GH which is just about coolable with 3x Quiet F9 92mm 1800rpm fans.
I'd rather distribute the heat over more units and not have a headache! Bitfury2 now shows another 25% GH improvement too, but it looks like it's going to be short lived as king, by the new 40nm Avalon 3 which promises 0.775W/GH, but still waiting to see real world results on that one.
* The Bitmine clock is 4x the Technobit clock.

This is also what I see. You get to crazy amounts of power usage and heat over 33... I don't think it is safe to run it over that with normal cooling. Even if you use coper coolers I don't think it is enough gain to be worth extra cost of power and price... When you get to about 950 MHz range necessary voltage to run the chips start going up like crazy... From 0,9 at 950 to over 1 at 1050...

EDIT: Looking at your signature. I think thus Hex16B are about as good as 8 chips A1 board... If we look at just a performance... Maybe even batter


Did anyone till now get close to promised numbers? So I know is it my board (and HEX8 too) or just the chip...

My experience with the Technobit Pre-production A1 chips is that they don't fire up until 850mV (Turbo) and then you'll be lucky to see 25GH (Normal).
They're in a top-30%/70%-bottom package so you'll need a heatsink on both sides and the power/temps rise quickly so you'll need a lot of cooling to get anywhere near 33GH. You need to raise the voltage as you raise the clock or you'll get hardware errors but 33GH is doable at 1040MHz(4x260)*/1000mV. Marto even managed to push it to 35GH at 1100MHz(4x275)*/1050mV
* The Bitmine clock is 4x the Technobit clock.

Well that's sad to see… I've still been hopeful that I could get 40 to work someday… It actually does kinda work, I just notice that when I over clock the chip I start dropping nonces (when running a known nonce test case)…


I haven't played with increasing the core voltage beyond 0.85V - still on the list. 1V really is necessary that's going to be a decent chunk of power indeed!

Has anyone been able to get 33GH/s or higher to run?

The current pricing really does need to be adjusted. Several competitors are all coming online very soon at well under $2/GH. I'd really prefer to stick with coin craft as I have a working design and am very happy with the support I've seen (zefir plus free samples). But the current pricing level is going to make it very hard to be profitable for long...

One always null? (Alignment gap to detect breaks in registers, for error-checking, or an unregistered bit for future-use?)

Does it use error-checking-bits? (Like a check-sum or always high/low, or tick flip-flop bit.)

Might be used to differentiate between send/receive data, on the same stream. send=0 receive=1 dictating pass-along, or use, by the chip or the chip-reader. (Thus, addressable only by the hardware for internal use only.)

I have checked the SPI. It shows 16MHZ . So i'm running the chip at 1024MHZ clock.
There is one thing o don't understand in the datasheet.
Here is the register:

https://www.dropbox.com/s/45w9ta19af7y1fl/A1_registry.JPG

It is 48 bit but the numbering goes from 0 to 46. It is correct?
I will program the dividers

Add to that  10%  dead cores on average. Take into account chip price and difficulty. In reality that means around 25% hashing power + 25 %  increased power draw without price reduction? sorry to say it but coincraft will remain in history soon  if price is not reduced significantly
I love my craft boards for sure. I would like to have more but at that price it seems direct loss to me
Seems that great zefir efforts will be wasted once again

Did Bitmine change chips specs?

Wasn't turbo 40GH? Now it is 33GH... But still same price... But I do agree I don't see way to get them to 40GH with normal cooling...

EDIT: Yes they did... And forgot to change it on one page...

http://bitmine.ch/?page_id=863

http://bitmine.ch/?product=coincraft-ai-asic

So they change it down fro about 20%...

EDIT2: And from what I see you can't get 33GH out at 0.85V more like 0.985 to 1 V. Also 1W in Turbo mode is fantasy... More like 1,3 to 1,5W... Hard to say how much is the loss on chip power supply...

Are you guys setting-up the chips so they fire-up all at once, or in time-delayed succession? (To allow power to "restore", before reaching the amperage limits of the voltage-regulators and reference-voltages.) Turning-on all the lights in your house at once, and the AC and the fridge, with the coffee-pot, and the hair-dryer, would trip the main breaker. Even though the total running amperage is not high enough to trip the breaker, the in-rush of start-up amperage draw would trip the breaker. It is like a 150-lb man standing on a scale, then jumping on it. When he jumps, he weighs up to 10x his weight, in momentum, from acceleration. Electricity does a similar thing, as the brown-out from the dip, causes the VR's to kick-in high-mode, delivering 10x more power that it "detected" it needed, but once delivered, that starting resistance is no longer there, and the in-rush is too much amperage. (Or it does a full black-out, as the reference voltage falls out of spec, since all the available power was just absorbed by that temporary short-circuit of all the devices turning on at once.)

I realize the chips auto-run once started, pulling massive power as they run free... However, could that also be addressed by starting them in low-power mode, in addition to firing-up in sequence, then ramping them up to normal mode in sequence also?

One nice feature would be an AUTO-ADJUST feature. Tuning each chip to remain only powerful enough to operate within X-errors. (Which would really be a nice safety feature in the event of chip-decay from overheating. In the event of poor cooling or dead fans or unavoidable restrictions like dust build-up.) Every returned error would count as 1/2 of a potential speed reduction, which would decay after a few minutes. If two errors hit, within that time, the unit drops down a level. Thus, less errors, or none. Which it would attempt to auto-adjust back up, after a few more minutes... Just in-case it only needed a mini-break, or the errors were just a chance-occurrence. (1/3 for the more daring, and 1/4 for those trying to ride the upper-limit... 1/5, 1/6, 1/7, 1/8 for those who may have better management of heat, but purposely want to get that extra few hashes.)

P.S. Due to the speed of the chips... You might also want to hard-code a minimum hash-rate "Share limit". Not even attempting to broadcast diff-16 or lower... For instance... Because it would just flood the output-stream with too many collisions that would appear to be errors, only because they are not arriving completely through the data-lines. Most pools now have auto-adjusting hash-rate diff-levels, and also compensate for returned diff-levels, based on the difficulty actually returned. I am not sure how much diff-16 as a minimum would save you, but diff-128 to diff-256 would be what most suggest for a THs miner, and most auto-adjust to diff-512 anyways. (Diff-16 from a 1THs miner would surely saturate most internet connections with a stream of data that would end-up being buffered anyways.)

I think the naming for post and pre divider were reversed in an earlier version of documentation / software, but basically yes, this should set the sys_clock to 250MHz.

If unsure, you always can double-check by scoping the inter-chip SPI clock, which is sys_clk/64, so at 250MHz you should get ~3.9MHz.

Note: take care to configure your host SPI clock below the inter-chip SPI clock, i.e. in this case configure your RPi SPI interface below 3.9MHz.

Oh, Things are getting more interesting. Thanks.

I have configured the chip at 250MHZ system clock and 16MHz reference clock.
pre_div=2, post_div=4, fb_div=125: 0x88 0x7d 0x21 0x84 0x00 0x00

So the hashing speed is supposed to be 250MHZ* 32 cores = 8000 MH/s

Am i correct?
So maybe i need to reconfigure at a lower system clock? To make it work stable.

Well, that's the standard problem with unstable power blocks: once you feed the chip with work, it starts hashing and with that consuming serious power. When that happens and your DCDC is not capable to provide the required power, the chip resets itself.

You need to scope your supply voltage and the reset line; then you should be able to notice significant voltage drops below critical values (800mV supply, 1.6V reset).

Once the chip resets itself, it gets inaccessible and you need to re-issue a HW-reset and send initialisation command sequence again.


Hope this does not demotivate you, but getting this issue stable was the hardest part with the boards at Bitmine.


Good Luck

We are not basing our wasp firmware / software off cgminer we used bfgminer / bifury as our starting point.

Review our thread for updates and progress on the Wasp & Hive Project.

I'm using raspi driver with one chip.
The chips starts hashing ok but after a while (30s) it stopes.
I have examined the log and i have found that the problem is in the scanwork function.

Here is the log:

We see that the registry values cannot be read.

Bichnellski, do you have cgminer fork for wasps or any board - e.g. 2-chip DIY-board developed by Zefir ?
And what is the progress on this ?

Our 2 chip test A1 Wasps will be housed in the Hive pictured below. Testing of the A1 Wasp board will start on Tuesday.

Hive P0rn. BP-1. Rev0.

I am selling 2 extra boards that I had made at cost ($25). I am in the US, quick shipping.

That needs to be forked from someone actually having that kind of hardware (I don't). The driver that is in cgminer upstream repository is specifically for the CoinCraft Desk modules, every derivative will need to have its dedicated driver.

Can 2-chip cgminer fork with RX buffer patch be uploaded to Bitmine Github ?