Topic: New Official AMT Thread - page 90. (Read 149472 times)

So........
It's been awhile since any sort of update from AMT.. Have emailed them plenty of times with no responses.... Has anyone got through to them? Still waiting on my machines........
Thanks!

Cool. Wish I had my miner...  'Course I'd like to fly and spit diamonds as well.  So far seems as likely

Doesn't the backplane have a coms chip on it for the GPIO/hash cards coms?

Quick update. I got the raspi hooked up with the pidora distro....backplane not detected. I need to figure that one out. Might just mean its a SPI driver issue I have to sort out. I will post more tomorrow if I solve the issue.

I would be using "Lexan", the brand-name... Unless I can find another source.

Managed to work-out a shroud that fits perfectly into a 12" x 24" pre-cut piece, perfect shape for the 5-card setup. Seems to fit nice and snug. Just have to remove the back-plane, slide it in, and insert the back-plane back into place.

Ordered more Nichrome wire and another LED-dimmer for a controller, to bend it better. Blew my last strand of wire a few weeks ago. Left it on a few hours too long.

Ordering some nice thick Lexan to build two separate custom cases, and some thinner sheets for shrouds. Just need to find some 20-pin ribbon-cables and get some "U" channel to support the cards. Getting rid of that whole metal frame. Was going to chop it into two halves, but it is setup so awkward. Now all I need is more individual cards!

Reduced the whole 5-card setup into a 16" * 7.5" * 14.5" (deep) setup... smaller than my home PC. xD

Going to make it smaller when I split it into 2 units of 3 and 3, and chop the large heat-sinks down by half, snugly packing them together. (The 24-pin ribbon-cables will give me the ability to move them closer together.) The excess length of the backplane just extends over the PSU, since 2 connections are not being used.) I got me a little GH-Cube. Though, adding the ribbon-cables will make it about 9.5" * 9.5" * 14.5" (deep). Ras-Pi also relocated to being above the boards, with the back-plane and ribbon-connectors.

Polycarbonate (Lexan is a registered TM) great choice. Not only good temp before warping but also unlike acrylic it is hard to shatter.

On those mysterious 10k resistors - def a mistake on them. No way in hell are they right. Even from a size standpoint, those would be 2w resistors or more and @12v 10k would dissipate a max of 14mw. And no, no polarity. With resistors the band is just for vision systems on the pick-and-place to know how to verify the number faster. Assuming it uses one that is...

On the PSU switching and PSU in general, is the 12vdc only used to power the hashing cards? If so, ja using the power_OK line is a good idea. Better yet (and cheaper) - how about a lil 5vdc supply for the pi and whatever else needs it and a 1U server PSU for the cards? Leave the Pi on and use the PSU remote on/off to switch it.

For my Ant pharm I use ones from Amazon. $51 for 1,200w HP server psu (runs 2 OC'd s1's pulling 850w per-pair) and @ >95% efficiency they are better and smaller than any PC supply.

Sounds like with better attention to assembly and thermals the AMT/BMch rigs can do fine. The SPI chain thing scare me a bit though. From what I gathered per that SPI coms wiki it doesn't do ring config, just serial chain and at best a star & chain config if all 4 master lines are used. No matter what, all I/O goes through the 1st (master) chip..

Local decoupling caps is but one way for best-practice. Note at least one other major difference:

The bitmine/AMT boards use 12? Vcore supplies (I swear somewhere I have a pic showing 6) that from I can tell are all feeding one very large set of power planes. Unless the Vdd and return planes are incredibly thick there will be large local circulating currents. There will be interaction between the chips ergo the pretty much mandatory need for the larger 6uf local caps. to reduce that problem. The myrid of tiny local 100nf caps. surrounding the A1's take care of the dips & rebound spikes cause by the logic switching. ISa: are the power planes segregated at all?

Now from what little I've seen of it the TechnoBit version only addressed the thermal needs by making it easier to clomp a single large heatsink to the top. But they kept the same idea of the paralleled regulators feeding everything.

Sorry Isa but I gotta say bad idea. For one, if there is a major fault either the whole board goes down if like from a screw shorting things (a bit more care building and testing takes care of that barring something shaking loose in shipping). Or, rather more serious from liability standpoint, because by necessity that very large single source packs a lot of punch if a chip shorts, well we've seen the results of that. Burnt boards with hopes the PSU shuts down before fire.

The new InnoSilicon board addresses those points. They kept the slew of tiny low value caps around the chips and yes have the needed larger caps out at the edges of the sectors leaving access to the chips for the top heatsink.

However. Note 2 things:
1. Power wise the ASICs are actually in sets of 2. Each ASIC pair has its own set of regulator FET's and buck inductorss. Not 12 feeding 8 chips. If a chip goes kerfluey hopefully you just lose the one or at worst the set. Also as you mentioned they may have split up the SPI chain as well Good idea. Again at worst you only loose 2 chips.

2. Those power planes appear to be gold flashed/plated. Between splitting up the supplies and the gold the circulating currents I spoke of are taken care of. Good choice there and not really *that* expensive. Depending on how thick maybe 5-cents per board tops. Electrically silver would be best but with all that aluminum in close proximity can cause corrosion problems unless passivated (even more cost).

3. Goes fuzzy before zooming in much but looking at some of the holes for heatsink screws the board looks pretty thick. Evey with 4 layers too thick for being made of FR4. Tlam or such perhaps?

Given all that attention I'd like to see the contact side of their top sink... Bet it has copper riser plugs pressed into the aluminum.

AMT, I suggest that you run these for a full 24-hours, after assembly... Not just each card on a bench-test... actual assembled units...

After you have given them a once-over, for quality control. Looking-out for those long screws, which seem to be drilling into one of the mounted boards. Also looking for backwards mounted fans. (All fans should be mounted so you DO NOT see the sticker on the fan.)

(Drill holes in a metal jig, and cut all the screws to length with a die-grinder or cut-off wheel... Even tin-snips! They are all waaay too long. Better yet, get nylon screws, or push-in clips... or use cheap "L" trim to hold all fans in place, mounted with velcro.)

Then, wait 2 hours, with the unit off... and power it back up, running it for another 24 hours. That allows the components to burn-in, settle-down, and any post-issues to crop-up. Issues that are seeming to crop-up after bench-testing, then assembling, then shipping.

Also, get a thermal imager. You can easily see the problem components, before they get out of control, on the bench. Just print a photo of a "normal view", and use that to compare the view in your imager. You can set that up with a simple assembly-line rig, and a cheap 300w PSU. You don't need a HD thermal imager like I have. Just a cheap $300 imager with a 20x40 screen would suffice, once mounted to a testing-slide. You can get decent used HD thermal imagers on ebay, for about $600. Mine cost over $3000. It was overkill for my purposes, but worth it.

The heat-sinks are adequate, with shrouding. Might want to invest in some cheap high-temp plastic (Thin lexan would work, and is cheap 155c melting temp.). Cover the holes in the back of the tower, where the cards are. That, or pull the shroud up to the heat-sinks on that side, so air comes IN through the holes, and is guided out to the big fan. Same for the front-side of the tower, and the side with the backplane. As it is now, hot air is just blowing all over inside the case, and some air is being sucked-in, from the back, and instantly pulled right out, by the larger fan on top.

A simple "U" shaped shroud. (I actually use cardboard) Wrapped around the three exposed sides, to create that air-guide/channel, helps dramatically, to ensure cool operation. Running at 780w, with only 3 cards, the exhaust on my unit is barely even warm. Though, the small copper heat-sinks are the hottest component, they are adequate, until they are not. Working on a spring-tension to place between the copper sinks, and the larger heat-sinks on the opposite side. (On the last card, where the copper sinks are exposed without a matching large heat-sink, to get pressure from... I am adding a metal plate as a shroud. That will keep better pressure against the copper, and allow some heat to dissipate through the spring-tension, which also increases the surface area.)

Other than that... These things get an A- grade in my book. Would have been an A if there was a power-switch, jumpered to the PSU connection, or a splice-kit switch, which splices to the power-ready line, so we can safely turn-off the power without power-cycling the PSU by removal of the plug. Since, not all PSU's have a separate power switch, and not all power switches operate at "Safe" levels, as they were intended for emergency power killing when the computers power button fails.

P.S. The large heat-sink is twice as large as it needs to be. The fins could be half the height, and the backplane junctions could be spaced a lot closer together, if that was done. Though, ribbon-connectors would have been better, to the cards. You could also reduce the airflow to six fans, for six cards. Not to mention, setting it up for two PSU's, you would save about 50% in PSU costs. Six cards can easily run off two 800W PSU's, instead of one 1500W PSU. There would be ample room for two PSU's, and that large fan would not be required at all. The PSU's could safely exhaust or be intakes, for 100% of the required air-flow. (Being intakes, would be preferred.) Once I get another heat-sink, I will build a unit as described, to show you. xD. Got a spare heat-sink?

LEXAN 1/32" * 24" * 48" = $10.00 in bulk (Might get down to $7.00) This is enough to do 4x units. That is roughly $2.50 a piece.

Anyone want me to make a shroud, let me know. I'll make them for $10.00 shipped. I won't be getting bulk prices, obviously. I'll be getting home-depot prices. lol.

EDIT: Screw that, Home-Depot charges $67.00 for a $10.00 sheet only 24" * 36"... What a rip-off. lol. I'll get them online. Found a nice seto of assorted springs for $5.00 that may work for adding tension to the copper heat-sinks. Looking for a better alternative, spring-steel "U" mounts.

The COUT's connections are a mix of CAPs and Resistors, it seems. However, though the R+ labeled SMD seems to be a resistor, it has a polarity? Might be a cap/resistor combo. That, or the polarity indicator is just redundant. (This is on all my boards that function without issue. Except the one where the A1 chip is not mounted flush, and has thus, overheated.)

I suspect that the short across COUT46, by a screw from the fans, has just shorted the card, enough to keep it from initializing.

The board with the A1 that was not mounted flush, I assume had "shut down" due to the internal over-heating, once it had gotten to that point. In any event, the failure of that single chip, has led me to believe that the through-connections, the serial in/out on each chip, demands that the chips are chained. Failure of one, would, in some cases, stop the whole board from responding.

Removal of the chip, would seem to imply that a jumper be soldered, to bridge the through-connections, where they would normally be communicating "through the chip", to get to the other chips down the line. However, without the jumper, it seems like it should be possible to communicate with the chips up to that removed chip.

The work, I assume is coming in through the U8 chip, and returning back to the Ras-Pi, through the U5 serial linked chip. U8 is top-right, by the four other chips, I assume setting-up the distributed workload through U1, U2, U3, U4. The U5 chip is at the end of the serial connections from the chips, going directly back to the Ras-Pi, which is why I assume it is returning work/solutions/shares directly back, to be processed for diff-targets, to determine if they are solutions or just shares.

However, I could have that backwards too... U8 might be sending the work to the chips, and U5 may be the return of processed data. However, I am about 90% sure I have the previous paragraph correct this time. xD

The chips are sort-of isolated in groups of four, by two separate clock-timers or frequency controllers. Chips U60 and U61 are tied to four independent 1-wire connections, which are tied to a chip that has the clock-crystal.

All the circuits in the center, seem to be just the voltage regulation. Odd that they all seem to be built as individual regulation units. Might be why the new versions just made one whole power-regulator that was simply more robust, as opposed to 8 individual mini-regulators. There is power-isolation, as the chip on mine, which overheated, had a matching voltage-regulator which was also overheating.

I wanted to remove the caps and resistors from that board, which had the non-flush overheating A1 chip, and place them on the other board with the previously shorted connection... However, I had to craft a special soldering head to desolder the SMD components. That is taking longer than expected. lol. It was easier to do that, then to attempt to rework the non-flush A1, which I assume is actually burned-out and shorted inside anyways.

I am not sure how easy it would be to jumper/bridge the small connections for the data-lines... However, there are tiny inline (resistors or caps, can't read the numbers on them)... in the two center datalines, of the four. Not sure if that would impact the integrity of the data transfer. This is a question only AMT or BITMINE or TECHNOBIT can answer. Again, I don't have the schematics, and have not taken a strong look at the pinouts on the A1's in a while. I had no intentions to EVER get this deep into the design process.

The A2 board does have a few sets of caps and resistors, of a similar form-factor, just outside the heat-mounting area. There also seems to be a lot less micro-caps for filtering, directly around the chips. Roughly about half. However I do not see any crystal clock components... They may have settled for on-chip clock controls with self-oscillators. The controller chip also looks more like an ATMEGA-32 or ATMEGA-64, or a RISC processor. I would assume RISC, due to it being cheaper, and having more sister-components.

Ordered a new PSU for my computer, so I can rape the PSU from that, since it is way overpowered for my PC, but ideal to run the last two card with heat-sinks, that I have. Once I transfer the heat-sinks to the other cards that came without heatsinks. Should be here by tomorrow. Then, hopefully I will have almost 1GHs total. 580GHs + ~400GHs = ~980GHs (Not including errors, which, oddly are counted towards the GHs output, even though they were not valid, thus, not a hash. Makes determining the actual output a pain in the ass. Always says 580GHs, but when errors start, and get worse, the output is actually only about 480GHs in the pool. Just requires a reboot, and the errors go away for another day. xD The program needs a frequency tuning area. Though you SET a specific frequency, the output is not "calibrated", thus, it is ballpark. Sometimes you have to tell it a higher or lower frequency to operate, to be that actual frequency. That is how you tune cards. That is something that 7970's demand, in order to operate at ideal speeds. Not sure why that was just reduced to "Power-save", "Normal" and "Turbo"... Those should be user-settable. However, limited to the MFG MIN and MAX, for obvious reasons. xD, Guess I just gotta PuTTy to get it done. )

Here is the A2 board for Innosilicon that has the same electrical characteristics at the A1:



I don't see any large capacitors near the A2 chips.

I just realized one difference though,  notice that it is clustered 4 at a time.   This may be hinting at an issue that the SPI daisy chain should be max 4 in lenght and not 8 as in the original specs.

You can underclock through software. ISAWHIM is more familiar with what range we have. But, won't help with this. We are concerned with the rise/fall times of the switching load - not its overall freq. It is a square wave so the transition times will be the same. If one has boards with the resistors in place of caps the only solution is to remove the resistors and put in caps.

Actually it takes the right kind of heatsink for proper efficiency too...thinner but more numerous fins can work quite well. I have passively cooled Dual XEONs. I have intake and outake in the cases. To some degree you are correct however. Larger heatsinks. BUT if IMET or whoever made the heatsinks had made smaller thinner heatsinks and better milled they would have worked well, as it is they look like shit with all sorts of grooves that have room for air pockets. And I am referring to the part contacting the chip. While thermal grease will help with that, its only so much. Milling needs to be done right. I might lap down to as close to a mirror finish on one or two of the heatinks with a before and after test on one of my 3 working cards at some point if the cards don't crap out on me first. I might localize it to the specific chip areas as those are some big ass heatsinks to do the whole thing to.

But if the heatsinks had thinner more numerous fins and even a smaller size (looking at the bitmain s2 asic boards heatink) they could manage much better. Less heat to hold and transfer off. Copper would be ideal but very expensive just due to volume alone. Or even a heatpipe design to transfer heat off effectively.

The overall design of the DIY kit as I got it was lacking in one area. There was the "small heatsinks" side of the card on one card that got no air. I had to cool that using a separate fan. It got no air and inevitably also failed. Altho that could have been for any number of reasons.

Okay... so if you think its a problem with switching speed... why don't we just plop in a slower oscillator?  You know,  just to see if it works.

Don't bet on it. At the clock frequencies the ASIC run at those spikes/dips can cause all kinds of havoc if not decoupled at the chip. Methinks that kind of oversimplified thinking is what got BM.ch into trouble in the first place. Power distribution at high switching speed loads is a very localized thing.

reserved...

I think they blew it because they didn't layout the board in a way that you can have large heatsinks!

Bitmine design have heatsinks only a little bit bigger than AMT.

It is not like the Technobit and Dragon design where the heat sink in much bigger.

Man... I really don't want to do electronics... but heck... is this not dirt simple?

BTW... Innosilicon A1 chips are going for $45 a piece.   The A2 is around $185 but use the same layout.

If anyone has got some minor experience in this,  we can layout a board and do our own manufacturing!!! Wee!

Okay... so if there happen to be resistors and not capacitors that should be there... a temporary fix is just to remove the resistors.

The capacitors are there to mitigate against voltage spikes, but it is not entirely necessary.

Not sure how well the reference boards work but if they do, then yes, just get a bunch and string'em up. Is what the Ant's do. Looking at their boards each is made from 4 identical sectors. Each sector even has it's own Vcore supply.

I assume that BM.ch just wanted a 'clean' design for the 8-chip boards. And blew it royaly.

Anyone know anything about the GPIO daughter board?

Pin layouts on it?    

What I really don't get,  is that there's a reference board with layout and BOM.  Is it not obvious how to actually create a board for this?  All that's on the system is are A1 chips and a voltage regulator.  

Just string up 4-8 of these and you have your system.  The driver code for the PI is also available.

What the hell am I missing here?  How complicated can this get?

SMD resistors come in the same package. The ones isa found that begin with R+ are 10k resistors. Those should be caps.

Those orange thingies...  I checked with digikey... http://www.digikey.com/product-detail/en/T520D337M006ATE010/399-9780-2-ND/3724870   those are capacitors.

Correct me if I'm wrong here... it's been over 20 years since I worked on this stuff.