If you really want to, but I probably won't wear one.

If I hadn't burned about a week and a half on the first regulator design we'd be a lot closer to "on schedule". That stupid thing never worked right, but the prototype for the current version, Novak had me a board etched before the parts arrived and it worked perfectly on the first shot. Stupid cut-up QFN monolithic junk.

I had a week off when I was waiting for the PCBs to come in. Then I jacked with 'em until midnight and worked on 'em all the next day. I did relax the whole weekend though, sat on the couch all Saturday and Sunday reading comics and pretty much didn't leave the house. So now I have to be dilligent and work hard for the next 20 days or so again.

The hub is not the limiter. 5V is driven directly off my bench supply, and it worked fine running a U2 and my crapstack prototype. I built that hub to run half a dozen overclocked USB BEs. The CP2102 enumerates fine.

Stencil? Yeah a stencil would be nice but we've never had stencils for anything. It'd sure be nice. Also, every board we've sold in the last 9 months or so is one that I hand-placed and soldered (I can do about 12 DPS2000 boards start-to-finish in 100 minutes including cook time). We're budgeting an actual pick-and-place and reflow oven for when the Compac moves into production (we need it for other builds, not just miners - yes, we do other things too...). I'll look into that link for cheap stencils, that might come in handy. The ASIC is the only part giving me grief right now with soldering, but since that's the only part that really matters (and also about 20% of the total parts cost) it matters the most by far.

I'm not the hero - I'm sidehack, the hero's hacker sidekick. That's literally what the name means.

Yeah, you see us designing stuff specifically to educate. I don't really blame under-educated consumers nearly as much as I do lazy consumers - the guys for whom answers are readily available if they'd take the 30 seconds to look for them instead of crying for help on everything. People who appear stupid sometimes genuinely are, but for most people being stupid is a willing act they've been allowed to continue. I'm constantly annoyed by stupid customers but we sell to 'em anyway. Then we hand them enough resources to solve 90% of their problems themselves if they wanted to not be lazy and that's it. But not everyone can get away with that, especially with diverse and complex product lines.

Especially because a lot of customers are, themselves, no-brainers.

Yes, behind. I was hoping to be on TypeZero design by now, and I haven't even gotten working Compac prototypes distributed for testing. Just because I'm not racing other independents doesn't mean it's not a race. I'm also racing new chip generations, and the network as a whole, and my own opinions about how long a thing should take to get done.

If by "on a roll" you mean "several weeks behind ideal projection and frustrated these things didn't work last Thursday", then yes I'm on a roll.

At least the LED works. Some people won't really care if it mines or not so long as the LED works right.

If it was BGA I wouldn't be messing with it. The QFN is bad enough by itself. I wish this chip had a larger cross-section of pad at the edge, which would make checking solder jobs better because it'd wrap around a bit. I think the most problem I have right now is being able to stick the chips down reliably. The alignment is difficult because there's basically zero reference for when it's in place (and it has to be correct within a quarter millimeter on two axes) and I can't trust floatation because if there's enough solder on the belly pad to float into place the pins might not get soldered, and if I press it down to connect those the belly solder could bulge out and short across to anything, including the corner Vcore pads and short out basically anything.

I've succeeded in getting two of four sticks hashing unstably and unreliably but I have seen submitted shares. I can't say for sure what the issue is that gets the chip working haphazardly at about 10% capacity, because the breakout boards were pretty much all-or-nothing. Could be signal integrity issues, I really don't know but I hope that's not the case because if it's a result of PCB layout we got problems. The signal traces are between ground planes - at least they should be if the etch fab didn't get layers out of order.

Anyway, here's a sneak peek at the sticks.



Got ya a fully populated boar with heatsink temporarily attached. The final heatsinks will probably be that dimension but anodized green. Note that about 2/3 of the visible area is the bigass regulator, complete with external FETs and 360uF of output capacitance for handling about 20A of low-volt DC. The adjustment pot is in the bottom corner.



Here's some crap plugged into my jankety powered hub. The crapstack in the back you might recognize as a 1384 breakout board on the test regulator and running off one of Novak's USB adapters. The thing in the foreground is literally everything I just said except on one board quite a bit smaller. You might note a lot of flux crust and such around the ASIC. That's because I've pulled the thing, restuck the thing, wicked pads, pushed it around and whatever probably close to a dozen times. Ain't fun when every piece of the circuit seems to work just fine except for the core. Let's think of it as "in a vegetative state".

If you're wanting over 20GH from 10W you'll need to wait for the two-chip stick. It should actually hit slightly better efficiency than the one-chip.

The TypeZero will have 4-wire fan headers. How they operate and/or are configured is in Novak's domain right now.

Bluetooth? Not on anything I ever run, which probably means not on anything I ever build. Our boards are USB tethered, so you can hook them to anything you want and run that wirelessly if you feel like. But I'm not going to limit people artificially to "you have to use this type of networking" because nobody wants the same thing. Slap a cheapo USB-enabled controller on it and do whatever you want from there. Sure I know people seem to like having things they can plug in an ignore forever, but building something that you can only plug in and ignore is stupid. Flexibility and configurability are not. That's where the fun comes from, and most of the learning.

Well, so far the prototypes are tickin' me off. I have two fully populated and neither works properly. One seems to be powering the chip properly, but the chip's never really lighting up. It's getting serial data but not responding at all.
The other, I honestly don't know what's up. Sometimes it enumerates in cgminer and starts to hash around 300MH/s, usually craps out after a few seconds but a couple times it's run for minutes and once or twice I briefly saw the hashrate over a few GH/s. If you look at the 1BURGER stats on Eligius you'll see a slight lump recently; most of that is from a U2 I was using to test the powered hub setup but the lone share about 19:00 was one single 128-diff that got submitted while the thing was temporarily approximately working. So TECHNICALLY a success. A stupid one.

I'll keep working over the weekend and see if I can't get things ironed out. Those BM1384 are really hard to get stuck down right, and since the pads don't wrap at all there's no good way to inspect them for proper solder joints. I need to get a friggin' microscope or something.

Actually, you reduce resistance. I haven't tested on an S2, but based on my tests on the S1 it should be possible to get 0.8W/GH out of them, board-level.

The buck in any PSU worth anything is running a transformer-based forward converter behind an active PFC boost converter. The active PFC uses an inductor to push regulated input current uphill into a bulk storage capacitor, adjusting on a cycle-by-cycle basis to keep the driven current proportional to the input voltage (so it appears to your power lines to be a resistive load with a power factor of 1), and adjusting the I/V ratio between every mains cycle based on how much energy needs to be pushed into the bulk cap to keep the voltage relatively constant even if output power requirements fluctuate.

What this means is, the buck part is always taking from a constant voltage and shifting down to a constant voltage. The part that works better with an increased mains voltage is the PFC boost circuit. They're usually taking the mains up to something like 380VDC, which it takes a lot less push to get that from 340Vpeak rectified mains than it does from 170Vpeak. Switching duty cycle is reduced, and more importantly the overall current is reduced, which decreases your conduction losses (which are proportional to current squared).

If you had something outputting a 350-400VDC bus you could wire it straight into the bulk cap, bypass the PFC entirely, and probably run your PSU at about 95% efficient.

Yes, the S5 has 30 chips. This was discussed extensively in the first few pages of the S5 thread. It gives 0.8Vcore average per chip so they operate up to 400MHz typical.

If you're warming ambient conditions for humans while outside conditions change, you might want an adjustable thermostat. Something that could restart cgminer occasionally with changing parameters to alter the heat generation from your miners (and calculating optimally profitable hashrates for the desired heat output) would be pretty slick.

Additionally, I got the Compac test PCBs in about 7PM. There's exactly one UPS driver on our route that has realized we're not here at 9AM, and the other fifteen that seem to cycle in and out appear at exactly 9AM to leave door tags every single time. And then they don't come back by on their way to the depot at the end of the day (which, by the way, we're one mile from and they drive past us anyway). I should request we get put on the same route as the tractor store next door that gets daily deliveries in the afternoons. Anyway if we ever want to actually receive our stuff we have to wait until the day's end when the truck has driven past our building probably several times without stopping and pick it up from the depot. So that's why it's midnight and I'm still working on the prototype instead of having it going ages ago. I'm also getting annoyed because when you couple the stupidly laggy browser with the junk Novak dumped in my keyboard today it's really freakin' hard to type anything. I will not be posting pictures tonight. Y'all just gotta wait and remember that I'm the curmudgeonly hardware guy, not the nice one.

A well-designed miner will run with an input voltage below 12V. Depends on the topology, but the Spec1 should be designed to work off a Vin probably 8V or so. Spec2 may not, since its power requirements are different. A buck regulator also works more efficiently with a smaller Vin-Vout difference so running at a lower Vin is actually better (to a point; reduced Vin means increased Iin which can cause increased wire losses). For a well-designed miner.

Almost all of your questions have been asked and answered more than once already. Most of them are addressed in the first post.

The chip I'm looking over for the TypeZero power systems should take in over 12V easily. Limitations might come from accessory components (management rail power, input capacitors, etc) but I'll probably make it work off at least 16V safely.

Also, the S5 has 30 chips not 32. And with software-adjustable clocking, you don't have to take the machine apart and put it back together to make changes, which is probably good for an industrial or immersion install.

Dev's gonna depend on feature set. Right now we're looking to refit the Prisma formfactor with a 30-chip board that'll be software clock and voltable. The primary goal is the Spec1 board to fit on S1 chassis, but once that design is done it won't take much to extend it to the Spec2 Prisma refit board. If you want higher density I can probably work on condensing it with more chips per node and a higher current power system. The easiest thing will be a straight string topology, but for your needs the ability to alter the operating points (core voltage and clock) could be worked into an adjustable thermostat.

I do like the idea of a miner-based heating system.

One issue with a board-cooled string design is that each node in the string is at a different ground potential. If there's any short through the heatsink from the ground plane of one node to another, the best case is your PSU trips out and the worst case is stuff catches on fire.
For an anecdotal reliability comparison, the Prisma used board-cooled string and the S5 uses top-cooled string. Not many S5 catch on fire. As much as it'd be easier to manufacture and probably easier to install, I don't like the idea of a board-cooled string because it also makes it easier to fail.

I was thinking of mentioning to you that we're looking into a Prisma refit board with the BM1384 - same size and stock power level. I know you were playing with immersion cooling on those several months ago.

Socketing seems to be an unlikely way to go. Given that (with the exception of BitFury R1 and R2 chips) nobody has ever made a pin-compatible next-gen chip, given that sockets for the fairly customized footprint may not exist, given that if they do (or can) exist they'll cost more than the chips themselves and the chips themselves would cost more than the PCB and every other part on it, it's probably not feasible.

My primary interest is still to build things geared toward individual miners. If one of the big guys building industrial stuff wants to repurpose for immersion cooling, that's fine. But unless the big guy specifically commissions a board design and pays all the dev costs, I'm probably not going to do anything specifically for his benefit.

Compac test PCBs should be here Friday. I have some sample heatsinks I can play with arriving about the same time, but some better samples (possibly even anodized, possibly already drilled and tapped) might be another week out. If the PCBs come in Friday I'll do initial testing over the weekend and might get some to my first wave testers early next week.

Another thing arriving in the same proto-etch batch are some L boards made for mounting two BM1384 with stringed comms. Using one leg I can do two chips talking direct at the same voltage level; using the other leg I have two chips in string mode. If the string leg works, I add that to the Compac PCB, alter a few parts and send out for a batch of Amita PCBs.

We were running around and in meetings most of yesterday, and out of town all day today (7AM until 6PM) so I haven't gotten the test work done on the TypeZero power that I wanted to do. I might be changing up how I do that, fetching a different controller chip that has some drawbacks to the current system but also some good benefits. I'll probably do competing designs and see if the performance from the one outweighs the annoyance factor - honestly, for this board performance is the key. Efficiency needs to be as high as possible.

But if the Amita/L-board tests work, and I have a workable power system within a week or two, the only concerns left are integrating temp sensing, fan control and core voltage adjustment. If that's fairly straightforward I can have a TypeZero board ready to prototype by month's end. We'd still have to get a crap-ton of money, and get word from Bitmain about what we can do about chips, but I sincerely hope the TyepZero is less than 4 months away.

But BGA is a terrible idea for all the reasons we've listed extensively. I hope it doesn't become the standard, especially if string topology does become the standard. But the CPU model like what Spondoolies has mastered (and everyone else that's tried has gone bankrupt or made terrible hardware or both) has a definite efficiency ceiling and complexity floor which leave a lot of room for improvement. Most of which improvement can be made quite easily by using a matrix of small chips. This also makes design and manufacture a lot easier too, since you don't need a friggin' X-Ray machine to inspect your solder joints. The kind of chips Bitmain and Avalon are doing already are the best option and, unless the cost savings from chip fab outweigh the cost increases and efficiency hits from every other aspect of the process, I don't see a good reason to do any different.

If we do a backplane, it'd be be basically a PCB and edge connector version of PCIe cable + USB cable. The only thing a backplane might do differently is integrated USB hub chips so there's only one cabled connection between the backplane and a controller. All our boards will have USB connectivity, so making a box with a bunch of 'em in it would be possible by sticking a several-port hub in there and getting a bunch of short cables.

Yeah, one of your guys emailed me about a year and a half ago about power supply something or other and I never got a reply back from him.

I don't really care about liquid cooling, since it's a rich man's game and screw those guys anyway they got enough help.

I didn't look much into Wasp but I did remember reading something about that backplane idea, which is why I asked specifically about it. When Novak and I were discussing ideas back in December I mentioned that idea. We like being able to use a generic controller, which is one reason we're doing USB connectivity on our boards instead of some application-specific comms and cabling which requires a specific controller build, and having a power+USB backplane fits right in with that.

We probably aren't going to build backplanes, though, since the demand for them will probably be small (by definition, the number desired will be a fraction of the number of boards) so the cost will be high, and cabling is already pretty easy to come by.

Translation: you want an S2 upgrade kit?

Unless someone made a backplane with power lanes capable of pushing at least 150W per socket (250W if you want the 30-chip board on it) and also have a USB data pair to each port, it wouldn't work. Did WASP build something like that?