Oh hey yeah, I sent you an email about some buck chip options and ideas from the last six hours or so of research and calculation. Think about that for a while and I'm gonna fall asleep.

But yeah the test chip, I think it was at 125MHz and it was definitely submitting shares at about 580mV at one point. Awesome.

Yep. I did some quick calculations and it should only take about 220uF to handle the burst transient tolerably before the regulator can kick in. We had something like 360uF on the output and it still bottomed pretty good. Has to be overcurrent protection kicking in (at about 11A). The new design I'm working on right now (literally) I can set the overcurrent pretty much wherever I want, so I'm gonna make sure it'll be good for those kinds of burst transients. If we start testing higher voltages and frequencies and find that the burst requirement gets worse, I can just take the overcurrent a bit higher with a change of one part. The FETs will be good for a heck of a lot more current than this thing will ever need. We're aiming to make it capable of maxing out a 1.5A port in steady state, which it should hopefully find around 300MHz (16.5GH). I mean, it'd be pretty great if it could do more than that. Since we can do pretty much whatever the heck we want, I'm gonna give it a full range of 600-800mV so you can clock it up to 400MHz if you can keep it cool enough. Whatever.

Yeah, the schfifty-three we put on there I think has... 3x 68uF output caps? And it behaved admirably even at starting the chip on about 640mV 250MHz. I didn't test the chip higher than that but I probably will tomorrow. It really wants a higher voltage to run at 300MHz stable anyway. I'm kinda surprised it started at 640. Our homemade regulator (well, except for the inductor) will be about as beefy as a schfifty-three circuit (by which I mean the TPS53355DQP) which has been the favorite of miners for a while now - S1, S2, S3, AM Cube, AM Tube, BTCGarden AMV1, New R-Box and a few other things use 'em without any problems. But they're like five bucks per chip. I can build an equivalent non-monolithic circuit for under $2 in silicon, with better power dissipation since it's less dense.

It's likely the controller and blade connected to the proper PCIe are still good. The blade connected to your 8-pin, I don't know what kind of reverse protection they have but it's pretty likely that's been roasted by the reverse current. My guess would be, best case you shot the PSU and half your S1.

The burst was probably around 12A output, for maybe a tenth of a second. It's long enough that I could see it on the mechanical gauge on my old benchtop supply, so not a microsecond pulse or anything. I'm gonna try to hammer a lot of input capacitance onto this thing and help that out. The schfifty-three is running off a USB jack plugged into a port with a 470uF adjacent, powered off my benchtop supply with about a foot of wire, so probably not too dissimilar to expectations from a decent powered hub.

The baby chuckwagon I'll work on design tonight and we'll send off for parts tomorrow. Between tomorrow and Friday I'll probably test two chips running in parallel off a common core voltage since I have a 30A output regulator handy, then crank up the schfifty-three to output in the range of 1.2-1.4V and see about putting two chips in a string. Once parts arrive for the baby chuckwagon I'll assemble one of those and see how well it handles driving a single chip up to, oh I dunno, 300MHz or so.

The FETs I'm looking at should be good for more current than the inductor will handle, and the inductor is good for more current than the chip will ever ask for. We'll probably end up shifting to a 40mm heatsink. If the regulator beats about 88% efficient at around 600mV we should be able to run this thing at 150MHz (8.25GH) on a stock USB port, which means 0.30W/GH device-level. Which is pretty great. And since we're building the regulator to handle balls current, there's no real reason why we shouldn't allow the voltage into a higher range so you can push the thing even farther. If it'll run 300MHz off the stock heatsink (with adequate airflow) you'd be looking at 16.5GH off about 1.5A from the USB port, which is doable with a good hub, and still more efficient than an S5 stock settings.

Okay, so, the BM1384 appears to really like current while initializing. Even with a fairly stable regulator, the starting problem for higher frequencies didn't go away. I once had it running at 175MHz, but I think I had the input voltage to the buck cranked up. Nothing I did within reason to the circuit could get it to handle the brief burst current required to start the chip above 150MHz. This included adding 22uF bypass capacitors to all four corners of the chip, increasing output capacitance of the regulator to 8x 47uF ceramics, and building a socket onto the regulator board so it'd plug straight into the breakout board and get rid of wiring impedance.

To make sure it wasn't the chip's fault, or the breakout board's fault or whatever, I rigged up a stupidly-overcapable regulator by cutting a chunk off a dead AMV1 Garden blade courtesy of CrazyGuy. The 53355DQP on there is rated for 30A, I figure that should be good enough. I modified the board to socket straight to the breakout, powered through a USB jack, and with a voltage range adjustment of 600-750mV and tested the chip successfully up to 250MHz (13.75GH) with that attached. The current meter on my bench PSU touched 1.8A during the chip-init burst transient, but the running current was around 1A.

So what I'm thinking is, screw that IR3899 anyways it doesn't seem to be able to handle that burst transient. If it's pushing 9W of 5V at 85% efficient down to about 630mV that's briefly touching an output current of 12A. I knew output current was the cause of my undervolt issues, but the outputs I was seeing didn't make sense for overcurrent protection when compared to the datasheet info. Given that I ruled out literally every other fault condition though, that has to be it. I'm not sure why the chip was reacting in 2.4mS instead of the 20mS listed in the datasheet, but whatever. So yeah, screw that chip anyways.

I think we're gonna go with what we're calling a "Baby Chuckwagon" strategy, which is instead of trying to jack with some unnecessarily-complex freakin' tiny-pin-pitch (or more likely, cut-up no-lead) package chip, we're gonna work out a design using a basic regulator controller with integrated FET drivers, and a couple pretty good super-low-Rdson external FETs. It'll require more board space, but the parts cost should actually be a bit cheaper, and the packages are decidedly non-evil.

While we're waiting for those parts to come in for testing, I can crank up the voltage on the schfifty-three board and start testing chip strings to prep for Amita design. How sweet would it be if I posted something tomorrow about how I had a two-chip string pushing 20GH or whatever?

So I slapped the regulator around some more today. I was really hoping that compensation issue would be the end of it, but nope. Not the case at all. I built a powered USB port and put a jack on the regulator test board to simulate the input conditions it'd see from a powered hub, and fitted a socket to the output so it can mount directly on the breakout board to minimize output wiring effects. Rehashed the compensation with higher input and output capacitors to help buffer load transients. I beefed up capacitor bypass on the VCC lines and tied them to power through a 4.7ohm resistor to help keep out burst current discharges on the bulk caps from affecting their voltages. The Enable line looks like it never gets below about 2.4V, twice the threshold for turnoff. Pretty much the only thing left is to isolate the soft-start pin and see if it's getting jerked low and causing an output discharge. Hopefully I can get it worked out tomorrow.

Also, I did the math and if I just stick a linear regulator on it set for 0.6V, the chip should still reach about 1GH off a 2.5W USB port which means that, at the absolute worst case, it's still three times the efficiency of a Block Erupter.

Please elaborate on which part of that multi-person rambling you're responding to, as well as the idea itself.

Also, looks like I might have found the issue with regulator stability. There was an error in my crossover frequency calculation (partly due to a really poor explanation of some things in the data sheet) which put me off by a factor of 6.5, shifting what I thought was a reasonable frequency to base calculations off of about an order of magnitude high. I'm hoping that explains the poort transient response I was seeing, as well as the high-load instability. I've recalculated everything (this'll be regulator iteration six I think) based on the feedback values I calculated earlier today for the 600-750mV adjustment range and the altered crossover frequency calculation (which affects the calculations for five of the six compensation components).

I'll scrounge up the closest parts I can find (most calculated out to pretty straightforward values) in the morning and see if I can't get that regulator stable. I'm really looking forward to reduced output ripple and high-current stability. I should be able to push this thing enough to ask for about 1.6A off the USB port. If it works how I expect it to work now that I'm pretty sure I know why the last five iterations didn't work (since I've been using the flawed initial formula from the beginning), we should be ordering test Compac PCBs later this week and I'll be able to run out a set of hashing efficiency curves for the thing using the prototype boards. Hopefully I can talk Novak into integrating the arbitrary frequency code into a cgminer driver in the next day or two so I can run curves with a lot tighter data points than every 25MHz. I guess I could just crack open the source code, add to the lookup table and recompile, but software is his job anyways.

Also, if you haven't seen TheRealSteve's Block Erupter Hack-off go check it out. We're already scheming on a couple ideas.

Phil - we should have some IceFuries inbound. I can't tell from an end photo what is what in that picture so I can't answer any more than that.

Meech - maybe, if it weren't a completely different protocol and also garbage controller software that'd be impossible to work with. We'd end up having to build a new backplane to fit in the case, as well as new boards, and it'd end up being us purpose-building a lot of new and fairly complex stuff in order to use the box it came in. Not really the best thing to put effort into, so no.

If anyone else asks about whatever else hardware we could look at building for, I'm going to direct them to this post:

NO.

I've already got enough on my plate with the FOUR miners I'm already designing solo and from scratch, and I'm not going to add a fifth, or sixth, or seventh thing to a list that's already probably longer than we can feasibly pull off with the resources currently available. Please don't ask again.

No, probably not. As I've mentioned a couple times so far in this thread, it's not likely we'll build S2 upgrade boards since the hardware design is basically a one-man team, the funding for the whole dev side of things is fairly low, and that'd be a FIFTH thing for me to work on. If Bitmain bails on the design that they already have completed and are apparently testing, maybe we'll give it a go but I probably won't have time until late enough in the year that the BM1384 has already been replaced.

So, updates. I built a USB port with current metering built in. It's got a pin header for taking in external power, or jumper a pair to use current from the USB upstream. I checked the calibration on the current measurement with about nine data points and it looks to read 3.5% high, so that'll be taken into account on all calculations.

After talking with Novak and looking over the hashrate extrapolation charts from last night, I'm going to do some more shifting on the regulator. We do still like that low end of the range, around 625mV, and there's no practical reason for the Compac to run higher than 750mV unless you want to pull more than 2A from your USB port for marginal hashrate gains. Which seems dangerous and not recommended.

With the wiring I have now between the regulator and breakout board, I'm seeing a fairly significant voltage drop (10mV or so) at the Vcore pins on the breakout board. So my starting voltage right now is definitely not the same as my running voltage. I've got it running 125MHz/6.875GH right now, starting voltage 635mV running voltage 625mV at a current draw of about 483mA off a 4.85V USB line. This is about 2.34W, so about 0.34W/GH which isn't too bad. I need to recalculate the regulator AGAIN (iteration 5) to shift the adjustment range to 600-750mV.

I think some of the instability I'm seeing from higher current outputs is more related to the compensation loop than temperature problems. I squeezed my feedback resistor down a ways to get 600mV, which throws off the frequency response of the compensation, and now when the chip tries to start at 150MHz/640mV it wigs out like it was doing during the upper ends of my load tests. So that's something I'll have to play with. I'm also thinking of, for the test board, pulling the feedback sense directly off the breakout board instead of the regulator board, which will take the wiring into account for regulation and should cancel out that difference between running voltage and starting voltage. Improve load regulation, that is to say.

So, maybe tomorrow afternoon I'll rework the regulator back to stable at the new adjustment range and get some good efficiency curves. Novak's gonna be looking into integrating the arbitrary frequency code I wrote some months back for S1 workups into a cgminer driver so we aren't limited to the stock settings like 100/125/150. The code takes in an arbitrary frequency and generates the hex value for the nearest compatible frequency that the chips' PLL multiplier can operate at. So once that's in, we'll be able to get efficiency curves with a much higher granularity - every 3-5MHz instead of every 25. And hopefully the regulator will be plenty stable up to 5W input, which at the bottom end 625mV would be about 6.5A output. I haven't tested it successfully up that high yet.

"All the new circuitry" amounts to changing some one-cent component values from the original spec, and adding about twency cents in components for the altered LED driver. It'll take a lot more than that to change the estimated $20. Right now that's still an estimate, but one we really hope to hit. The only real question remaining is heatsink cost, which I just sent an updated spec out on Thursday so I don't know yet what a price would be.

Nope, it's going in our mining hardware museum.

Text? No, he only gives his phone number to friends.

Yeah, the pins on that regulator chip are quarter-millimeter wide with quarter-millimeter spacing. I don't know how you pulled that off. And then there were the traces running between pads on 0603 parts. Ridiculous.

I'm already working out in my head how to build up a power-metering USB port I can run off an external supply. I'll throw it together Saturday afternoon and get some numbers out of it. I think I'm gonna need to beef up heatsinking on the ASIC though, before stress-testing too much. That crappy little sink that's on it, especially being surrounded by cables and jumpers, won't do 3W by itself.

Also, Novak, if you beat me in to the shop tomorrow the thing's running on my bench and the LEDs turned out sexier than I could have hoped. Take a gander. The drive circuit was also pretty simple. The RF pulses were so short you couldn't even tell anything was lighting up, so I used a double FET inverter with an RC delay on the second gate to hold it closed longer and extend the LED pulse a bit. I was afraid I'd have to use a comparator circuit but it works well enough with the FET gate threshold and such. I didn't scope the RF pulse but it's on the order of 50mS; my driver extends that to probably 150-200mS.

Surprisingly enough, there has been zero smoke so far. I did throw off a part or two while trying to debug some issues with one board, but I think they're probably actually still functional. I was really surprised the regulator didn't blow up, but that PCB was literally the best toner-transfer etch Novak has ever done. Super precise and not a single cut trace. I might throw together an LED board so I can have my breadboard back, and get some updated pictures. But the more pressing priority will be getting actual hashrate efficiency curves off of it, instead of extrapolations and estimates.

That particular toy in the picture will never be sold. It's a landmark thing for us. Maybe you can buy one of the finished product Compacs, but not the test hardware.

Also, I tend to fetch cheeseburgers from the Waffle House. There's something nice about being able to watch food being cooked before you eat it - you know exactly how fresh it is, that that patty hasn't been sitting under a heat lamp for the last five hours. Also, for the cost of a large fries at a fast food joint I can get about a pound and a half of hash browns, which are better anyway.

So, the new regulator iteration has been tested to its limit. I did get over 5A out of it at several tested voltage setpoints, but that's still not really great. I'm hoping when it's mounted on a proper board with better heat dissipation it doesn't wig out quite as much.

Here's a chart of actual regulator performance. The current output is an estimate based off using low fixed resistance loads and not directly measuring current, so the actual current (and therefore actual efficiency) could be skewed, even though I was using 1% tolerance resistors. Wiring and solder joins can have a large effect on overall resistance when the overall resistance is pretty low to begin with.



And now, the fun parts. Based on the current outputs and measured regulator efficiency at various voltages, and extrapolating from Bitmain's efficiency and performance charts for the BM1384, here's some expected performance data on the Compac.





Looks to me like 7GH off a standard 2.5W USB port is attainable at 675mV. The regulator as designed should give us 650mV as a lowerbound, but I had to approximate some of the component values for this test which has shifted the feedback slightly high. It's actually higher than I'd estimated (675mV versus an expected 660mV) but that's not terribly surprising. At 650mV, if the chip'll even start (should, but no guarantees) 8GH should be possible from a stock USB power. This is, unfortunately, not taking into account the added power consumption of the CP2102, IO and PLL LDOs, and LEDs. This all shouldn't add up to a substantial amount of power, but when the ceiling is 2.5W everything needs to be taken into consideration. But that's a job for another day.

I just finished rebuilding the test regulator for the third time (which makes this iteration number 4) and am about to load-test it. The present design adjusts between 650mV and 800mV and hopefully remains stable at currents above 4A output.

If it behaves well enough, tomorrow I'll look at running it on the hashboard and see about getting some initial W/GH estimates.

Phil, you're on the short list for receiving prototypes to field-test. If the current iteration of the regulator design is good enough, we'll probably be sending off for prototype PCBs early next week and start populating for testing the week after. Which is going to be a pain to do by hand. The bottom half of the board is freakin' dense. In one square inch there's the CP2102 USB/UART chip, a UART level shifter, two LDOs, the oscillator, two LEDs, a couple FETs, and the entire regulator circuit with its approximately two dozen support components. But that's what happens when you build an 8A VRM onto a stick miner for a chip that needs 3 different voltages and takes 1.8V IO and requires a topside heatsink. But we designed the whole thing with nothing smaller than 0603 parts so it's actually possible to do by hand until we can get a decent robot. The worst part will actually be the ASIC itself, since it's a no-leads chip with a belly pad. If there's a hair too much solder on the belly pad (or the VDD corner pads) the thing floats too high and the data pins don't make contact with their pads. So it requires very careful pasting and very precise placement. Which if we don't have good machinery is going to make assembly suck balls, especially on TypeZero boards with a whole bucket of chips.

You already quoted the answer to your question:

"I'm also going to rig up a USB extension, possibly externally powered, that'll allow me to directly measure input power so we can get some sample efficiency curves off the prototype boards. Since the various modules (right now two PCBs, one of Novak's Prisma adapters and some breadboarding) are basically identical to the Compac as it stands, the power requirements should be pretty close to accurate."

Which is to say, I have not tested this yet but it's on my list for the near term - probably this weekend. Also, the watts and hashes per volt are a whole field of curves dependent on the operating frequency, so there's no simple answer. Once I have the power metering set up I'll probably have to spend a whole day testing different operating voltages at set frequencies (hashrates) to see what's stable and get a series of efficiency curves.

As an initial estimate, I believe the 6GH point I'm at right now should be pulling around 2.1W but I haven't measured that, it's based mostly off theoreticals. Yes, that makes this stick miner more power-efficient than the S5.

I'm not sure how much it's a problem of hot restarting or just that the prototype's pretty jankety right now with wires on headers flopping everywhere, but sometimes when it cuts out it decides not to come back online for a few minutes. I've noticed it's not as stable at 125MHz as it was at 100MHz, which could be heat-related or voltage-related. It'll take some more testing to be sure of which.
I pulled the test board this morning to play a bit with the LED drive, so that cost a few minutes of hashing, but it's been running alright basically continuously for the last two days and some. The regulator has no heatsinking, and the chip has about a 2cmx3cm thing taped onto it not very fancily.

I'm pretty sure we won't be selling these in cases. It adds to the cost and how many folks are going to just chuck 'em anyway?