Yes, sweet tea.

I typically eat 3 cheeseburgers, 5 hashbrowns and between 6 and 8 glasses of tea.

So yesterday evening I finally got the wire cutter parts re-machined and reassembled, but today was a maintenance day for hosting and stuff around the shop. So very little progress. We've had the protoCompac mining on the 1Burger at Eligius at 125MHz (about 6.8GH) for the last 24 hours without any problems. I have noted the chip doesn't like to start below 0.65V, which might help explain what some people were seeing with S5 undervolt testing. That's about 10V across an S5 board. I'll probably change up the regulator voltage range to start about 625mV and run up to about 800mV.

I'm working on using the RF pin to trigger sent-share LED like on the U1/U2 but having no luck. The flag is pulsing with a short-enough duration that persistence of vision keeps me from seeing an LED change with the setup we have. I can make it work with a one-shot circuit but it'll take some extra dev time and a bit more board space. Hopefully I can keep things simple. Having a flashing sent-share LED is way better than not having a flashing sent-share LED.

I'll probably be on Compac all day tomorrow. Hopefully I can get enough ironed out that we'll be sending off for actual prototype PCBs early next week.

Also, tomorrow is Friday which means Cheeseburger Day. So I won't actually be working all day. Gotta make time for sammiches.

So, I currently have three tested functional BM1384 breakout boards. I didn't get any further work done on the regulator for doing that and other things. But one of the breakouts is heatsinked and running on Eligius pointed at the burger address (http://eligius.st/~wizkid057/newstats/userstats.php/1BURGERAXHH6Yi6LRybRJK7ybEm5m5HwTr)

I think the regulator problems are temperature-related. I hope to have time to look at them tomorrow, but our wire cutter went down today and I have to weld/machine some parts back together to get it going so we can meet orders. Grumble grumble Chinese steel grumble grumble.

In case anyone's wondering, here's a chart of the regulator efficiency as tested. It still needs work, as I'm having trouble getting more than about 3W out of it before it starts to cut out. I've already ruled out undervolt protection, and it's certainly not overcurrent, so it might be temperature-related. More testing is needed, but I want to be able to push it to at least 6A at 0.75V and right now I can't get 4A at .75V out of it.

Also, yes, the regulator is currently attached to a populated BM1384 breakout board and capable of operating. We exchanged high fives until our hands went numb. Tomorrow's Tuesday sandwiches will include celebratory ice cream.

Once I get the regulator ironed out, I have a few changes to make to the Compac PCB. We'll need to test chaining and see how hard it'll be to get two chips in series to talk to each other.

Nothing except time, currently, and you're not really capable of that kind of thing. I'm probably gonna focus on hosting maintenance tomorrow and hand off Compac dev to Novak for software stuff, maybe get initial Amita board design done and get back to the TypeZero regulator work which I haven't touched in a few weeks. I guess I do have some paperwork to take care of for Bitmain as well.

We're also in Missouri, probably a bit south of you though. We've been hosting machines fairly small-time (50KW, with a 100KW upgrade in the works) since about last August.

Yeah, I like to think of the TL494 as the 555 of SMPS controllers. I'm not familiar with the part you mentioned.

The TL494 has dual error amplifiers, so you can wire up an overcurrent protection on it natively, or overtemp, or whatever. And an external deadtime control pin that basically allows for additional error amplifiers to be tied in to reduce the duty cycle when necessary. Switching frequency control up to 300KHz and dual outputs that can be run in parallel or push-pull. Buckets of those chips were used as half-bridge forward converters in old ATX supplies, I don't know if they're still that common anymore though. They can be driven by an external clock, but the only thing I don't like about the chip as a whole is the external clock has to be a sawtooth. Makes sense since that's what PWM is compared to, but it'd be easier to tie 'em together if they could synchronise off a square pulse. I really like that they don't use a fixed internal reference voltage, which is what makes software adjustment possible when you use a DAC. And then you're not limited to whatever lowerbound the fixed internal reference gives - typically 0.6-0.8V which you mentioned 0.25V so that'd be out by 99% of driver chips.

Working on the miner project is the most fun I've had at work since about a year and a half ago, when I was spending the first 8 hours of the day on hardware refurb stuff (the real job) and the next 8 hours designing and building a 7V->150V push-pull forward converter for a nixie tube power supply around the TL494. That was a good week.

The chip being used on the Compac is an IR3899, which I noticed last week is the big brother of the 4A-rated regulator on the Bi-Fury stickminer (I forget its part number).

Snazzy idea there. I have zero electronics publications, so you're definitely winning.

The AntMiner U3 has software-definable voltage, but it uses some crazy TI digital driver that's buried under NDAs just to get a datasheet, and so is the probably-all-digital FET driver it chats with to get things done.

Yeah our engineering philosophy is definitely minimalist - or more precisely, minimal modularist. Find a simple and reliable thing to do the job well and use it; if the job is manifold, split it up and find simple reliable segments. Microcontrollers can be the simple thing, but you have to consider that though it's only one chip, a microcontroller is an incredibly complex and intricate circuit all its own. When we were looking at multiphase regulator design we looked at something like the TI driver chip on the Habanero, which can handle 6 phases with all digital monitoring and something something potatoes. It's pretty fancy. The thing incorporates a 32-bit ARM processor on top of a bucketload of ADC circuitry and some DSP segments I think.
But I think instead we're gonna use TL494s to do the heavy lifting, with a small microcontroller to generate synchronised clocking signals and control phase shedding where needed. Small code, few errors; easily modularisable, flexible and reconfigurable; anything time-critical is entirely analog. There's probably more compact and integrated ways to do it, and with a fast enough processor we could probably replace every piece of silicon short of the switching FETs with a single chip, but I'd have to be convinced that it was legitimately functionally better, not just easier.

I really like the TL494. I don't know how popular it is anymore, in a world of highly integrated and automated everything, but that chip is a beast. So many functions and basically all of them are broken out to pins. Lots of control, lots of flexibility, zero programming requirement. For the record, 2/3 of my college education was Computer Science and Computer Engineering so I'm not scared of the programming, I just tend to look for non-computational solutions first. A couple weeks ago Novak combined an entire board design of mine into a $0.70 microcontroller and about five interface parts, and it's clearly the better option. These things happen.

https://bitcointalk.org/index.php?topic=902305.0

There is an efficiency chart in the first post of the thread about miners with that chip, which has been on the front page of the hardware forum for four months. That's what I calculate off of.

Right, switchmode is used in just almost everything these days. But a mathematically assured analog hardware solution I will always prefer over a software-only solution, especially where high speed frequency-dependent response is concerned. Why do an ADC sample and FFT and whatever else convolution processing for transfer functions when you can just put proper RC filtration on the line and be done with it? Everyone loves complexity these days, but complexity usually substantially decreases reliability and I'd rather have something reliable than something fancy any day.
Our TypeZero boards will have software-settable voltage control, but not because we have a fancy digital regulator. So far I'm designing around a flexible and reliable analog SMPS chip and probably put a DAC reference on it to do software adjustment. The stick miners don't need software-defined voltage. The complexity for that scale is stupid, so I'm just putting a smal trimpot on instead.

Yeah, efficiency on that circuit would be pretty bad especially for low-duty-cycle applications. Also it's microcontroller-based, which can get tricky with software-defined frequency response if you want it to be really stable. A proper dev and testing of the code would take me a lot longer than just using a TL494 to do the same thing at least as well.

I measured the peak-to-peak ripple on our regulator, and also the peak overshoot, which would be on the charts. Typical instantaneous overshoot (which is a peak maybe 50-100nS wide) was less than 100mV. A bit of topography change and a high-frequency shunt capacitor, and some more bulk capacitance (I found my 47uF 1206 caps after I'd already full-range-tested it with some 22uF) would help smooth it out quite a bit.

One noteworthy thing was when the driver chip shifted into synchronous mode instead of skip/diode mode, usually around 650mV at low loads but as the load currents increased into the ranges these things will actually be running at, that threshold dropped somewhat linearly down below 600mV and by the top-end measurement it was starting in synchronous mode. This had a measurable effect on both efficiency and output ripple, as continuous conduction with a synchronous rectifier is way better than low-side-diode pulse skipping with hysteresis.

All my boards are designed for integrated current measurement. But people (especially in Bitcoin) are cheapskates and usually didn't want to spend the extra three dollars for a feature that wouldn't get used.

I think we've got one of these at the shop. Novak spent more time playing with it, I'll see if he's got any notes on the thing.

You know the BM1384 can do under 0.35W/GH, right? I worked over some numbers a month ago speculating S2 upgrade kits, with suitable chip density they could make a 3.6TH miner off a 1KW PSU with the same power density (and therefore same fan volume) as the S2 if they really felt like it. It's possible to get 0.21W/GH from 28nm if done right, and pretty easy to do with 22nm.

I'd be surprised if a 6-board kit came in under $500, and honestly $600 seems more likely. Looking at around 2.6TH at about 0.3W/GH would be a steal at under $400, when you'd have to put down about $750 for that hashrate (at worse efficiency) from S5.

I believe there's timeout data in the datasheet, for what that's worth. I'll have to clean up the Vcore regulator a bit for you but the thing should be pretty movable from one bench to another. Those breakout boards turned out pretty nice. Remains to be seen if they work properly though.

I guess at some point I should make a chart of the page of power data I got from the regulator. I'll probably make it better and then have to run all the tests again, but today I got output ripple and efficiency info from about 65 current/voltage setpoints before jacking with the BM1384 breakout board.

I'm really hoping there's nothing wrong with the chip itself and there's some trivial reason I'm not even seeing serial return data. I might reflow the chip tomorrow and shift it around a bit, make sure all pads got good solder contact. Any software poking that'll have to be done to cgminer driver will probably have to wait until at least Monday, if not later.

So, the chip is mounted on the board and receiving serial data and running off the Compac regulator. But it's not sending any serial data and cgminer doesn't recognize it. That's not terribly surprising, really. I'm hoping Novak can software up an answer for convincing cgminer to recognize it, or I find yet another hardware issue with my serial level shifter or something.

It did once report an Icarus error wrong device detected something or other, and occasionally it throws LIBUSB stop errors on unplug, but nothing has been consistent.

Anyway, right now it's quittin' time.

Tested the regulator up to 750mV 5A, which should be able to run between 9 and 10GH. Regulator power at that point is 4.34W, so factoring in power for the rest of the stick you're probably looking at 4.5W or so to get there.

At 650mV I tested up to 5.2A, which should run 10-11GH off a regulator input of 3.94W, so probably about 4.2W input. Past 650mV with the .125ohm load, I'm not sure what was kicking on (internal OCP shouldn't have been a factor yet, nor should OVP) but the chip was shutting down then ramping back up rapidly.

I'm still not satisfied with the output high frequency ripple, nor am I impressed with the load regulation. Probably an artifact of the switch noise on the output getting through the feedback compensation is causing the output voltage to increase with output current. I didn't notice it until about 2-3A output, which means it shouldn't be a factor for a stock USB power range, but this thing's supposed to be designed to obliterate USB power if the user wants it to. I just don't want someone to crank up the clock on a particular voltage setting and then find out the voltage jumps up to shutdown point.

It's probably good enough to start initial testing with the chip, though. I'll wire up a breakout board and see if I can't get cgminer talking to an ASIC before clocking out tonight.