Topic: GekkoScience BM1384 Project Development Discussion - page 86. (Read 146665 times)

Bitmain could probably do it if they wanted, but with a fixed-voltage string it comes down to balancing high hashrate versus chip density. When I first ran the numbers for a potential S2 upgrade kit I was eyeballing something like 3.5TH off 1KW with 100W blades, but their prototype was more like 2.4TH off 6 blades at about 700W probably because you could do it with a lot fewer chips. To get low power means clocking the chips at their bottom end, and to get high hashrate at bottom clock means a jillion chips.

Please have my bitcoin babies if you're telling me we can push about half the w/gh compared to current miners. That's like 40% better and it is beautiful

Don't play with my emotions.

If I can work out strings at those voltages, it means the bottom-end efficiency for the TypeZero should be comparable, if not better.

Nicely done sidehack.  Thanks for all the work.

Love seeing the .3X's on efficiency.  That would make a nice miner

my lousy eyes skipped a link and blended in

8.25
9.63

getting

8.63

Which is why I skip a line when typing.

Yeah, it's pretty much what I was expecting. The Amita numbers should be slightly better, since the regulator will probably be around 90% efficient at 1.2-1.4V and auxilary losses (CP2102, LEDs and such) will be constant per stick, so halved per chip.

Also, 150MHz is 8.25GH, so yes 8.63GH is better than what's possible and whoever told you those numbers is wrong.

Tomorrow my priority will be to finish the Compac PCB design and give it a good once-over. Hopefully we can get those ordered on Friday. If possible I'll draw up a two-chip breakout board so we can test parallel and string better, which will be good for Amita and TypeZero work. While waiting for that stuff, I'll probably shift back to inline regulator design and stuff for the TypeZero. Depending on how much I like the buck chip on the Compac, I may try to build around it for the TypeZero as well.

So, I ran efficiency lines for 675mV and 700mV today. But it's super boring and I could discern general trends - plus running above 700mV is not "recommended" as at that point you can already push a chip far enough to draw over 1A from the USB. Suffice to say, it appears to work well.

And now the moment you've all been waiting for - actual performance data. The breakout board I used for testing has a bad RF pin so the LED doesn't work properly, but other than that sporadic flashing load (which is probably more than made up for by the LEDs on the USB/UART adapter) I have what should be a pretty accurate mockup of the final Compac circuit running behind my USB meter. Apparently the U3 code in cgminer only recognizes 100-250MHz so that's the range I tested in until we are able to alter the code.



Specified hashrate vs measured power for 600-675mV. Testing above this voltage range was moot, as I could not take the hashrate higher than 250MHz (13.75GH) with the software available.



Here we have a chart of the actual measured W/GH for the Compac. I was surprised but I did get the chip to light up 200MHz (11GH) off 625mV so I'll leave it there for the night and see how it works. This corresponds to 3.52W or 0.32W/GH on the performance data. My ideal stock setpoint, 150MHz 600mV, yielded me 511mA draw which puts it at 2% higher than USB power technically allows. I think the final board should be slightly more efficient which may pull this into acceptable range though 511mA is actually probably good enough. If we're able to adjust the code for arbitrary clocks, 8GH corresponds to 145MHz which should pull about 495MW and I'd just stick it there for defaults.

I did get the chip to start at 150MHz off 595mV, which dropped my current consumption to 504mA as measured. And once started, the running voltage should be adjustable to a bit lower setpoint.

The turn-on current burst for higher frequencies continues to be a killer; I need to adjust the overcurrent limit resistor on the regulator board because when I tried to start 675mV at 200MHz and above it actually dropped out and had to be power-cycled a few times before it survived the burst. I think I saw the bench supply tap 2.4A out briefly, which is pretty toasty.

So yeah, here's the raw data.

600mV	GH	Iin	Pin	W/GH
100MHz	5.5	372	1860	0.34
125MHz	6.88	448	2240	0.33
150MHz	8.25	511	2555	0.31
625mV
100MHz	5.5	390	1950	0.35
125MHz	6.88	464	2320	0.34
150MHz	8.25	544	2720	0.33
175MHz	9.63	620	3100	0.32
200MHz	11	704	3520	0.32
650mV
100MHz	5.5	410	2050	0.37
125MHz	6.88	495	2475	0.36
150MHz	8.25	575	2875	0.35
175MHz	9.63	660	3300	0.34
200MHz	11	760	3800	0.35
225MHz	12.38	850	4250	0.34
250MHz	13.75	945	4725	0.34
675mV
100MHz	5.5	440	2200	0.4
125MHz	6.88	530	2650	0.39
150MHz	8.25	610	3050	0.37
175MHz	9.63	715	3575	0.37
200MHz	11	810	4050	0.37
225MHz	12.38	915	4575	0.37
250MHz	13.75	1020	5100	0.37

It's not difficult, especially if you have a dummy hub without built-in power protection. Cut the 5V trace from the upstream jack, add beefy wire (preferably in a star pattern) from your power input to the 5V pin on each port. Ideally, make sure there's some good buffer caps (470uF or so) near each port. There's surely enough information running around here (and who knows where else on the internet) from the last two years of stickminers being a thing, so I'm not going to gather and encapsulate information that's probably already readily available.

We're definitely going to add power pads to the Amita for running in external 5V at higher current. I haven't decided yet if we'll do that on the Compac, but probably not. Should be able to get around 16GH out of it off the 1.5A the USB jack is rated for, which is pretty toasty. Stock settings for the Amita will probably be in the 800-1000mA range, but it'll probably be built to take up to 2.5-3A if you can keep it cool. That'll be a step for another week, is recalibrating my regulator prototype for 1.2-1.6V output and seeing how well it behaves.

If these units don't have a way to provide power from an auxiliary source, then I hope you will put together a tutorial on the proper/correct methods of modifying existing hubs to handle the top end power potentials.
I've seen people suggest all kinds of things from adding crystals to resistors or just plain 16AWG wires over the positive traces.
With no mention of protecting the desktop from being hit with power from the aux. source.

 This pretty much makes it a perfect stick for me.  My hubs do 900ma per port.

I've already got code written to calculate the PLL hex values for close-to-arbitrary clocks. Software's Novak's job, and if we implement a cgminer driver for these instead of just using the U3 driver already there, we can put that code in it and get about an order of magnitude more granularity in clock setpoints.

Right now it looks like 11GH should be possible on 750-800mA of hub power. That's sturdy, but not difficult.

thats flexible for a 1 chip usb stick should allow for lesser hubs to do well with underclocks to 6.875gh or 5.5 gh  with better hubs doing 8.25  9.625 maybe squeeze out the top end at 11gh

Well, I have the regulator marked up to 650mV and 8.9A output. I've used the numbers with some of Bitmain's BM1384 numbers to get some updated extrapolations for performance data. The regulator's not running as efficient as I'd hoped (I was hoping to break 87% peak but the best I saw was 85%). The board-level efficiency data is assuming 100mW toward auxilary stuff, which may be an underestimate.

Actual regulator efficiency data based on this afternoon's measurements:


As you can see, the peak tends to be around 3.5A but efficiency between 3.5 and 8A is still pretty respectable.



This is an extrapolation of expected hashrates versus power required from the USB. Each line is peaked at the rated maximum hashrate for that voltage, as given by Bitmain's chip performance chart, and the power requirement is solved from the regulator input vs output power, Bitmain's chip-level efficiency numbers, and the expected 100mW maintenance power.



This is basically another visualization of the above, with the extrapolated hashrates divided out for board-level hashing efficiency (W/GH). The range is somewhat deceptive, as with the current driver the minimum operating frequency is 100MHz so a minimum hashrate of 5.5GH; it also has a minimum increment of 25MHz, so 1.375GH steps - 5.5, 6.875, 8.25, 9.625, and 11GH for these chart ranges.

We've got a meeting in the morning so I'm not sure when I'll get started on working over the numbers for 675mV through 800mV, but if I can keep getting the performance out of it that I'm seeing so far, I'm definitely going to approve the design and finish putting it in the Compac PCB for prototyping. I'm fairly certain the Compac regulator will be slightly more efficient and/or reliable than this prototype, as I'll be a bit more careful with the layout, and the output will be directly tied to the chip. One thing I noticed when testing was load regulation wasn't so great; going from 500mA to 8A output typically required an increase of 30mV of measured output voltage to stay in range for the measurement. I took my voltage measurement from the regulator board right at the output socket, but the Vsense line on the regulator chip is tied closer to the inductor so it's possible some voltage was dropped in the output trace (though I can't see it being 30mV). The chip actually has a sense pin for GND as well, which was tied locally, so it wasn't taking into account ground-plane drops either. The input voltage to the regulator board was measured below the USB jack, so that does take into consideration any terminal impedance losses and such.

Once I get the regulator marked the rest of the way up, I'll post some updated charts and then rig up a full mockup of the Compac to get actual expected performance data for board-level hashing efficiency. That'll probably take a couple days to get all the data, since for each voltage level there'd be potentially a dozen workable hashrates (and that's if we don't implement my generic clock code so we're stuck on 25MHz increments, min 100MHz) and I'll need to keep a watch on error rates.

Sorry, I don't have any photos of the test rig all set up. Maybe tomorrow. The only additions from already-seen stuff are the jerryrigged FET dummy load and my worked-over scope probes. It is kinda cool to see everything socketed together though.

if you meet the 2.5w to 8gh goal you will be at .3125 watts a gh. pretty nice to see it.

I laid down plenty of solder in the parts that are going to be handling upward of 10A, especially the parts I want low resistance to keep down ripple voltage. The corner is cut on the regulator board because there's a capacitor on the BM1384 breakout board it needs to accommodate.

I do have a scope, two in fact though one's trigger functoin doesn't work. It's about 30 years old. I have a 50MHz Rigol. I'd like to get a 4-channel one of these days; it'd make the efficiency testing I'm doing right now a lot quicker. I'm having to swap probe pairs to measure each meter board. That does cut down on ground-loop errors though, which is good, but it's certainly not making the task fast.

I have completed 550mV and 575mV curves from 500mA up to about 8.5A out. Now just to get 600-800 in 25mV increments, at 500mA output steps. Something like 160 more data points, each one recording 2 numbers, switching probes, recording two more numbers, then doing 5 calculations on those numbers to get actual current, actual power and efficiency. I might be done today.

Also, the 550mV and 575mV both peaked at just shy of 83% at 3.5A output. If the higher voltages stretch that a bit (say, 86% at 625mV) that could put us pretty close to the goal of 8GH off 2.5W input.

oh man! lookies at the soldering, you obliviously subscribed to the webmail "More the better!" and what with the edge cut on the big guy? trying to let the electrons squirt out?!

got a oscilloscope? love to get you one of them 160GSamples/s 20GHz but all i can think i could round up is a cheap pocket 20MHz 48MSa/s thingy..

I'm working up a one-FET dummy load right now, nothing fancy active control, just a multi-turn pot on the gate. Coupling that with the two power meters (source and load) I'll be able to get efficiency curves for the regulator pretty quickly and I should have some fresh extrapolations for the hashing efficiency later today. And yes, I'll take more electroporn pictures for y'all.

I'll admit, all the time wasted trying to iron out the old regulator circuit was fairly disheartening. But I'm much more confident about this one, partly because it's less complex and more modular - both of which things I really like in a design. I'm starting to wish I'd gone this route to begin with instead of looking for a monolithic solution. It does save quite a bit of space, but at the cost of plenty of suck, apparently. And there are almost no options for going below 0.6V natively in a monolithic buck regulator.

side hack and novak i am really enjoying seeing the updates and seeing the progress is looking very nice and promising cant wait to buy one or a few of these to use and would love to test this out from a novice user's prospective