Hacking the S7 - improving efficiency through minor hardware manipulation

[BryanC]

Thanks for the additional info... It looks like 630 might not be the ticket either. I've got one board only showing 44 chips, and a half dozen HW errors.

[sidehack]

I'd watch for a HW percentage below 0.01 - that's when I started noticing the effective hashrate dropping out. Make sure to keep fans low too; these boards really like to be kept above about 45C to run stable and at low powers it takes almost no fan to do that, especially with the oversized heatsinks in the LN.

[IITravel01]

Decided to try and see if I can do this myself.  Ordered a Pickit 3 and have a S7 that's not doing anything right now.  

Goal:
To try to get the 3 blades running at the same speed one of my S7-LN with sidehack's mod. is running which is 660/630 mV I believe (I'd initially set the modded S7 at the sweetspot of 550M=806w and increase speed until HW errors start showing up, or at 670/640 mV running at 587.5M=900w).  With the 2 blades/boards on the modded S7-LN it's running at 562.5M at 0.0001% HW errors.  This brings about 2,531 GH/s consuming about 550w, so with 3 blades that would bring about 3,8 TH/s consuming about 825w (of course 575M at 3.88 TH/s and 843w with near 0.00% HW errors would be even better goal at 630 mV).  That connected to a EVGA SuperNOVA 1000w P2 ($180) should be a pretty efficient setup for a S7 without the need to purchase any more splitters or adapters as there is 10 PCIe connectors with the power supply (oddly the T2 model doesn't come with the same cable set).  (If money isn't an objection, Thermaltake has a unique 1,250w power supply for $350 that can monitor the electrical usage via a celphone app. that might be interesting to some miners out there.  You'll have to get 2 of these adapters https://www.amazon.com/Express-adapter-inches-sleeved-Angeles/dp/B005H3KH9E/ref=sr_1_1?ie=UTF8&qid=1468701899&sr=8-1&keywords=EPS+to+PCIe as well as figure out how to convert 4 of the PCIe 8-pin connectors to 6-pin connectors.)

I bet this would become a pretty popular hack to the S7's (batches with 135 ASIC) if it's easy to do and obviously these hacked models would sell for more after you're done with the S7.  Anyone know what a S7 pulls from the wall at 562.5M unmodded?  Edit:  Pulled out my Kill-A-Watt.  It's about 1,175w at the wall connected to a Corsair 1,500w Titanium.  So that means about 350w reduction, that's about 29.8% I think (default on the S7 is 690 mV or 700 mV I believe, that's how they get 1,293w=4.72 TH/s at 700M?).  Who would't rather have a S7 that uses about 30% less electricity at the same speed, even if it did only 3.8 TH/s, unless you've got free electricity?  If BitMain sold the power supply bracket separately, I bet you could sell a S7-LW (Low Wattage) version with the EVGA 1000w P2 together as a set.

[sidehack]

I've got a few I'll be doing similar work to, with intent to resell. I should probably be doing that today but it's Saturday and I want to sit on my couch instead. Probably start on it tomorrow.

At 825W you can get away with 2 cables per board (assuming they're not garbage cables), so a less-well-outfitted PSU can still do the job.

[Gahs]

If this S7 improvement venture is successful, are you going to have an equipment that will perform almost as good as a S9? Because that will mean not having to buy a S9 in long run.

[sidehack]

No, it's really not possible to get the S7 reliably below 0.2J/GH which is still twice the power of the S9. Check the chart on page 4.

[NotFuzzyWarm]

Arrg...
Got the PICkit3 programmer,, dl'd the latest software for it http://pic-microcontroller.com/download-mplab-ide/ All well there, installed to Win7just fine. Problem is with "You'll want to set the IPE "Device" drop-down to PIC12F1572".

Guess what device is not in the drop down list? Closest match the sw will let me do is PIC12F1501 and that throws a mismatched device ID# error when trying to connect.... Possibly that device was removed in a recent 'update"?

Ideas on getting maybe an older device lib for the programmer to use?

Also details: you mentioned disconnecting the 18pin cables to control board.
Done that.
Using the PICkit to power the device with 3.375v.
So, do the hash boards need their normal +12v power to them or is the MCU for each strictly powered from the programmer?

EDIT: Looking at microchips site for the PIC12F1572, http://www.microchip.com/wwwproducts/en/PIC12F1572 its data sheet starts with:
"PIC12(L)F157X microcontrollers combine the capabilities of 16-bit PWMs with Analog to suit a variety of applications"

I'll try that tomorrow as I recall seeing the LF version listed... Really wanna get my my 22 online s7's knocked down in power. Looking to run using 625MHz vs default 700 for clock as that produces 4.1THs vs 4.7THs, should be a good target for when I can get the Vcore reduced

[sidehack]

You could probably get 625MHz at around 660mV, which would be a decent reduction from stock power use. The two I worked up today look to be pretty good; one of them is hitting 550MHz at 630mV with almost zero errors. I couldn't quite keep boards from dropping out at 575.

The MCU is programmed only from the 3.3V line from the controller. No power needs be connected to the hashboards or anything else for programming to work right, but disconnecting the 18-pin will prevent voltage from flowing back from the hashboard into the controller which overcurrents the PICKit.

[VonSpass]

Interesting; I have not been following the BTC mining tweaks lately - good old Bitmain always performs beyond advertised numbers!

[sidehack]

Not sure if anyone's interested, but I'm having good luck working up S7s to run 4TH under 1000W off a DPS1200 PSU so I've got some available in the sales board. I'm working on a deal with a guy who is retiring a mine, so I should probably have more of these in the future.

[NotFuzzyWarm]

Found my problem (no PIC12F1572 device listed): Microchip has at least 2 different programming solutions and the first one I found is version 8.08, the programming software you need is v3.35 found here http://www.microchip.com/mplab/mplab-x-ide

Starting to play with me s7's and as rather expected looks like each board can need different Vcore settings. On my 1st one modded, a batch 16 1-fan, I can run board-1 @ 630mv all the way up to at least 600MHz. For boards 2 & 3 I had to use 650mv. Power dropped from 1,330w down to 995-1,010w and that is with a 120-220v step-up xmfr between my UPS and the PSU.
Most excellent.

[IITravel01]

Found my problem (no PIC12F1572 device listed): Microchip has at least 2 different programming solutions and the first one I found is version 8.08, the programming software you need is v3.35 found here http://www.microchip.com/mplab/mplab-x-ide

Starting to play with me s7's and as rather expected looks like each board can need different Vcore settings. On my 1st one modded, a batch 16 1-fan, I can run board-1 @ 630mv all the way up to at least 600MHz. For boards 2 & 3 I had to use 650mv. Power dropped from 1,330w down to 995-1,010w and that is with a 120-220v step-up xmfr between my UPS and the PSU.
Most excellent.

Thanks for the update.  I ordered my Pickit 3 through amazon and didn't realize it would be shipped from Hong Kong, so it's taking longer than expected.  You answered one of my questions in your post, do different boards need different voltages and can they work at different voltages in the same S7.  What kind of HW errors are you getting?  I'm wondering if different boards need different voltages, then the sweet spot for one board (like 630 mV is 550M) won't be the same as the other (like 650 mV is 600-625M), but this might cause one board to start getting errors (like the 630 mV might not do so well at 600M).  There's no way to set the speed for different boards in the same S7, is there?

[NotFuzzyWarm]

Running different voltages for the boards in the s7 is fine. What I was seeing during tests is that when voltage was to low a board would be stable for maybe 5-10 min and then dropped out. Once the voltage is where the chips on a board are happy then the s7 performs as normal. Once I got the boards dialed in the overnight HW errors for this s7 was 0.005%.

No you cannot use different speeds between the boards in a s7.

[philipma1957]

If this S7 improvement venture is successful, are you going to have an equipment that will perform almost as good as a S9? Because that will mean not having to buy a S9 in long run.

No but a quiet 950 watt machine getting .22 watts still has some value.

My buddies office  can use 1 s-7 and 1 s-7ln this winter  with power at 2.4 cents.

Right now I run 1 s-7ln  and I may grab a modded s-7 from sidehack in early sept.

they would run from sept to may using only 2.4 cent power.

[IITravel01]

http://gekkoscience.com/img_stash/S7LN/S7LN_HEX.zip

The above link is a ZIP archive containing some 17 hex files. The one called "S7LN_PIC_FIRMWARE.hex" is the original unmodified firmware found on the hashboards in my S7LN.

The remaining files are called "S7LN_userselect.X.production.AAA-BBB.hex" where AAA represents the cold running voltage and BBB is the hot running, in millivolts average per node.

I have cold-running files from 580mV to 690mV, where the hot is 30mV higher. I also have three files that set both hot and cold to the same for 700, 710 and 720mV - at those points, a clock rate that would require a higher voltage to start would also draw enough power to smoke out the whole board. Pretty sure anyways. I've tested up to 775MHz and 1020W wall (stock setting was 680W on my unit) without finding an unstable clock on even 710mV and I don't want to catch my boards on fire.

Anyway. If you want to make use of this, find yourself a PIC programer (I use a PICKit 3 that I've had since about 2013) and grab the MPLAB IDE software. What you'll actually need is the IPE side (Programming, rather that Development) but I'm not sure it's possible to download just the IPE without getting the whole thing.

If you remove the outlet fan from the unit, at the end of the boards you will see eight pinholes - one pair, and one set of six. The six are the ISP header into which you'd plug your PICKit.


The lowermost hole of the six corresponds to the header-socket hole on the PICKit that has been flagged (on mine there's a while triangle pointing to it).


You'll want to set the IPE "Device" drop-down to PIC12F1572. Under the "Settings" select Advanced Mode, and once in there click Power, set VDD to 3.3V and check "Power Target Circuit from Tool". For that to work right you're going to want to unplug the 18-pin cable from the board before you program it.

Once that's all configured, plug your PICKit into the board (you might need some wires or pins to connect the header socket of the programmer to the holes on the board) and click "Connect". If you did everything right, there ought not be any errors. Might get a box pop up saying something about low-volt programming and voltage mismatches or some such; just click "OK" and carry on. As far as I can tell it doesn't matter for this application.
You should be able to read existing code off the chip, which you can see in Memory View (View -> Show Memory). If you would like, click "Read" to pull all the existing firmware off and File->Export->Hex to save a backup copy just in case.

Click the "Browse" button associated with the Source path, and navigate to the new hex file of your choosing. Once that's loaded, click "Program" to push the new program to the board.

Make sure you plug the 18-pin cable back in before trying to mine again.

The program will turn on your board to the hot-start voltage value upon power-up of the controller board. The S7's software, once the OS is booted, will run cgminer for about a minute to get things warmed up and then restarts it fresh. My board firmware will start the cold running voltage about 35 seconds into the second cgminer run. If you're watching on a kill-a-watt, the power consumption should drop about 10% or so when it kicks over to cold-running.

When you write "...If you would like, click "Read" to pull all the existing firmware off and File->..."  Do you mean click "Read Device ID" in the "View" upper menu bar?

Do we need to program each board individually (3 of them in the S7) or is one enough (as in are we programming the controller board on top via the board or programming each board)?  I ask this as I'm examining my S7 and notice that of the 3 boards, the header has a few of the holes soldered closed (one has all holes closed and the other has 3 of the 6 closed) on 2 of the 3 boards.  Nevermind, I realize that that doesn't make sense if we have to disconnect the 18 wire cable.  Why would the holes be soldered closed?

[RichBC]

When you write "...If you would like, click "Read" to pull all the existing firmware off and File->..."  Do you mean click "Read Device ID" in the "View" upper menu bar?

Do we need to program each board individually (3 of them in the S7) or is one enough (as in are we programming the controller board on top via the board or programming each board)?  I ask this as I'm examining my S7 and notice that of the 3 boards, the header has a few of the holes soldered closed (one has all holes closed and the other has 3 of the 6 closed) on 2 of the 3 boards.  Nevermind, I realize that that doesn't make sense if we have to disconnect the 18 wire cable.  Why would the holes be soldered closed?

Read means the Read button in the centre of the screen that will read the contents of the Pic on the hash board. This can then be saved using File & Export.

Each of the 2 /3 Boards has to be done individually as there is a Pic on each board. If the holes are soldered closed it means that Bitmain has not properly masked them during the flow solder process. Perhaps if these were early boards they did not program the Pic after assembly?


Rich

[NotFuzzyWarm]

re: holes soldered closed
So far I've ran into that on 3 boards in different miners. Does not follow batches, it just seems random. I first saw that on a b6 and last saw it on a b18. No idea why Bitmain did that.

What I did was just leave those boards as stock. I suppose one could use a solder-sucker to clear the holes...

No we are not programming the BB controller -- we are programming just the MCU on each board.
Yes EACH board has to be programmed but they do not have to use the same settings. eg. a miner can have 2 boards set to 650/680mv and another can remain as stock (holes plugged) or it can be at a different voltage setting that keeps it happy.

[sidehack]

I've also seen it on a few boards; I cleared the holes and programmed them as usual. Most of the miners I've sent out didn't have the same voltage across each board; they were all tested individually for coldest stability at a given frequency and then tested as a whole. It'd be the same concept as running several Compacs off the same hub but each one set to a different core voltage, whatever that stick needed to work best.

[IITravel01]

When you write "...If you would like, click "Read" to pull all the existing firmware off and File->..."  Do you mean click "Read Device ID" in the "View" upper menu bar?

Do we need to program each board individually (3 of them in the S7) or is one enough (as in are we programming the controller board on top via the board or programming each board)?  I ask this as I'm examining my S7 and notice that of the 3 boards, the header has a few of the holes soldered closed (one has all holes closed and the other has 3 of the 6 closed) on 2 of the 3 boards.  Nevermind, I realize that that doesn't make sense if we have to disconnect the 18 wire cable.  Why would the holes be soldered closed?

Read means the Read button in the centre of the screen that will read the contents of the Pic on the hash board. This can then be saved using File & Export.

Each of the 2 /3 Boards has to be done individually as there is a Pic on each board. If the holes are soldered closed it means that Bitmain has not properly masked them during the flow solder process. Perhaps if these were early boards they did not program the Pic after assembly?


Rich

Thanks, I missed that middle "Read" button there on the "Operate" page.

[aarons6]

i picked up one of these modded s7s from sidehack..
with the stock fans i was having trouble keeping it running at night.. even lowering the fans to 20% was not enough to keep the boards warm enough.. once they got to 40c or so they will shut off.

we are having pretty good weather here, but it still cools off at night.. upper 50s at night and 80s during the day.


i put on a pair of AFB1212-SHE fans i got for my s5s.
https://www.amazon.com/Delta-Electronics-AFB1212SHE-120x120x38mm-connector/dp/B004Y1HLA8

these fans do have a bit more of a whine to them than the stock fans.. but the air flow is slower and the s7 doesnt whistle, over all i think they are more quiet. like a s5.

running them at AUTO they seem to be running pretty good at 600 speed. right at 4000 ghs and using 1023 at the wall. my unmodded s7 at 600 speed uses 1280 watts.
they speed up to about 3400 rpm during the day and slow down to around 3100 at night. they max out at 3600 so there is still room for when it gets hotter.
the temps are holding right at 55-64c.



i think you can flash the 1-fan rom on it and just use one stock fan.. not sure tho.. i didnt try. i might try it on my other s7 when i get the pickit3 i ordered from china.


overall the mod is well worth it.. you dont lose much speed (actually i was running my s7 at 600 already) and you lose almost 300 watts. you do have to do something with the fans tho, the boards need to be above 50c to stay stable.