Hacking a BFL Jalapeno to 32GH and beyond....(???)

[miaviator]

so bfl is selling 10G jallys at $574?  If those are three chips we could get what 18-30Ghs out of them just with better cooling and an update?  or am I way off base here?

[bcp19]

so bfl is selling 10G jallys at $574?  If those are three chips we could get what 18-30Ghs out of them just with better cooling and an update?  or am I way off base here?

They're $374 on the website.  I doubt you could push them to 18+ with only 3 chips, for the most part they seem to run at about 5GH max per chip.

[HellDiverUK]

P.S. the FW version written on back is 292..... hmm board rev. C same as others. This has to be a FW issue somehow....

My "7GH" Jalapeno that I finally got on Christmas Eve (late June order) came with 292 on it.  It's running around 8GH.

[HellDiverUK]

I notice on BFL's site they mention some sort of 'teaming' system:

Additional Bitforce SC products can be added to the chain via a USB hub for linear performance multiplication with no overhead cost. Each additional unit is auto-configured and folded into the workforce without any user intervention required.

I'm using BFGMiner, the two Jallys I have are running off a single USB hub.  What am I looking for in regards this 'teaming'?  Is it important?  Does it actually help?  How do I know if it's working?

[lightfoot]

so bfl is selling 10G jallys at $574?  If those are three chips we could get what 18-30Ghs out of them just with better cooling and an update?  or am I way off base here?

They're $374 on the website.  I doubt you could push them to 18+ with only 3 chips, for the most part they seem to run at about 5GH max per chip.

4gh is the most you will get out of a chip in normal operation. The Chili I have gets 5 (39gh total, 37gh submitted after errors) but that is using exceptional chips, is running Mr. Teal's software, and adjusts core speeds per chip based on chip temp and actual chip frequency. It is amazing.

Right now my 3 chip turbo jally is getting 12gh, and my 4 chip is getting 15.9, and my five chip jally with a really crummy chip is getting 19gh. So it depends.

This morning's job is going to be boosting my 4 chip to 5 and maybe flashing the 12gh one to see if it goes faster or slower with the level 9 code.

C

[lightfoot]

By the way Drug: 1.3% errors is really pretty reasonable, have you tried reflashing it with a speed of 5 or something?

All my new jallies have two chips, so yours may be a special case. However I'm curious to see if you can flash any of the others, the boards do not look different and can take chips with this software load.

C

[Photon939]

I notice on BFL's site they mention some sort of 'teaming' system:

Additional Bitforce SC products can be added to the chain via a USB hub for linear performance multiplication with no overhead cost. Each additional unit is auto-configured and folded into the workforce without any user intervention required.

I'm using BFGMiner, the two Jallys I have are running off a single USB hub.  What am I looking for in regards this 'teaming'?  Is it important?  Does it actually help?  How do I know if it's working?

It just means that if you have two 5GH/s Jalapenos and you plug them into one computer your hashrate is 10 GH/s. It's just marketing jargon that says performance scales linearly with the amount of units you buy.

[HellDiverUK]

I notice on BFL's site they mention some sort of 'teaming' system:

Additional Bitforce SC products can be added to the chain via a USB hub for linear performance multiplication with no overhead cost. Each additional unit is auto-configured and folded into the workforce without any user intervention required.

I'm using BFGMiner, the two Jallys I have are running off a single USB hub.  What am I looking for in regards this 'teaming'?  Is it important?  Does it actually help?  How do I know if it's working?

It just means that if you have two 5GH/s Jalapenos and you plug them into one computer your hashrate is 10 GH/s. It's just marketing jargon that says performance scales linearly with the amount of units you buy.

I thought is was more than that.  I did check bfgminer more closely, and it's saying something about 'parallel queue processing' which I'm thinking sounds a bit more like what BFL were alluding to.

[lightfoot]

\
I thought is was more than that.  I did check bfgminer more closely, and it's saying something about 'parallel queue processing' which I'm thinking sounds a bit more like what BFL were alluding to.

Well, yes. BFL runs the chips as units, within each chip is up to 16 sub processors that the work is farmed out to. So each chip shows up as a computing node, not the sub processors.

Chili takes this a slight step further by dividing all the work to all the chips as one big shot. Thus it looks to bfg like one big-ass chip. It's more efficient in some ways, probably about 4gh extra per 8 chips I think.

C

[lightfoot]

Well, more chips in the boards. Because I am waiting on heat sinks (always something...) I am running my crummy jally (the one with one temp sensor, the one that blew up the FTDI chip, etc etc etc) with a total of five chips, 4 running and one not. The not was the one I need to reball; I really should have taken it off the board. Because board temp is at 73.

Now it was going to 80, but I put this super fan+heat sink+pipes on the unit and it's down to 73, which is good but still high. And it's only hashing 14gh because at least one of the factory chips was crap (it hashed 6gh stock). Oh well, you take what you can get in this world, and this keeps me from selling the poor little thing on Ebay.

In fact one of the side fails of all this is I am not going to be able to sell this stuff on Ebay; it's requiring bigger sinks and a lot of thought. In retrospect this will probably be a mistake but fuck it, such is life. Actually I could turn off the extra chips in software and sell them in stock boxes as three chip jallies. We'll see.

I have two cooling units on the way; another 120v unit and a $70 corsair one with two fans and a big heat sink. That's going on the 7-8 chip unit I think... :-)

Never dull.
C

[lightfoot]

Just a quick recap: What do I have here:

Jally 1: 19gh unit, 5 chips, the first jally. Running with a BFL type 1 sink, runs great.
Jally 2: 14gh unit, 4 chips. This is the crap jally, one I bought from a UK person for full value that turned out to have the blown FTDI chip.
Jally 3: 16gh unit, 4 chips. One of my new jallies, 4 chips in a cross pattern.
Jally 4: 12gh unit, 3 chips. The other new jally, I don't want to push the power supplies if I can help it.
Jally 5: 8gh unit, 2 chips. This is a stock "control" jally for a friend.
Chili 1: 36gh unit, 8 chips. Bought for .7btc, this is a cool little thing.

At this point I'm getting good enough to start thinking about offering this as a service. I have done 8 chips in a row here with no problems, and the combination of flux, heat, preheat, and my loupes to ensure every chip is perfectly on the balls has reduced my failure rate to zilch.

Need to figure out a price though; going from a 2 chip to 3 chip jally seems to add about $200 in value on Ebay. No one has sold a 4 chip jally.... So message me; what would be a fair price in your mind to turn your jallies into more money?

Any thoughts?
C

[Drug5bitz]

I'm not worried about the HW errors on that one look at this lol. 2x4chips 1x6chips 1x3chips and 1x2 chips. My six chip carefully hand picked for being the most stable coldest unit. Adding chips on these gravy with a steady hand. I wouldn't suggest above 3 chips on the blocks. I vaguely recall you saying the front light would blink if the chips didn't have enough current. My 3 chip will only run on speed 1 only! I have it on my PSU now and thought maybe it could handle a speed boost.  Here's what I have from last night.

[lightfoot]

Note to all: I am going straight to jalapeno hell :-)


This is the 4 chip+1 dead chip "Danger Jally with a much... bigger... cooler. 5 chips of heat.

Temps went from 80c with the stock fan down to 38c and holding. Now wait till I get water cooling. :-)

Note: You do have to do a bit of stuff. First you need to bevel the edges so it will not touch the caps right around the chips. Then you need to drill and tap two new holes, try to center it as much as possible so you don't hit one of the heat pipes with the drill.

But man, this solves the heat problem. I'll try taking this one up to seven chips as soon as I get the little heat sinks for the FETs.

C

[MrTeal]

If you want to go crazy, I'd love to see you use something like this and cover the whole back side, instead of little press-on heatsinks (with less than ideal tape).
http://www.ebay.com/itm/2x-New-Large-Rectangle-Aluminum-Heatsink-w-screw-mount-DIY-POWER-AMPLIFIER-LED-/251261395030?pt=LH_DefaultDomain_0&hash=item3a8058a056

Drill some clear holes to mount on that big heatsink for the ASICs, and drill and tap some holes to screw the PCB down to the backside heatsink.
I think you'd have no problems running 7-8 chips with a setup like that.

[lightfoot]

If you want to go crazy, I'd love to see you use something like this and cover the whole back side, instead of little press-on heatsinks (with less than ideal tape).
http://www.ebay.com/itm/2x-New-Large-Rectangle-Aluminum-Heatsink-w-screw-mount-DIY-POWER-AMPLIFIER-LED-/251261395030?pt=LH_DefaultDomain_0&hash=item3a8058a056

Drill some clear holes to mount on that big heatsink for the ASICs, and drill and tap some holes to screw the PCB down to the backside heatsink.
I think you'd have no problems running 7-8 chips with a setup like that.

Hm. I have thought about cutting up an old 486 heat sink; the key here is surface area.

What would really win would be water block cooling of course, and one of the reasons I bought a big corsiar block is the thought that I could mount it on the back of the board (the large bracket *just* clears the sides of the LS board) and then just put the stupid AL heat sink on the top.

The full surface area cooling ability of a block might then allow me to cool the whole bottom of the board including that massive ground plane. Since I'll move it forward as much as possible, it might be able to cool the fets from the bottom, and remove the heat sink need on top.

We shall, as they say, see. But I'm convinced that water cooling is the way to go.

[freddyfarnsworth]

Yeah... "Does not ship to Poland"

Who, Lent? Send him a private message, he seems to be a good egg.

C

I'm also interested, but he wont ship to Canada and I cannot even message him on Ebay (it wont allow me somehow?).
Do you know if he's on this board and what nick?
Thanks.

That's odd. I can email him myself and ask if he will ship to Poland and Canada. This is a big world of course.

C

Try using a US address, ebay has prolly blocked countries that have blocked shipping from even communicating with seller to stop spamming him with to many questions.

Use No-ip or something like it to get a us addy then set your pc to be no-ip's real destination.

[Brassguy]

It's been a while but here's an update...

I can confirm that having a dead chip lowers the output (shares produced). However your reported hashrate via cgminer will be consistent with whatever number of chips you have... for example...

I have 3 jallys with 6 chips installed. One was only able to initialize 5. Fair enough. It was hashing 19.5gh average for one week as reported by cgminer. I figured I would get around to reflowing it this weekend. I did a little math before I fixed it and found this...

It was only producing shares consistent with a hashrate of 15.4gh... which coincidentally was the approximate hashrate before chips 5 and 6 were added.

Another interesting fact is temp reported by cgminer when it was at "19.5gh" is the exact same as it is now after it was fixed ~58C at 23.4gh.... I only reflowed the chip that wasn't initializing...

Thoughts?

[Drug5bitz]

I'm more familiar with BFGMiner, but can you check the individual chip speeds and the voltages to them? Sounds like you got a junk chip. Trust me one is a wrench in what ever it touches. Individual info can help you dissect it far more.

[lightfoot]

Another unit up to 20gh. To minimize heat near the FTDi chip and the FETs I have been placing 5 chips as follows:

2,3,4,7,5 (use the map in the previous pages for locations). 60c temp with stock heat sink and a very warm power supply. Wish I had my darn PCIx cable from BFL....

Otherwise seems to work well. Tomorrow I should be having 4 more chips come in and the water cooler, we'll try for 7 on one jally with water cooling...

[Drug5bitz]

Another unit up to 20gh. To minimize heat near the FTDi chip and the FETs I have been placing 5 chips as follows:

2,3,4,7,5 (use the map in the previous pages for locations). 60c temp with stock heat sink and a very warm power supply. Wish I had my darn PCIx cable from BFL....

Otherwise seems to work well. Tomorrow I should be having 4 more chips come in and the water cooler, we'll try for 7 on one jally with water cooling...

I believe you've seen the molex to barrel adapter that takes 5 minutes to make. It's a lot safer power if you have the PSU for it. Yellow 12V line to the inside shielded wire, and the two center grounds on the molex to the outside braided ground.  I wired up 4 of these just to get my miners up while waiting on cablez. Not that he took long, 2 days maybe on his end and an amazing looking product!! 

So in the end you're going to hit 8 chips right? Have we thought about cooling the FTDI chip with a simple sink? If you need any help with water cooling let me know, I love the stuff. I didn't go that route because of ROI. Just be super careful if you drill into the water block not to permeate the chamber. Don't mix metals between sinks and radiators ETC. ionization will happen and slowly deteriorate everything. It's usually suggested to run the system for hours to leak test. The thermal conductivity of water exceeds air something like 10x so this makes me giddy.  Air vs water let my processor go from 4.0 Ghz to 5.2 Ghz in the end it is usually efficient enough that the next hindrance is hardware related as there should be plenty of overhead for heat.