Experimenting with Jalapeno firmware...

[philips]

The pads used have a thermal transfer figure in the teens (14?) paste is rarely even 10 ...

Don't forget the thickness - pads are thicker, so even with higher thermal transfer figure, they are worse than a few microns of paste.
In the Jalapeno case, the pads are preferable, because you may damage the chips when tightening the heatsink with just paste and if the two chips are not perfectly aligned the paste may even not be enough to make good contact with the heatsink

I ordered this one: http://www.frozencpu.com/products/17499/thr-181/Fujipoly_Ultra_Extreme_System_Builder_Thermal_Pad_-_60_x_50_x_05_-_Thermal_Conductivity_170_WmK.html?tl=g8c487s1288

Would that be good enough?

[jelin1984]

Ok you think will better 1mm or 0.5mm?

[philips]

Thinner (0.5) is better for thermal conductivity, but I never used a pad that thin.
I am a bit worry about its...whats the word...flimsiness ?

[KNK]

I think 0.5 should be fine, after all many people are OK with paste, so it seems the chips are well aligned.

When i get my Jalapeno i will be able to measure the original pads, but will still try the 0.5 first even i have 1.0, 1.5 and 3.0 at hand too.

My idea was to somehow connect the FETs to the main heatsink too (hence the additional pad sizes and some cooper plates collected), but it seems they are adding dedicated heatsinks to them now ... will see

[Flashman]

It's not going to be such an issue with teh jallies with only 2 chips, you get far better odds of them being near enough coplanar to use thermal paste for an advantage in thermal interface thickness.

With 8 chips under a sink though, it might get a little more critical to keep the pads.

Personally, I'll be checking mine with a straightedge to see how they lie, before considering pad replacement. It's not outside the bounds of possibility that I could consider making custom shims for mine, on a per chip to fit basis.

One thing that could happen is that for a chip that meets the sink at a slight angle, the paste fillets it, and seems good enough at first, but as it dries out in service it creeps further up the crack leaving half the chip out of thermal contact.

If I were putting together my own BFL based hardware, I miiight consider pressure mounting the ASICs to a chunk of copper, with a thermal epoxy (Which BTW suck as much as any other TIM) then mounting them to the board after, but that could turn into a huge PITA.

If the board were clear of other components though, you could shim each chip, then lap the tops of all the shims coplanar but that's probably impractical.

Also the less thermal interfaces the better, but also if you can get them thinner than they would be otherwise, that's good too, swings and roundabouts.

[danattacker]

It has always been my understanding that thermal paste is superior to thermal pads. But, the only reason I had to use thermal paste is because the thermal pads on mine completely fell apart after removing the heat sink about 10 times. I was also worried about the chips not being coplanar and damaging them, but it seems to work. I'm not using much torque when tightening down the heatsink.

[KNK]

I was also worried about the chips not being coplanar and damaging them, but it seems to work. I'm not using much torque when tightening down the heatsink.

The problem also is, that with the time the paste will dry and every gap may cause local overheating and looser the heatsink - higher the chance.
Can you tell the difference in temp with pads (before) and paste (now)?

[bitpop]

I used a ton of paste and torque. Good enough

[danattacker]

Can you tell the difference in temp with pads (before) and paste (now)?

I couldn't tell you for sure. The temps seem about the same, but they are measured from the PCB and not the chips.

[KNK]

but they are measured from the PCB and not the chips.

oh right, sorry.
I was planing to connect (some of) the sensors (and fets) to the main heatsink with the help of a cooper plate and forgot its not the case with the original board

[erk]

https://forums.butterflylabs.com/announcements/3282-bitforce-sc-firmware-version-1-2-5-a-3.html#post45157

Flashing your firmware will void your warranty yes.

Just thought I would make people aware of that.

[Trongersoll]

https://forums.butterflylabs.com/announcements/3282-bitforce-sc-firmware-version-1-2-5-a-3.html#post45157

Flashing your firmware will void your warranty yes.

Just thought I would make people aware of that.

I'll make sure to keep my top down then. geesh, BFL Prudes.

[kano]

but they are measured from the PCB and not the chips.

oh right, sorry.
I was planing to connect (some of) the sensors (and fets) to the main heatsink with the help of a cooper plate and forgot its not the case with the original board

Yes indeed - pretty much every mining device out that is quite unreliable at measuring a useful temperature.
So the reported before and after numbers wont necessarily give the the correct answer.

[KNK]

https://forums.butterflylabs.com/announcements/3282-bitforce-sc-firmware-version-1-2-5-a-3.html#post45157

Flashing your firmware will void your warranty yes.

Just thought I would make people aware of that.

That's pretty much the case with every device, so it is not a surprise at all.
There is a big fat warning about loosing your warranty when your overclock your GPU too

[Askit2]

https://forums.butterflylabs.com/announcements/3282-bitforce-sc-firmware-version-1-2-5-a-3.html#post45157

Flashing your firmware will void your warranty yes.

Just thought I would make people aware of that.

That's pretty much the case with every device, so it is not a surprise at all.
There is a big fat warning about loosing your warranty when your overclock your GPU too

See I see some of your statement as wrong.
No one said I was going to overclock it. The thermal protections should avoid issues with running too fast. But even then what I want is 1 job per chip. It is faster that way as all but one of my units (8-1=7) have mismatched chips. The worse they mismatch the more benefit from the newer firmware.

PS. Firmware was released 10 days before mine shipped. I have the old firmware. I suppose it being available and installed don't have to be the same thing.

[kano]

https://forums.butterflylabs.com/announcements/3282-bitforce-sc-firmware-version-1-2-5-a-3.html#post45157

Flashing your firmware will void your warranty yes.

Just thought I would make people aware of that.

That's pretty much the case with every device, so it is not a surprise at all.
There is a big fat warning about loosing your warranty when your overclock your GPU too

See I see some of your statement as wrong.
No one said I was going to overclock it. The thermal protections should avoid issues with running too fast. But even then what I want is 1 job per chip. It is faster that way as all but one of my units (8-1=7) have mismatched chips. The worse they mismatch the more benefit from the newer firmware.

PS. Firmware was released 10 days before mine shipped. I have the old firmware. I suppose it being available and installed don't have to be the same thing.

Except the 1 job per chip is bad due to the latency it incurs.
A single 4GH/s chip takes ... 0.93s to complete a nonce range, and the BFL MCU (with the same design flaw as the BFL FPGA) doesn't return results until it completes the work item.
(Note the latency/work loss factor in the 1 job per chip design is multiplied by the number of chips you have - more chips, the worse it is)

It's also not necessary.
(though I've no idea if Luke will take notice of how I suggested it's easy to fix this ... divide the work up so they all complete at roughly the same time - very simple maths)

So, although the gain from 1 job per chip is greater than the loss it incurs on any pool other than p2pool, it's still far from optimal (and optimal is not difficult at all)
However what this means for you is you would again want to re-flash it later when it's fixed properly

[Askit2]

I see your point. My point had been that with the frequent 10% mismatch there would be a fair (5%) gain overall and the double loss at LP wouldn't be more then the gain. Even then anything less then 20 seconds would be a loss, over 40 seconds I would be ahead.

I also see how with trivial maths I could keep both chips un-hobbled and a low time for work to finish.

I guess what bothers me is that it could have shipped ~5% faster on 7 of 8 Jalapeno's and for no cost to the manufacturer. Just flash the current firmware and life is good. Instead of keeping connectors uncovered BFL built the heatsink over them. Then to make matters worse I could flash on a 5% gain but it wouldn't keep the warranty. I intended to keep the tests on and likely keep the units running around 5.5Ghash as I want them quiet. But there isn't a good way to get everything. Now I have to figure out what the warranty is worth. So far I have 1 failed fan and 1 failed power supply. I emailed and haven't heard anything back.

[bitpop]

Well they did over delivery 1 ghs and as much as I hate bfl, they even allowed further free overclocking.

[danattacker]

I just got through spending most of the day messing with the jalapeno.

I made a custom peltier assembly and a temperature controller that regulated the temperature to around 65-68 F. I didn't want to go colder due to risk of condensation.

I also modified the firmware to make the same changes as ckolivas did in his, except you can just comment out in std_defs.h...

Code:

#define __RUN_HEAVY_DIAGNOSTICS_ON_EACH_ENGINE

...and it should have the same effect. I also reduced

Code:

#define __TOTAL_DIAGNOSTICS_RUN 10

...down to just 1. After these changes, I was able to get 30 engines running from 28 previously.

I also increased the voltage from 1.016v to 1.047v by changing the resistor at R11 to 7.5kohm. This resulted in the chips automatically clocking to a higher frequency. At frequency index 7, I was previously attaining 291MHz/271MHz on the chips. After the voltage increase, it was clocking at 305MHz/291MHz.

After these changes, my hash rate went up from ~8.2 to 8.3 GH/s at frequency index 7, to ~8.85 GH/s. I was able to achieve ~9.05 GH/s with index 8 (~4.5% errors), and ~9.23 GH/s with index 9 (~7.5% errors).

Unfortunately, the peltier assembly didn't seem to have much effect on the error rate. I put the old heatsink on and it handled the heat just fine. The only thing the peltier was good for was getting a high initial frequency setting. With the peltier drawing ~90w, it wasn't really worth it to continue using.

Right now I'm running frequency index 8 @ 8.93 GH/s (~2.6% errors). It is possible for me to go higher at that index, but that was when I was using the peltier. I stuck the unit in the freezer, but I guess I didn't get lucky.

Some pictures of the peltier assembly...

[goxed]

I just got through spending most of the day messing with the jalapeno.

I made a custom peltier assembly and a temperature controller that regulated the temperature to around 65-68 F. I didn't want to go colder due to risk of condensation.

I also modified the firmware to make the same changes as ckolivas did in his, except you can just comment out in std_defs.h...

Code:

#define __RUN_HEAVY_DIAGNOSTICS_ON_EACH_ENGINE

...and it should have the same effect. I also reduced

Code:

#define __TOTAL_DIAGNOSTICS_RUN 10

...down to just 1. After these changes, I was able to get 30 engines running from 28 previously.

I also increased the voltage from 1.016v to 1.047v by changing the resistor at R11 to 7.5kohm. This resulted in the chips automatically clocking to a higher frequency. At frequency index 7, I was previously attaining 291MHz/271MHz on the chips. After the voltage increase, it was clocking at 305MHz/291MHz.

After these changes, my hash rate went up from ~8.2 to 8.3 GH/s at frequency index 7, to ~8.85 GH/s. I was able to achieve ~9.05 GH/s with index 8 (~4.5% errors), and ~9.23 GH/s with index 9 (~7.5% errors).

Unfortunately, the peltier assembly didn't seem to have much effect on the error rate. I put the old heatsink on and it handled the heat just fine. The only thing the peltier was good for was getting a high initial frequency setting. With the peltier drawing ~90w, it wasn't really worth it to continue using.

Right now I'm running frequency index 8 @ 8.93 GH/s (~2.6% errors). It is possible for me to go higher at that index, but that was when I was using the peltier. I stuck the unit in the freezer, but I guess I didn't get lucky.

Some pictures of the peltier assembly...

Thank for the notes and your cool experminent