Pages:
Author

Topic: Hacking the S7 - improving efficiency through minor hardware manipulation - page 13. (Read 26494 times)

newbie
Activity: 28
Merit: 0
One problem with having a hot and cold setpoint are, since the hot setpoint is only delivered when the miner is first powered up, it's entirely possible that simply restarting cgminer after a board drops out will actually cripple the whole machine. It really requires a power-cycle - more specifically, it requires cutting 3.3V to the hashboards which would require either power-cycling the controller or some external hardware.

Since the micro listens on the TX line, I could have it just check in for traffic and if it detects there's no hashing going on (so, cgminer was killed) it resets back to hot setpoint to be ready to restart mining. I'll have to look into exactly what traffic gets sent when cgminer is still operating even though the board is down to know how to detect that condition, and I don't really have time for a while.
Possibly the code could increment the cold setpoint every time mining fails and is restarted. One problem with that is you couldn't optimize per-board since both boards would detect cgminer restarting and both increment even if only one needed it. Also, unless it was saved to EEPROM, the new setpoint would be lost every power-cycle. If it was kept, without a means of resetting back to defaults, over time the setpoint will drift higher and higher due to internet downtime, manual reboots from software and other issues.

I would say once you have some time the traffic might be a good place to get some more reference from. I really like the idea of hot/cold script If only there was relm for small flash memory instead of EEPROM.

legendary
Activity: 3374
Merit: 1859
Curmudgeonly hardware guy
One problem with having a hot and cold setpoint are, since the hot setpoint is only delivered when the miner is first powered up, it's entirely possible that simply restarting cgminer after a board drops out will actually cripple the whole machine. It really requires a power-cycle - more specifically, it requires cutting 3.3V to the hashboards which would require either power-cycling the controller or some external hardware.

Since the micro listens on the TX line, I could have it just check in for traffic and if it detects there's no hashing going on (so, cgminer was killed) it resets back to hot setpoint to be ready to restart mining. I'll have to look into exactly what traffic gets sent when cgminer is still operating even though the board is down to know how to detect that condition, and I don't really have time for a while.
Possibly the code could increment the cold setpoint every time mining fails and is restarted. One problem with that is you couldn't optimize per-board since both boards would detect cgminer restarting and both increment even if only one needed it. Also, unless it was saved to EEPROM, the new setpoint would be lost every power-cycle. If it was kept, without a means of resetting back to defaults, over time the setpoint will drift higher and higher due to internet downtime, manual reboots from software and other issues.
newbie
Activity: 28
Merit: 0
Not really. It was a "hey this works but it's 7PM and I want to go home and eat lunch so I'll leave it on overnight" kind of test, and when I came back in the next day it had dropped out and the lifetime hashrate average was a bit over half expected so I extrapolated about how long it ran before stopping while I was not there to observe it.

I think the problem with the 610mV profile is probably balance. By the BM1385 datasheet, the chips should run 425MHz at 600mV. However, since the voltage is dropped per node according to the current use at that node, there is no perfect balance. If one node is dropping even 2% lower than the average 610mV setpoint, it's gone down to 590mV which I've found is in a range 28nm ASICs don't particularly like. The 620mV setpoint gives more leeway in node imbalances at the cost of slightly higher per-unit power use, so for now I'm recommending the 650/620 profile as the best stable bottom clock.


Would be interesting to see if there would be a possiblity of some kind of "early warning" script that would detect the drop in volage for the node and adjusted the setpoint, haveing looked at the data sheet for BM1385 I am still wondering what other issues might arrise from the voltage drops if you persisted at the higher rates.

fantastic reading and some good development in this thread, Ill be keeping a keen eye on the progression of your project!

Thanks!
legendary
Activity: 3374
Merit: 1859
Curmudgeonly hardware guy
Not really. It was a "hey this works but it's 7PM and I want to go home and eat lunch so I'll leave it on overnight" kind of test, and when I came back in the next day it had dropped out and the lifetime hashrate average was a bit over half expected so I extrapolated about how long it ran before stopping while I was not there to observe it.

I think the problem with the 610mV profile is probably balance. By the BM1385 datasheet, the chips should run 425MHz at 600mV. However, since the voltage is dropped per node according to the current use at that node, there is no perfect balance. If one node is dropping even 2% lower than the average 610mV setpoint, it's gone down to 590mV which I've found is in a range 28nm ASICs don't particularly like. The 620mV setpoint gives more leeway in node imbalances at the cost of slightly higher per-unit power use, so for now I'm recommending the 650/620 profile as the best stable bottom clock.
newbie
Activity: 28
Merit: 0
This is a really interesting project you have here,  running @ 610mV - 450MHz @ 8hrs did you have recorded temps or any other recorded data you could share with us?


sr. member
Activity: 338
Merit: 250
Well the one blade's been running solid at 620mV (9.3V, down from 10.0V stock) and 450MHz (1TH) for about 15 hours now. This afternoon I'll do some more testing of the absolute bottom end efficiency and overall power use, generate some charts and plots for y'all. I'll go ahead and work up a stock S7 as well.

Just wondering, would anyone be interested in a group buy of S7 or S7LN pre-flashed to the desired improved efficiency setpoint for a moderate markup (maybe $20-30), tested and reshipped within the US within one or two days of arrival at my shop?

I'd go for a S7-LN or two.  I'd love to see how efficient that could be with a Titanium power supply (I've got a EVGA 1,600w and a Thermaltake 1,250w I'd test with, as well as a 220v step up converter).
legendary
Activity: 3374
Merit: 1859
Curmudgeonly hardware guy
I'm questioning the stability of the 610mV setpoint. My first extended test last week was 610mV 450MHz, and it kicked out after about 8 hours. I ran 610mV 425MHz overnight and it kicked out after about 3 hours. But running 620mV 450MHz ran all day with no issues. The 620mV is about 0.22J/GH versus just under 0.2J/GH for 610 so I wish it was stable but I'm not so sure. Might be 620 is the best we got.
legendary
Activity: 1456
Merit: 1000
It looks like my 640/610 profile is just fine for 400MHz, delivering 1.8TH at 340W for an efficiency of 0.189J/GH

I'm testing 425MHz on it now, and it's looking good. If it runs stable all night (I'm at zero errors five minutes into cold voltage), I can say 1.9TH at, according to my meter, 366W for 0.191J/GH. So this S7LN will go below 0.2J/GH machine-level. I was having trouble with the 630/600 profile doing anything right, but I'll check and make sure I just didn't screw up the code.

By the way, those numbers are for fans fixed at 20% (1560RPM); peak temp is 47C and it's as good as silent to me.

Well done on getting more efficiency with these chip's.  It's been fun to watch (I think you had first opened S7LN on site).   A efficiency upgrade to .2j upgrade is pretty amazing.  Either way congratulations it's a heck of a upgrade you have done on it.
legendary
Activity: 3374
Merit: 1859
Curmudgeonly hardware guy
It looks like my 640/610 profile is just fine for 400MHz, delivering 1.8TH at 340W for an efficiency of 0.189J/GH

I'm testing 425MHz on it now, and it's looking good. If it runs stable all night (I'm at zero errors five minutes into cold voltage), I can say 1.9TH at, according to my meter, 366W for 0.191J/GH. So this S7LN will go below 0.2J/GH machine-level. I was having trouble with the 630/600 profile doing anything right, but I'll check and make sure I just didn't screw up the code.

By the way, those numbers are for fans fixed at 20% (1560RPM); peak temp is 47C and it's as good as silent to me.
legendary
Activity: 3374
Merit: 1859
Curmudgeonly hardware guy
Well, looks like we got 5 for a group buy. I'll take that over to the right section then.
legendary
Activity: 4256
Merit: 8551
'The right to privacy matters'
I don't actually own a Kill-A-Watt and none of the stores I looked at in town apparently carry it, so I just threw together a metered socket with 0.1A resolution. The miner is currently powered off a Dell Z750P PSU. At hot voltage (650mV) I market 120V 2.1A, and when it switched to cold voltage (620) I mark a solid 120V 1.9A for 228W - that's the controller, both fans and a single board clocked at 1012GH for a device-level efficiency of 0.225J/GH

With both boards and the stock PSU, the current is bouncing between 3.4 and 3.5 amps. If we call it 3.45A, that's 414W for 2.03TH at 0.204J/GH - which definitely makes me happy.

I'll do more measures tonight (it's friggin' hot at the shop so I'll run it at home where I can also get a good "how annoying is this" sound gauge) and put up some more numbers on different setpoints.

EDIT - except that I moved some hex files over to a thumbdrive and then left it plugged into the workbench machine and didn't realize until I was already at home. So I won't be testing other voltage setpoints tonight. Whoops.

Also I have interest to the tune of 3 machines so far for a group buy.

while doing this, did you still had to power all 7 PCIe connectors or 4+1=5 was enough?
As far as wattage goes, it should be OK with 5.
Some people might be interested in selling the PSU on the aftermarket and powering two or even three with one EVGA 1300.
Using splitters one can up the number of connectors on 1300w EVGA to at least 12 easily.

I am looking to run 1. I am looking to use it with this psu.

https://www.amazon.com/gp/product/B01CE7NUIU/ref=oh_aui_detailpage_o01_s00?ie=UTF8&psc=1

I think it would do 395 watts.

I also may try to trick the fans the fans pull a lot of watts this fan below does less watts.


https://www.amazon.com/Focused-NF-F12-iPPC-3000-PWM/dp/B00KFCRATC/ref=sr_1_1?s=pc&ie=UTF8&qid=1465781203&sr=1-1&keywords=noctua+3000

goal would be to get under 390 watts
legendary
Activity: 3374
Merit: 1859
Curmudgeonly hardware guy
When I was bench-testing I only had 2 cables plugged into the board.

Here at the house, I locked the fan to 30% and it's about the same noise as a window fan on low sitting at the same distance. Pretty unobtrusive for my house. Temps at 44 and 46 in a 75F ambient.
legendary
Activity: 3892
Merit: 4331
I don't actually own a Kill-A-Watt and none of the stores I looked at in town apparently carry it, so I just threw together a metered socket with 0.1A resolution. The miner is currently powered off a Dell Z750P PSU. At hot voltage (650mV) I market 120V 2.1A, and when it switched to cold voltage (620) I mark a solid 120V 1.9A for 228W - that's the controller, both fans and a single board clocked at 1012GH for a device-level efficiency of 0.225J/GH

With both boards and the stock PSU, the current is bouncing between 3.4 and 3.5 amps. If we call it 3.45A, that's 414W for 2.03TH at 0.204J/GH - which definitely makes me happy.

I'll do more measures tonight (it's friggin' hot at the shop so I'll run it at home where I can also get a good "how annoying is this" sound gauge) and put up some more numbers on different setpoints.

EDIT - except that I moved some hex files over to a thumbdrive and then left it plugged into the workbench machine and didn't realize until I was already at home. So I won't be testing other voltage setpoints tonight. Whoops.

Also I have interest to the tune of 3 machines so far for a group buy.

while doing this, did you still had to power all 7 PCIe connectors or 4+1=5 was enough?
As far as wattage goes, it should be OK with 5.
Some people might be interested in selling the PSU on the aftermarket and powering two or even three with one EVGA 1300.
Using splitters one can up the number of connectors on 1300w EVGA to at least 12 easily.
legendary
Activity: 3374
Merit: 1859
Curmudgeonly hardware guy
I don't actually own a Kill-A-Watt and none of the stores I looked at in town apparently carry it, so I just threw together a metered socket with 0.1A resolution. The miner is currently powered off a Dell Z750P PSU. At hot voltage (650mV) I market 120V 2.1A, and when it switched to cold voltage (620) I mark a solid 120V 1.9A for 228W - that's the controller, both fans and a single board clocked at 1012GH for a device-level efficiency of 0.225J/GH

With both boards and the stock PSU, the current is bouncing between 3.4 and 3.5 amps. If we call it 3.45A, that's 414W for 2.03TH at 0.204J/GH - which definitely makes me happy.

I'll do more measures tonight (it's friggin' hot at the shop so I'll run it at home where I can also get a good "how annoying is this" sound gauge) and put up some more numbers on different setpoints.

EDIT - except that I moved some hex files over to a thumbdrive and then left it plugged into the workbench machine and didn't realize until I was already at home. So I won't be testing other voltage setpoints tonight. Whoops.

Also I have interest to the tune of 3 machines so far for a group buy.
legendary
Activity: 4256
Merit: 8551
'The right to privacy matters'
I would guess any regulated-string S7 variant would be improvable.

I would also guess the S9 would be improvable. 63 chips is seriously dense though, so they might already be running them close to bottom end. I have an S9 being delivered for hosting this week that I could spend an hour with if the owner allowed it, but I don't like experimenting with other folks' hardware.

I'll try and work out some efficiency curves for the S7LN this afternoon, might bring it home and run it here at the house to get an idea of noise. Should probably bring a stock S3 as well for comparisons, since me saying "it's not that loud" doesn't mean anything to almost literally everyone else (I have a very high tolerance for fan noise).

So, something like a hacked S7LN group buy might be in order?

Sending pm.
legendary
Activity: 3374
Merit: 1859
Curmudgeonly hardware guy
I would guess any regulated-string S7 variant would be improvable.

I would also guess the S9 would be improvable. 63 chips is seriously dense though, so they might already be running them close to bottom end. I have an S9 being delivered for hosting this week that I could spend an hour with if the owner allowed it, but I don't like experimenting with other folks' hardware.

I'll try and work out some efficiency curves for the S7LN this afternoon, might bring it home and run it here at the house to get an idea of noise. Should probably bring a stock S3 as well for comparisons, since me saying "it's not that loud" doesn't mean anything to almost literally everyone else (I have a very high tolerance for fan noise).

So, something like a hacked S7LN group buy might be in order?
legendary
Activity: 4256
Merit: 8551
'The right to privacy matters'
Well the one blade's been running solid at 620mV (9.3V, down from 10.0V stock) and 450MHz (1TH) for about 15 hours now. This afternoon I'll do some more testing of the absolute bottom end efficiency and overall power use, generate some charts and plots for y'all. I'll go ahead and work up a stock S7 as well.

Just wondering, would anyone be interested in a group buy of S7 or S7LN pre-flashed to the desired improved efficiency setpoint for a moderate markup (maybe $20-30), tested and reshipped within the US within one or two days of arrival at my shop?

maybe an  an s-7ln for me works,

 but do you think the s-9 can do this?  and drop to 9.3v if so  it may end up being an amazing value.

I have 4 s-9's two coming on tues-weds  two in 12 days or so. I think I will run 3 s-9's in the array and one with my buddy in his office.

Then run your s-7ln in house for fun along with the eth coin gear.

Maybe some research on s-9's will allow for them to go down to .090 or .085 per gh
legendary
Activity: 3892
Merit: 4331
Actually, 4TH would be more like 1100W.

there is also a 'funny' S7 batch 6, which was a prototype for later batches.
This batch is 4.05th at 1042W +extra at the wall per specs.
Do you think that this one can be improved or is it efficient already?
it's rated voltage is 11.6-13v.
thanks
full member
Activity: 214
Merit: 100
1KippERXwH1PdBxKNt1ksgqh89WBv6CtWQ
Well the one blade's been running solid at 620mV (9.3V, down from 10.0V stock) and 450MHz (1TH) for about 15 hours now. This afternoon I'll do some more testing of the absolute bottom end efficiency and overall power use, generate some charts and plots for y'all. I'll go ahead and work up a stock S7 as well.

Just wondering, would anyone be interested in a group buy of S7 or S7LN pre-flashed to the desired improved efficiency setpoint for a moderate markup (maybe $20-30), tested and reshipped within the US within one or two days of arrival at my shop?

My S7 already sitting in your shop.  Tongue  Feel free to take a look at it if you like.  Smiley
legendary
Activity: 3374
Merit: 1859
Curmudgeonly hardware guy
Well the one blade's been running solid at 620mV (9.3V, down from 10.0V stock) and 450MHz (1TH) for about 15 hours now. This afternoon I'll do some more testing of the absolute bottom end efficiency and overall power use, generate some charts and plots for y'all. I'll go ahead and work up a stock S7 as well.

Just wondering, would anyone be interested in a group buy of S7 or S7LN pre-flashed to the desired improved efficiency setpoint for a moderate markup (maybe $20-30), tested and reshipped within the US within one or two days of arrival at my shop?
Pages:
Jump to: