Hi soy,
If you really looking to improve efficiency then you have to down-clock the S3, the sweet spot is 150Mhz.
Once you drop the clock you should be able to lower the voltage even further down than 9v on your power supply.
I like the idea of simply moding the power-supply instead of moding each of the 16 step down voltage converters on the s3 board, which is a major p.i.t.a.
I would really like to see your results at 150Mhz and see if your power supply mod can match my S3 pencil mod efficency (180w @ 305 Gh/s on the wall)

Regards,
-Jay

Yes they look very similar, the man difference is that you have to take the S3 pretty much completely apart.
You have to get access to the back side of the blades since the resistors ar on the back not in the front like on the S1. So just popping of the case won't be enough.

Hi Phillip,
Correct, I run 6x S3 at 150Mhz undervolted at exactly 1090w on the wall (GHash.io reports a 1.85 TH/s dally total ). Not sure if 200 watts headroom is save ( I think your suppose to stay 25% under the maximum rating at continues usage ).

I tried undervolting 193MHz-200Mhz but it only yielded 380Gh/s at 280w was not worth it in my opinion. ( the gain was only 20w ) and noise and heat was not much lower.

If you shot me your email in a pm I canemail you some photos and simple instructions. If you done a S1 pencil mod you should be fine, but if not it might be a little more you bargaind for.

I used the new firmware with the "Advanced" tab to adjust the clock frequency down to 150mhz, with that the consumption goes down to 240-250w on the wall ( at 310 Gh/s ).
The resistors are located on the back, or the inside where the big sink is attached. You will need to completely remove the board to get proper access to the resistors.
The default value is 2.083 Kohm. And I use the pencil mod to bring the value down to 0.900 to 1.050 Kohm. I have most of mine set close to 0.950 Kohm.
After that the consumption will be down to 180w on the wall ( still at 310 Gh/s ). Noise from the fans and temp will be way down after this mod. And you can use a single PCIE cable to power one S3 ( two hashing boards )

0.68v seams to be the physical limit of the chips ( at 150Mhz clock ). A little lower and you get all hardware errors.

That chart seams to be purely theoretical, or may been results of the chip design simulation.

I did a whole batch of undervolting my S3 to different voltages trying to replicate the efficenty in the chart.
At 0.72v the max clock could not be higher than 193Mhz and resulted in 380GH/s at 280w on the wall. That's only 20w improvement over the unmodified voltage ( 300w at 380GH/s )
The gain by undervolting is not comparable to the S1 where you could get almost twice the efficency with a careful pencil mod.
Although for me it was still worth it, because of my power restrictions and heat problems in my small apartment.
Without intend trying to keep the max GH/s In mind I found my sweet spot at 150MHz clock rate and 0.68-0.69v producing 310GH/s at 180w on the wall.
Also I really wanted to get the consumption per S3 below 200w to be able to use a single PCIE power cable per unit so I can have all 6x S3's running of only one 1500w SilverStone PSU.
The end result is 1.8-1.9 TH/s at 1100w on the wall (80plus Gold PSU) which is still only 0.6w per GH.
So that may be one of the reasons why ther is no "official" S3 pencil mod. It's simply not worth the trouble.
In my case it was more about the power restriction and heat output. Which is now more like a warm breeze on the fan exhaust rather than standing in front of a space heater on max in my small Texas heat 625 sqft Apartment. As nice side effect the fan noise is a lot less now (from 2100-2400 rpm Down to 1000 rpm). With some F12 fan replacements I could make the S3 completely silent now.
The mod is a big hassle and very tricky with the having to remove the case and boards. The location of the resistors ( on the inside/back of the board ) is also tricky to work with.
And the PCB seams to be very fragile, I managed to kil two boards. I barly touched the board and yanked off a tiny capacitor this was sticking out from the board. After carefull soldering I tought I fixed it but must have shorted something and resulted in a dead board. Even only one bad voltage regulator ( 1 of 8 per board ) results on a completely dead board.

I made plenty of pictures during the mod, but I feel the gain in efficency and General PITA to do the mod is not worth the write for a tutorial.

-Jay

Block 266975 shows up double on Blockchain, one of them is showing as (Orphaned) ?
Lets hope at least the first one is Valid.

ok, it must be variance.
block 20642 and block 20656 round length was exactly the same ( 1hour  10min )
in the same round time I had 10.000 shares less in round 20656 and 20% less payout.
so I guess on top of the bad pool luck I now also have bad luck and a constant low variance 

I wish Slush would have a user hash rate graph ( like WeMineLTC ) that would show the variance over time.

hmm,
my block reward dropped from 0.025 to 0.0215 progressively in the last 10 hours, that's close to a 20% decrease !
please correct me, but I don't remember the pool hash-rate being much lower 10 hours ago, or was there really a almost 20% increase in hash power in recent hours ?
my 3 AVALON's hashing  constant at 240 - 245 G/hash per sec.

anybody else ?

Why cant the luck stay around 100% for a while. Why from 192% straight down to 40%  
Almost like the Texas weather, from one extreme to the other in a day.

  new diff is 390 million !

correct,
but it comes pretty darn close, specially for shorter rounds ( if mined steady for at least 1 million shares )

Block 20608 fixed itself,
Slush must have a script running a couple of hours after a block was found that recounts the shares.
 
if you want to calculate your correct payout use the following:

25 divided by the total round shares
then multiply with your number of submitted shares

(25/total shares) * your shares

Looks like everything is working for me again ( startum.bitcoin.cz ).
only had a quarter of my regular income this morning ( last 5 blocks )
but the last two blocks should make up for that with almost double the payout for me ( pool fell way below 200 Th ) and on top we solved the last block in 41 seconds 

Thanks, received my refund. Relieved 

Any idea why I am not in this new payout sheet ?
I filled out the from that was posted and verified my return address with bigbeninlondon.

Unconfirmed addresses outstanding for the following people:

I confirmed my refund address with you via PM, so I shouldn't be in the unconfirmed list ?

Info sent...

20 @ .017BTC

Sorry, its not instant.
Right now it is hiding the "Unconfirmed reward", so they are missing from the "Total reward".
Once a block reaches the 100 confirmations then the correct reward is added to the "Confirmed reward". So no need to panic, its only a display issue.