Supermicro H8QGi/6 and H8QGL Next Generation OC BIOS

[tear]

So, not knowing the MB architecture very well (actually not well at all), and no block diagram, does each CPU control one of the dedicated I2C busses?

Yes.

On the 4P, is there one IR3521 per CPU or does one control multiple xPHASE3 chips which in turn provide core voltage for all CPUs?

One per CPU. If you visually inspect the board you'll find them.

I will spend some time looking for a simple micro I'm comfortable with that can hang on their HS bus. I'll need to consider whether to use an on board clock or if it's OK to use the built in micro clock.

That's I2C so clock always origins in the master and is only driven when the master
is transmitting (both lines are high when the bus is idling) -- unless AMD's tweaked the
protocol so clock is always on (unlikely).

Also, one doesn't really need to perform HS transmission (3.4 MHz); only one CPU rev
supported that (C3) and our CPUs of interest (D1) only support 400 kHz.
May even be able to use on-uC I2C...

And should I/we start a seperate thread for this since I've managed to hijack this one badly?

I'm fine with it so far but if someone else feels we've crossed the line then we could
probably ask a mod to extract this discussion into separate thread.

 

[tear]

Hi.

With this new OC bios, is the CPU voltage the only limitation for high refclock or not ?

Pretty much, yes. We're still struggling with minor HT issues on some setups
but I hope this can be resolved soon.

 

[Jeanjean]

Pretty much, yes. We're still struggling with minor HT issues on some setups
but I hope this can be resolved soon.

Hi Tear.

Very good news and incredible job !

Can't wait to test the new OC bios for my H8QGL when it will be available .

Indeed, I have got some headroom with my CPU voltage.

 

[Punchy]

Also, consider that as TPF decreases, PPD increases non-linearly.

As do watts. Has anyone been measuring power as they increase the clock to see if PPD/watt is increasing, decreasing or holding steady?

 

[Kendrak]

R-Type said it better....

 

[R-Type]

As do watts. Has anyone been measuring power as they increase the clock to see if PPD/watt is increasing, decreasing or holding steady?

This is incorrect. Power consumption increases non-linearly from voltage increases, of which there are none here. Straight increases in clockspeed at the same voltage should be linear and most importantly, rather small.

 

[Core32]

I'm ordering a kill-a-watt to make some measurements on the OC 6128 rig.
I'm very surprised at how little temps have gone up, at 18.5% OC now, and the fan scream is still much lower than they create during the POST. I have them on "balanced" control.
The fans could be running faster than stock but my audio volume detection has decreased dramatically over the years.

 

[Punchy]

This is incorrect. Power consumption increases non-linearly from voltage increases, of which there are none here. Straight increases in clockspeed at the same voltage should be linear and most importantly, rather small.

Okay, I was assuming that voltage increases were included along with frequency increases in order to maintain screen colors other than blue. If you can keep a system stable with any significant frequency increase and no voltage increase, great! I guess it depends how much margin AMD builds in to the voltage settings as they bin the processors and program the voltages.

 

[Parja]

This is incorrect. Power consumption increases non-linearly from voltage increases, of which there are none here. Straight increases in clockspeed at the same voltage should be linear and most importantly, rather small.

More specifically, power increases as the square of the voltage increase. But yes, all else being equal, CPU power consumption increases linearly with clock speed.

So, anybody finding a particular motherboard to overclock better than others? The thought of being able to buy the cheapest dodecas available and put a healthy OC on them is mighty tempting.

 

[Nathan_P]

More specifically, power increases as the square of the voltage increase. But yes, all else being equal, CPU power consumption increases linearly with clock speed.

So, anybody finding a particular motherboard to overclock better than others? The thought of being able to buy the cheapest dodecas available and put a healthy OC on them is mighty tempting.

The 4p mafia have only managed to overclock one type of board so far, you can't go far wrong with the one mentioed in the thread title

 

[Parja]

The 4p mafia have only managed to overclock one type of board so far, you can't go far wrong with the one mentioed in the thread title

Well, there's at least 4 motherboards in the H8QGi/6 family, so mostly just wondering if any of those is particularly better than the others.

 

[Core32]

Well, there's at least 4 motherboards in the H8QGi/6 family, so mostly just wondering if any of those is particularly better than the others.

I believe those are just subsets of the same board.
Some have SATA only.
Some have SAS as well.
Some have onboard video or not.
Some have one PCIE slot and some with multiples.
The board I received has SATA only, onboard video and multiple card slots.

 

[tear]

+1

The only possible variable is presence of SR5670 (extra PCI-E slots,
extra HT link in use).

As of now, however, we have not seen any issues related specifically to SR5670.

 

[Vaulter98c]

Wow this thread just got wicked smart, way above my pay grade lol

Core32, did you ever get PPD numbers for your 4x 6128's at 2.3 Ghz? (230)

Any clue what they were getting at stock 2 Ghz?

Just curious to see what a bargain 4p build would pull....

 

[Parja]

Wow this thread just got wicked smart, way above my pay grade lol

+1

I think Tracy Morgan as Pumpkin Escobar said it best.

 

[Vaulter98c]

Yea, I feel like that black guy from the ipad video on the homepage, its got this and this and this, and I have no idea what any of that shit means but I aint gonna be the jackass walking around with an ipad2, thats for damn sure

lol

 

[Core32]

Core32, did you ever get PPD numbers for your 4x 6128's at 2.3 Ghz? (230)
Any clue what they were getting at stock 2 Ghz?

I bumped up the OC mid WU unfortunately, but the combined 225-230 PPD of the last 6904 was 275K+ for a total of 550k+ points. The final TPF was 27:11.
At stock, the PPD was on average 255K.
I am now running a full 6904 at 235 (2.35GHz) end to end so I will have a more accurate number to report.

 

[Vaulter98c]

Sweet, thanks man. Looks like a budget build could come in below 1500 dollars and perform just under the $5000 SR-X lol

 

[Core32]

Price is about right. I bought one barebones server for $950 and another coming in at $850, added 32GB ram for $240 and 4 x 6128s at $100 each.
The servers come with dual/redundant 1400W Gold PSUs, all cables, 4 x heatsinks w/thermal compound, 6 x chassis fans plus fans in each PSU unit, CD/DVD drive, 2 GigE ports, rack mount slides if you want them and 6 hot swap trays.
They also have built in video (VGA connector) so I did not need to add any cards.
I had a used 160GB SATA HDD, keyboard, mouse and 17" LCD.
The built-in fans are loud and can be VERY LOUD when the heat cranks up.
Fortunately I have a section of my basement that is far away from the finished area and stays about 65F year round.

 

[Core32]

One per CPU. If you visually inspect the board you'll find them.

I'll dive in and look for them when the second board arrives next week. If they are close enough together a single micro with multiple I2C sections might be enough.
The difficult part (probably over simplifying everything else by a lot ) will be making this useable/installable by all who would like to impement it.
One layer of the onion at a time though

That's I2C so clock always origins in the master and is only driven when the master is transmitting (both lines are high when the bus is idling) -- unless AMD's tweaked the protocol so clock is always on (unlikely).

I actually meant the clock to run the micro, but point taken. Almost not worth running a micro on external clock or crystal any longer since most can be had with decent internal.

Also, one doesn't really need to perform HS transmission (3.4 MHz); only one CPU rev supported that (C3) and our CPUs of interest (D1) only support 400 kHz.
May even be able to use on-uC I2C...

For "on-uC I2C" you mean non-bit banged? That's what I was thinking, to use the controllers built in I2C function. Maybe I misunderstood that point.

 

[tear]

I'll dive in and look for them when the second board arrives next week. If they are close enough together a single micro with multiple I2C sections might be enough.

Unfortunately, they're not packed together...

The difficult part (probably over simplifying everything else by a lot ) will be making this useable/installable by all who would like to impement it.
One layer of the onion at a time though

Agreed. Rome wasn't built in a day

For "on-uC I2C" you mean non-bit banged? That's what I was thinking, to use the controllers built in I2C function. Maybe I misunderstood that point.

Yes, I meant non bit-banged. We're in sync.

 

[Core32]

Ok. I think I have all the up front information I need to study while I wait on the "patient" to arrive.
One other item I've also wondered about is emergency OVP for this. I'm sure something must be in place already on the MB but we may need to adjust it closer to the limit or sharpen the roll-off so to speak. Any idea if this should be a concern?

 

[402blownstroker]

After the BIOS has been flashed and the 'Optinal Defaults' have been loaded, do we still need or should apply the settings from musky's G34 checklist:

BIOS settings to check
- PowerNow set to Disabled (see above for the check)
- NUMA/SRAT set to Enabled
- Node Interleaving set to Disabled
- Bank/channel interleaving set to Auto or Enabled

 

[tear]

Nope. These are taken care of the moment optimal settings are selected and saved.

 

[tear]

Ok. I think I have all the up front information I need to study while I wait on the "patient" to arrive.
One other item I've also wondered about is emergency OVP for this. I'm sure something must be in place already on the MB but we may need to adjust it closer to the limit or sharpen the roll-off so to speak. Any idea if this should be a concern?

Protocol limitation is 1.55V. For OV caused by circuit malfunction see page 24 of IR3521's
spec. Re CPUs -- SOIs are pretty difficult to burn, I've _heard_.

 

[Core32]

Still waiting on the new board to experiment with the OV device. It's due to arrive Friday.

In the mean time I experimented with pushing the OC on my existing quad 6128.
I am able to fold 6904s with a few retries at 260, hottest core at 42C.

When shooting for 262 I accidentally typed 272 and hit enter before noticing.
I received an error stating something to the effect " 272 too high, OC must be between xxx and 262" or something similar. Did not write it down unfortunately.
I did not power off but next ran the OC command for 262, which it accepted, then powered down.
At that point the rig would not POST .
Scared me a little but resetting the CMOS worked fine. I went back to 258 and it has been folding 6904s for two days with a TPF of 24:14, about 340K PPD.

So now I am wondering if the non-POST was due to the 272 typo or was the 262 accepted and that's my breaking point?
And is 262 a hard coded limit set in the new BIOS like the message said?

 

[Vaulter98c]

Still, 340k on 400$ worth of CPU is great

 

[Core32]

Still, 340k on 400$ worth of CPU is great

Totally agree and amazed at the temps on air. Was initially thinking of H2O cooling but now I'm not sure, except for the fun of working on it.
Now I am trying to decide whether to get higher speed 8c (6136?) or low speed 12c (6166?) for the new rig if the money is equal.

 

[Wyluliraven]

12core, definitely.

 

[ChelseaOilman]

So now I am wondering if the non-POST was due to the 272 typo or was the 262 accepted and that's my breaking point?
And is 262 a hard coded limit set in the new BIOS like the message said?

I've run into non-post situations twice on my 4P machines. The 1st time I tried to get my 6180SE CPUs to run at 230 fsb but cooling is poor in that machine or those CPUs just won't clock that high. Took me a couple tries clearing cmos to get the machine running again. It's been running trouble free at 220 fsb ever since.

The second time I tried to run my 6174 CPUs at 235. That machine booted and I was able to fold a few frames before the F@H client crashed. Linux was still running ok so I ran the script to back down the fsb to 232. I shut down the machine then it wouldn't start. It took several tries clearing cmos before the machine would boot. Oddly the fans would spin up each attempt but I didn't see the green power led on the motherboard light up and the board didn't boot. Finally when the board booted I set the fsb to 230 and it's been running like that since.

 

[jebo_4jc]

I'm so jealous of you guys.

I'm about to go put my tyan board in the garbage.

 

[Core32]

12core, definitely.

Obviously that's what I want
The more "affordable" low speed 12c are stock at 1.7G to 1.9G.
So a very optimistic 25% OC would result in 48 cores at 2.125G to 2.375G.
The more "affordable" high speed 8c are stock at 2.3G to 2.4G.
So a very optimistic 25% OC would result in 32 cores at 2.875G to 3.0G.

I don't have the intimate knowledge to compare the two for PPD, though I can see that
48 times 2.125 is greater than 32 times 3.0 and that PPD increases on a somewhat greater slope than pure linear.
Maybe I answered the question for myself

 

[Core32]

I'm so jealous of you guys.
I'm about to go put my tyan board in the garbage.

Hey! Five weeks ago I thought my 560Ti was a killer folding machine
Then I got really lucky and found the knowledge and effort being put forth on this site.
Keep that board running because in a short amount of time, someone with some new HW or trick will have all our 4P rigs looking lame I'm sure

 

[Jeanjean]

Hi

Any news about new OC bios for H8QGL ?

 

[Wyluliraven]

Hi

Any news about new OC bios for H8QGL ?

Under development last I heard, but you may need to hold your breath.

 

[firedfly]

Any news about new OC bios for H8QGL ?

I don't believe much time has been spent on it yet. Tear was trying to research issues with the current Gi/6 OC bios before porting the changes over to GL. The good news is that Tear has a GL board in his hands to work with! (thanks musky!)

 

[jebo_4jc]

"Gadgetsea" on eBay appears to have three h8qgi-f-o in stock and ready to sell for $750 shipped.

Anybody want to buy a tyan s8812 for $425?

 

[R-Type]

"Gadgetsea" on eBay appears to have three h8qgi-f-o in stock and ready to sell for $750 shipped.

Anybody want to buy a tyan s8812 for $425?

I haven't has time to do much research on 4P's, what is the model with 3 EPS connectors? I'm assuming this is the most desirable?

 

[402blownstroker]

I haven't has time to do much research on 4P's, what is the model with 3 EPS connectors? I'm assuming this is the most desirable?

The H8QGi-F is the model with 3 8-pin EPS connectors.

 

[Activate: AMD]

great googly mooglie you guys are nuts! Making me really jealous