"The Stilt" sets new AMD LN2 OC records

[BeepBeep2]

Major props to The Stilt for these mind-blowing scores. He deserves lots of praise!

SuperPi 32M: 9 min 34.078 sec
http://hwbot.org/submission/2426495_...min_34sec_78ms

SuperPi 1M: 9.218 sec
http://hwbot.org/submission/2426494_...370_9sec_218ms

Also congrats to I.nfraR.ed for his scores, very impressive as well.

[Evantaur]

waiiitt....he took 1st place with AMD hardware? :O

[BeepBeep2]

No, but fastest scores ever set by AMD hardware.

[Satanello]

:-O excellent result!!

Inviato dal mio GT-I9300 usando Tapatalk 4

[FlanK3r]

Its much better than the ebst results before...Than Phenoms II days. And Phenom II has x87 coprocessor, FX not. Wow...Hope, Stilt cvome here and tell more us

[The Stilt]

The 9370 is indeed a tough chip.

The second CU is simply superior to anything else ever made on 32nm node.
The remaining three are also way above the average.

The leakage is very high on FX-8 series standards (over 25% higher than the 8320 I've got), yet ultra low on FX-9 series standards.
The 8320 is much better on conventional cooling methods as it runs so much cooler than the 9370.
The 9370 needs a high end custom loop to outperform my 8320.
In those conditions the 9370 gains around 150MHz over the 8320.

On extreme cooling the 9370 really scales alot better.
That is not directly because of the higher leakage, but the so called voltage wall (the critical voltage, nearly the same for all of the chips, no matter the leakage characteristics).

It took some serious binning to find it thou..
It's the first and the only 9370 I've tested

Any guesses what is the maximum validation with all 8 cores enabled?

[i found nemo]

was this with or without the patch that u made stilt? also was 1.8 real volts for 7.6ghz? wow that's pretty decent 4.2ghz per volt AND benchable :p

[The Stilt]

was this with or without the patch that u made stilt? also was 1.8 real volts for 7.6ghz? wow that's pretty decent 4.2ghz per volt AND benchable :p

The whole idea of the challenge was to use the new patch, so indeed it was.
The real voltage was 1.88V during 32M calculation and 1.93V during 1M calculation.

[demonkevy666]

that Batch So close to 1337 lol

1322 PGS [Pretty gruesome settings]

chip begs to be awesome LOL

[BeepBeep2]

The 9370 is indeed a tough chip.

The second CU is simply superior to anything else ever made on 32nm node.
The remaining three are also way above the average.

The leakage is very high on FX-8 series standards (over 25% higher than the 8320 I've got), yet ultra low on FX-9 series standards.
The 8320 is much better on conventional cooling methods as it runs so much cooler than the 9370.
The 9370 needs a high end custom loop to outperform my 8320.
In those conditions the 9370 gains around 150MHz over the 8320.

On extreme cooling the 9370 really scales alot better.
That is not directly because of the higher leakage, but the so called voltage wall (the critical voltage, nearly the same for all of the chips, no matter the leakage characteristics).

It took some serious binning to find it thou..
It's the first and the only 9370 I've tested

Any guesses what is the maximum validation with all 8 cores enabled?

Is it often that the second CU is faster than the other three? In four CPUs I've tested now (only Zambezi) the second CU was the fastest.

Based on your results I'll guess 8.06 GHz for all 8 cores.

[I.nfraR.ed]

Well deserved, Stilt! Great knowledge about what is going on behind the scenes and a great dedication.
Also this was not possible before without the BDC, so even without your scores, you still would had been the big winner.

I'm only sorry for the 32M. Had to run CPU-NB at max +0.15V, otherwise it was freaking out.
That's why I settled for relatively low CPU-NB frequency, but lost a lot of time and LN2 while trying to figure out what is going on.
While I could run 6-9-6 on the FM2 board, it was impossible on this one, plus the board restored its backup bios several times. I don't understand why changing some mem timings leads to corrupted bios (or at least the board thinks so).
At the end, I managed to complete just the control 32M runs (without wazza) and the 7840 run crashed in the middle. Unfortunately I had no more LN2 left, so I had to submit the 7770 run.

1M score is bad. Maybe it's due to voltage being higher than the tolerance, but I was getting slower scores at lower frequencies and voltage, probably because already corrupted OS. The slow mode switch was not working this time (I guess the wire got accidentally disconnected). Will have to take a second spin sometime in the future. At least I managed to get in the 8GHz club, completely unexpected .

Is it often that the second CU is faster than the other three? In four CPUs I've tested now (only Zambezi) the second CU was the fastest.

Based on your results I'll guess 8.06 GHz for all 8 cores.

Second module was best on mine, too, while my oldest chip has almost even CUs, but the first is slightly better than the rest and 4th being the worst.

[Lightman]

Excellent news and great work by The Stilt!
Congratz!

I wonder if going LHe would help a little bit more. Benching anything around or above 8GHz is fantastic!
Remember the days of P4 where most of the people thought 10GHz era is around the corner? Thanks to AMD we are a bit closer but still not quite there

[FlanK3r]

I think validation with all cores 8.1-8.2 GHz (cause my died FX-6300 has with all cores 7880 MHz and still there was limit in voltage. I used only 1.95V)

Do you have some results with air/watter etc? Show something . Im looking forward. My FX-9370 coming in 2 days. There was nice sale and for dew days was price for FX-9370 near FX-8350. So I bought it . Yesterday was price 50% higher, so I was lucky for this moment .

[The Stilt]

Is it often that the second CU is faster than the other three? In four CPUs I've tested now (only Zambezi) the second CU was the fastest.

Estimates based on my logs on all parts I've tested:

In 80% of cases = CU1-CU0-CU3-CU2 (best to worst).
In 15% of cases = CU0-CU1-CU3-CU2 (best to worst)
In 5% of cases = CU2-CU1-CU0-CU3 (best to worst)

Usually the CU which does the highest benchable frequency is not the one which hits the highest validation frequency
This 9370 represents the first pattern for benchmarks, yet the third one applies for validation.

Zambezi is a bit different, however the worst CU is usually the CU2 there too.

Well deserved, Stilt! Great knowledge about what is going on behind the scenes and a great dedication.
Also this was not possible before without the BDC, so even without your scores, you still would had been the big winner.

I'm only sorry for the 32M. Had to run CPU-NB at max +0.15V, otherwise it was freaking out.
That's why I settled for relatively low CPU-NB frequency, but lost a lot of time and LN2 while trying to figure out what is going on.
While I could run 6-9-6 on the FM2 board, it was impossible on this one, plus the board restored its backup bios several times. I don't understand why changing some mem timings leads to corrupted bios (or at least the board thinks so).
At the end, I managed to complete just the control 32M runs (without wazza) and the 7840 run crashed in the middle. Unfortunately I had no more LN2 left, so I had to submit the 7770 run.

1M score is bad. Maybe it's due to voltage being higher than the tolerance, but I was getting slower scores at lower frequencies and voltage, probably because already corrupted OS. The slow mode switch was not working this time (I guess the wire got accidentally disconnected). Will have to take a second spin sometime in the future. At least I managed to get in the 8GHz club, completely unexpected .

Second module was best on mine, too, while my oldest chip has almost even CUs, but the first is slightly better than the rest and 4th being the worst.

Usually when your time suddenly slows down in SuperPi at very high frequencies (on Zambezi & Piledriver based parts) you are probably hitting a ECC flood for L1 or L2 caches. It might add a 0.5 seconds per a calculation phase in 1M Raising the voltage does not necessarily help, but the temperatures do. When pushed to the absolute limit, it is always a bit faster immediately after a proper cool down. Thats why I use the "Slow Mode".

Your efficiency in SuperPI is simply sick, there is no way around it.
I have absolutely no clue how you do it. There must be something in your OS setup, but I have no clue what it is.
Before I started making the competition runs I tested the difference of different timing and frequency combinations.
No matter what the sub-timings were, the fastest combination was usually still the highest frequency (in SuperPI).
The tighter timings came close but still could never pass the results achieved on much higher MEMCLK.

I tested on both, PSC and Hyper.

Do you use a XP installation with embedded SP3 and updates or do you install them after the OS installation?
If you do install them in the first place?

Excellent news and great work by The Stilt!
Congratz!

I wonder if going LHe would help a little bit more. Benching anything around or above 8GHz is fantastic!
Remember the days of P4 where most of the people thought 10GHz era is around the corner? Thanks to AMD we are a bit closer but still not quite there

On LHe the 9370 (e.g.) is highly likely to puncture the 9GHz barrier.
The benchable frequency for light loads such as SuperPI should be around 8.5GHz:ish.

On LHe the actual die temperature would drop around 60 degree (C) compared to LN2.
That should be more than enough to make up the remaining difference...

The only issue in LHe is the cost.
In Finland the required amount (around 300 litres) would cost around 8000 euros.
It could be hauled over from Sweden where it costs significantly less, but the cost would still be around 4500 euros.

I think validation with all cores 8.1-8.2 GHz (cause my died FX-6300 has with all cores 7880 MHz and still there was limit in voltage. I used only 1.95V)

Do you have some results with air/watter etc? Show something . Im looking forward. My FX-9370 coming in 2 days. There was nice sale and for dew days was price for FX-9370 near FX-8350. So I bought it . Yesterday was price 50% higher, so I was lucky for this moment .

Not nearly enough

I'll do some benches soon.

[I.nfraR.ed]

Estimates based on my logs on all parts I've tested:

In 80% of cases = CU1-CU0-CU3-CU2 (best to worst).
In 15% of cases = CU0-CU1-CU3-CU2 (best to worst)
In 5% of cases = CU2-CU1-CU0-CU3 (best to worst)

Usually the CU which does the highest benchable frequency is not the one which hits the highest validation frequency
This 9370 represents the first pattern for benchmarks, yet the third one applies for validation.

Zambezi is a bit different, however the worst CU is usually the CU2 there too.

I can confirm that usually the CU that goes highest for validation is unable to match some of the weaker ones for benchmarks,
however in my case with the FX-8350 from the competition the best CU for both purposes was CU1.
It was the other way for the 6800K - CU0 was slightly better for validation (I saw 7.8), while CU1 was better for benchmarks.
There's another thing - not all CUs are equally fast. Some are a little bit slower.

Usually when your time suddenly slows down in SuperPi at very high frequencies (on Zambezi & Piledriver based parts) you are probably hitting a ECC flood for L1 or L2 caches. It might add a 0.5 seconds per a calculation phase in 1M Raising the voltage does not necessarily help, but the temperatures do. When pushed to the absolute limit, it is always a bit faster immediately after a proper cool down. Thats why I use the "Slow Mode".

Yes, I've seen that. Perhaps my problem with 1M was exactly the slow mode. I was using PSCheck for lowering the PState and was keeping it opened, so I can immediately swich the PState when the run finishes.
Probably that crippled my performance. I will check the wire and resolder it, so it wil lwork next time.
Btw I gained more than 100MHz from last time, partially because of the low CPU-NB which was holding the CU back, partially from a perfect contact with the pot this time.
The first 7900MHz run that I submitted was really the highest I could go in that session. In the last session I could probably run 32M at 7.85 or even 7.9 if it does not throttle with about 2V.
7.82 is 100% doable.

Your efficiency in SuperPI is simply sick, there is no way around it.
I have absolutely no clue how you do it. There must be something in your OS setup, but I have no clue what it is.
Before I started making the competition runs I tested the difference of different timing and frequency combinations.
No matter what the sub-timings were, the fastest combination was usually still the highest frequency (in SuperPI).
The tighter timings came close but still could never pass the results achieved on much higher MEMCLK.

I tested on both, PSC and Hyper.

Do you use a XP installation with embedded SP3 and updates or do you install them after the OS installation?
If you do install them in the first place?

I tested Hyper and PSC, but my board (or IMC) can't do high memclk on any timings. 1250 or close to that is the maximum with PSC. With CFR I could boot at 2600 slack timings.
I couldn't even lower the timings to 6-9-6, while it perfectly fine on FM2 board with just 1.9V. 1300 6-9-6 should not be a problem.
In my testing Hypers were faster on my system and I was actually running 1060 6-6-6-18 tCWL 6, but CPU-NB was not stable at any settings, so I had to tear down. Then decided to give PSC a try.
Same problems with PSC, but finally managed to stabilize it, when LN2 was almost depleted.

About the OS - I use stripped OEM XP SP2. I don't install SP3 or any other component/update, just the .net 2 framework if I need it and vga drivers.
Then I do the usual optimizations such as maxmem, LSC, pagefile, etc. Then a manual wazza/cdt, although my AM3 scores in the comp are without - never managed to reach the point where I had to do a proper run.
As you know you need to have a good balance between available and cache after wazza. Initial available RAM is not important at all.
Also affinity and priority.
I also use an old Intel V25-X 40GB SSD, which is nothing spectacular in terms of IOPS or read/write speed.
Write me a PM if you're interested, I can give you my XP to try, however I don't think there's anything special,
I can do the same/similar with any other XP.

[Johan45]

I think it's absolutely awsome how you guy's can make these things dance !!
I've also gained some insight to the Superpi bench and should spend some time finding my best core(s)

[massman]

Yes, I've seen that. Perhaps my problem with 1M was exactly the slow mode. I was using PSCheck for lowering the PState and was keeping it opened, so I can immediately swich the PState when the run finishes.

I can confirm this. Having PSCheck open during SuperPI runs will hurt the performance like that. Always close it.

[chispy]

Well done the stilt , absolutely marvelous run , congratz.

[FlanK3r]

So, whats new the master (Stilt)?

[The Stilt]

Nothing much at the moment.

Testing the different A88X boards and writing some software for... ahem... the "friend"

[FlanK3r]

btw, did you tested before FX with different clocks at CU? Example I have very bad CU3 in my FX-9370. So my idea is 5100 MHz at this CU3 and after with PS check tried 5300 MHz for others CUs (not enable in settings all cores node)