A little birdie told me...

[i found nemo]

i don't know how this ended up being taken so personal. i have tons of respect for beep and chew, but for chew to think, beep, the one he taught.. would call him out on a fake result in a thread..i don't really know what to think of it, take your results seriously, but don't take questions/suggestions personally.

[Movieman]

I was there that day and all cores were at not only the amount submitted for the WR but for 8522 or 8528MHz. This I saw with my own eyes and grabbed a pic of the monitor with my camera,
If needed I will dig it up and post it here. End of argument.

[Don_Dan]

Please post it, this thread needs less arguing and more pics!

[Movieman]

Please post it, this thread needs less arguing and more pics!

OK, let me see if I can find it.. I know I kept it even though AMD didn't want it seen because that number was gotten after most of the press left.
I stayed an extra day and they only wanted to submit what had been seen by the 25 press guys that were there at the time.

[BeepBeep2]

i don't know how this ended up being taken so personal. i have tons of respect for beep and chew, but for chew to think, beep, the one he taught.. would call him out on a fake result in a thread..i don't really know what to think of it, take your results seriously, but don't take questions/suggestions personally.

What I said wasn't even about questioning a result, it was about the way a stupid article was written.

Like I said, the article was centered around a CPU-Z valid and basically there are 3 steps to CPU-Z...
1. Find best cores
2. Clock down other cores (send those into lower p-state), then push best core(s)
3. Validate highest frequency on said core(s).

The article, which was centered around a CPU-Z validation, was written as if all 8 cores were being pushed to or near the validation frequency. That usually doesn't make sense, so I asked for "proof". I meant that lightheartedly, as long as you understand the context around which I was asking. I know that most people that would be gunning for a frequency record, would be pushing hard on that one CU / 2 cores and not on the others. Typically also would be doing extra things to cool the CPU further like leaving it in a low p-state and low voltage until it was time to really push.

8 cores enabled makes it harder due to increased heat, I knew that already and I've done this before. And it doesn't take much more than a set of eyes to see that all 8 cores were enabled in the validation like chew* thought he was so gloriously pointing out to me.
But I wasn't questioning if 8 cores were enabled, or anything about the result, because you can't tell from the validation submission what cores were at what frequency.

IMHO, chew* just found it a great time to act like I was calling someone out and make me look like a retard...and while chew* taught me a lot, I also taught myself a lot, especially after he "retired". I haven't gone out and bought a ton of Vishera CPUs but I most definitely have enough experience on all of these platforms to know better than he is thinking.

I also have more respect for The Stilt than most anyone else here, and he knows himself how much respect I have for him so this whole thing is silly.

[The Stilt]

There seems to be plenty of misunderstanding around the 8.722GHz run on FX-8370.

Unlike some of the news falsely claim not all of the 4 compute units (8 cores) were operating at the frequency of 8.722GHz.
I have never claimed otherwise and honestly no one except Christian Ney even bothered to ask me.

The other compute units were clocked down to a lower PState, yet they were neither clock or power gated but in full power state (effectively C0 state).
Clock and / or power gating the other compute units or cores is quite common now days but thats not something I do.
Just as shutting down 3/4 of the compute units of the processor is just not my thing.
The way I currently do it mimics the normal behavior of Turbo Core. During low loads the idling cores will be clocked down.

ps. Some of you wondered how much power the chip draws at that voltage and frequency.
Based on the data measured previously the calculated power draw peaks at <140W immediately after the launch and then tapers down to 110W.

[Movieman]

There seems to be plenty of misunderstanding around the 8.722GHz run on FX-8370.

Unlike some of the news falsely claim not all of the 4 compute units (8 cores) were operating at the frequency of 8.722GHz.
I have never claimed otherwise and honestly no one except Christian Ney even bothered to ask me.

The other compute units were clocked down to a lower PState, yet they were neither clock or power gated but in full power state (effectively C0 state).
Clock and / or power gating the other compute units or cores is quite common now days but thats not something I do.
Just as shutting down 3/4 of the compute units of the processor is just not my thing.
The way I currently do it mimics the normal behavior of Turbo Core. During low loads the idling cores will be clocked down.

ps. Some of you wondered how much power the chip draws at that voltage and frequency.
Based on the data measured previously the calculated power draw peaks at <140W immediately after the launch and then tapers down to 110W.

I'm far from what I would call a "good" overclocker but I am a pretty good judge of people and have been following the exploits of yourself and others since 2005 when I came here.
You get a feel for people's character. Something I developed in over 45 years in the business world and when it comes from your mouth I take it as gospel truth and I don't feel that way about everyone.
There are people and some of them have been at XS over the years that in the pursuit of records forgot about the most important thing.
"It isn't if you win or lose but how you play the game."
Too many I have seen forgot that in their pursuit of records.

[FlanK3r]

I think, the OC is (for me most about have to fun together with Movieman sentence "It isn't if you win or lose but how you play the game." .
Im looking forward for all new FX results. So just relax guys. Nothing bad happens

[The Stilt]

So it is finally the second of September...

So far I have posted a picture of a crown (which once belonged king George XII of Georgia) and hashtags #kg & #Alphabet...

#kg which stands for Kilogram
#Alphabet... "Kilo" stands for "K" in the phonetic alphabet

So what is this all about?

Everything below is highly unofficial of course as is everything else I write here.

Piledriver module based Vishera die has been mass-produced in two different die revisions since the prototyping phase.
While all of the revisions have the same major die version (OR-C0), the minor revision has changed.
Initially the first mass-produced die revision was "India" (OR-C0i, prototype and ES only), the second revision was and still is "Juliett" (OR-C0j, retail) and now finally...
The "Kilo" revision (OR-C0k) a.k.a "King Vishera" a.k.a "Vishera Type-K" has arrived.

The "King Vishera" is initially only available in the new models, FX-8370E & FX-8370.
This is most likely the case with FX-8320E also, however I have not been able to test one of them personally.
The new version is likely to be phased-in at least in the other high-end models such as FX-9590 and at some point in all of the remaining models also.
The alledged metal tapeout of the new revision (alledgedly) occured in the beginning of July. So the only way to get a newer revision part is to get one of the new models, atleast in the beginning.

The differences?

- On average 18% less leakage*1 (0-38%) for FX-8370
- On average 53% less leakage*1 (14-106%) for FX-8370E
- Up to 300MHz higher overclocking margin *12
- 100mV less voltage required for the same clocks on average *1

*1 - Compared to an average FX-8320 or FX-8350 CPU
*2 - When not restricted by the cooling or the motherboard (VRM)

The E-version is the best choice for air or water cooling thanks to the ultra low leakage characteristics.
The non E-version does the same clocks however it might require use of a higher end motherboard (with better VRM) and high-performance cooling.
The non E-version has significantly better overclockability under sub-zero temperatures (phase, LN2) since the leakage levels of the E-version are too low for the purpose.
Having an ultra low leakage characteristics is great under normal conditions however under sub-zero temperatures the voltage requirements become a issue.
Basically the low leakage part exhaust the usable range of supply voltage prior reaching it's maximum frequency.

Based on my own tests, I would estimate that >95% of FX-8370 & FX-8370E parts will reach 4.8GHz frequency in 24/7 without a custom watercooling or a ultra high-end motherboard being a requirement.
As long as the temperature (see below) stays =<65 degree C or 149 degree F and the motherboard has even remotely a sufficient VRM you'll be fine.

On a high-end motherboard and a custom watercooling 5.0G - 5.2GHz+ should be doable in 24/7 use with a good specimen.

These chips still draws a vast amount of power when overclocked so the final overclocking potential is basically just the matter of cooling.
The maximum recommended temperature during the worst case stress is 65?C tCase.

Officially the maximum tCase temperature for the various FX models is specified to:

Infra A - FR (125W TDP) - 61.1 degree C
Infra B - WM (95W TDP) - 70.5 degree C
Infra C - OL (65W TDP) - 70.3 degree C
Infra D - HO (45W TDP) - 69.1 degree C
Infra E - SJ (25W TDP) - 70.0 degree C
Infra F - FH (220W TDP) - 57.0 degree C

The tCase temperature must not be mixed with the tCTL control value sometimes dubbed as the "package temperature".
The tCase temperature is also calculated and it represents the simulated case temperature, measured from the very center of the heatspreader (see the illustration).
Neither tCase or tCTL is the actual die temperature. The actual die temperature information is not directly available on these processors. The actual die temperature is significantly higher
than the tCase or the tCTL control value indicates.



The maximum tCTL control value on all of the FX-series processors is 70 units.
When that value is reached the processor HTC logic engages and starts to reduce the power consumption and dissipation by throttling.
Some of the motherboard manufacturers (such as ASUS) alter the limit manually to reduce the chance for throttling.

So when you are talking about the temperature always use the tCase temperature instead of the tCTL control value.

Here is the examples of the wrong and right values in various monitoring softwares.

Red = Wrong (tCTL)
Green = Right (tCase)









The AOD picture also explains the value dubbed as "Thermal Margin".

Also here is a reminder why Prime95 should not be used as a reference for stability when overclocked. It results significantly higher power draw and emitted thermal than any of the most stressful real world applications.If you still find it absolutely necessary to use Prime95 for stability testing please do it this way:

Run it on only two compute units at once (set the thread count to 4 and affinity accordingly) and decrease the cooling to simulate higher power dissipation.

Only the relative Stock or OC results are comparable.

[chinook]

That's some great info, The Stilt! And congrats on the world record
So, it seems that we'll still be getting some "new" toys to play with until the next iteration. Time to sell my procs and buy one of these new kings
I'm curious - how do you test stability? Do you use the Prime95 + 2 compute + lowered cooling thing or something else ?

[FlanK3r]

nice article man. Its as I thought, the Coretemp temps+15 C is around the real temp in chip?
Btw, can you post some OC result please

[The Stilt]

That's some great info, The Stilt! And congrats on the world record
So, it seems that we'll still be getting some "new" toys to play with until the next iteration. Time to sell my procs and buy one of these new kings
I'm curious - how do you test stability? Do you use the Prime95 + 2 compute + lowered cooling thing or something else ?

I usually set the voltage and clocks to the intended level an start encoding a MKV with X264 (which is the most stressful real world workload).
After around 15 minutes of encoding I check where the temperatures have peaked.
Once I have the temperature information available I start Prime95 Large FFT and set the thread count to four.
Before starting the torture test I set the affinity to the compute units I am going to test during the run (i.e. cores 0-3 = CU0 & CU1).
After starting the test I start reducing the fan rpm until the temperatures reach the same peak level they did during the X264 encode.

After you?re confident that the CPU is stable, repeat the procedure for CU2 & CU3 (cores 4-7).

It takes twice the time the normal method does, but it won't molest your motherboard or CPU beyond the potential breaking point.
Depending on the load line configuration I would recommend that you reduce the core voltage by 6.25 - 12.5mV during the test.
Thats just to compensate the difference in load line effect (droop) you will have from the reduced load from running two compute units instead of all four at once.

[The Stilt]

nice article man. Its as I thought, the Coretemp temps+15 C is around the real temp in chip?
Btw, can you post some OC result please

CoreTemp is tCTL only.
It does not support any of the external LPC/IO sensors (AFAIK) which are used to read the tCase (SB-TSI).

Some results inbound later today.

I am just waiting some water cooling components to arrive to dig in properly

[vario]

CoreTemp is tCTL only.
It does not support any of the external LPC/IO sensors (AFAIK) which are used to read the tCase (SB-TSI).

Some results inbound later today.

I am just waiting some water cooling components to arrive to dig in properly

First of all, thank you.So much good and concrete information.So it IS a new revision, albeit minor one.Thats a good thing.
On the prime95 thing, problem is, it IS a real world piece of software that does actually SOMETHING of use and is not a real power virus, you can join the community and help in Great Internet Mersenne Prime Search.So im not fully with you on that, although im more of a realist, so if your method will net me a 100% stable PC im all for it, until now only when i was prime 95 blend AND large stable none of my software hanged sooner or later, sometimes it could take 12h of encoding while doing another work but it would freeze/hang/quit the encoder.
Another thing is, there are 2 ways to set up an overclock at least on my mainboard, 1.APM ON i get turbo but fluctuating CPU clocks, when set at 4ghz they hover between 3.7 and 4 ,this results in lower power consumption also. or 2.APM OFF, i get no turbo rock stable 4ghz and higher power consumption.
So ,not everyones results are comparable :-/.Disabling APM is also a part of official AMD perf tuning guide here:
http://sites.amd.com/us/Documents/AM...ning_Guide.pdf

Anyhow, awaiting your results and a question, could you test NB clocking in this revision ? anything upwards 2400-2500 is really hard to get fully stable on mine.
Thank you again for loads of info.

[The Stilt]

First of all, thank you.So much good and concrete information.So it IS a new revision, albeit minor one.Thats a good thing.
On the prime95 thing, problem is, it IS a real world piece of software that does actually SOMETHING of use and is not a real power virus, you can join the community and help in Great Internet Mersenne Prime Search.So im not fully with you on that, although im more of a realist, so if your method will net me a 100% stable PC im all for it, until now only when i was prime 95 blend AND large stable none of my software hanged sooner or later, sometimes it could take 12h of encoding while doing another work but it would freeze/hang/quit the encoder.

Another thing is, there are 2 ways to set up an overclock at least on my mainboard, 1.APM ON i get turbo but fluctuating CPU clocks, when set at 4ghz they hover between 3.7 and 4 ,this results in lower power consumption also. or 2.APM OFF, i get no turbo rock stable 4ghz and higher power consumption.
So ,not everyones results are comparable :-/.Disabling APM is also a part of official AMD perf tuning guide here:
http://sites.amd.com/us/Documents/AM...ning_Guide.pdf

Anyhow, awaiting your results and a question, could you test NB clocking in this revision ? anything upwards 2400-2500 is really hard to get fully stable on mine.
Thank you again for loads of info.

I don't think the normal use of Prime95 result as high power draw as the torture test, or does it?
AMD has some specially developed power viruses created to simulate worst case (unrealistic) workloads and these softwares result lower power consumption than the most recent version of Prime95 LargeFFT torture test.

The trouble with Prime95 is that the stress it creates can actually damage the CPU, the motherboard or even both.
The CPU package current rating is exceeded as are the ratings of all of the other VRM related components.
Even the EPS12V connector and cables, in case you don't have the ATX12V connector to share the load is at melting point.

Even if you don't manage to burn the VRM for example you will be putting out a lot of extra heat.
If your CPU pulls 220W during Prime95 the VRM alone is dissipating 55W of power.
In overload conditions the VRM efficiency drop to ~ 80% or even below.

Regarding the CNB I haven't tested it at all, only on LN2.
For me it doesn't matter much as long as it is able to do syncronous with 2x8GB DDR-2400 memory.

Regarding Apm(Master): When enabled the Apm estimates the power consumption of the CPU and modulates the clocks in case the TDP limit is reached. Turbo requires Apm as otherwise it doesn't know the power consumption and when to engage or disengage the Turbo Boost. Since the highest boost is only available when half of the CUs are gated the C6 must be enabled in order to make Turbo Core boost work as it should.

You can actually have all of the above without having any TDP restrictions, however thats one of the tricks in my sleeve ("Infinite Turbo").
HPC is actually quite a bogus feature as it's only purpose is to fool to Apm / Turbo.

When the TDP limit is reached the Apm will shuffle down the PStates until the power consumption drops below the TDP limit.
Enabling HPC disables all of the PStates between the Pb0, Pb1, P0 and the throttling PState so that Apm cannot shuffle them down.
That way the CPU frequency stays at P0 (minimum) unless the thermal throttling limits are reached.

btw. The clocks for FX-8370 models:

FX-8370E = 3.3GHz (Base), 3.6GHz (Boost, 4CUs active), 4.3GHz (Boost, 2CUs active, 2CUs gated)
FX-8370 = 4.0GHz (Base), 4.1GHz (Boost, 4CUs active), 4.3GHz (Boost, 2CUs active, 2CUs gated)

[vario]

I don't think the normal use of Prime95 result as high power draw as the torture test, or does it?
AMD has some specially developed power viruses created to simulate worst case (unrealistic) workloads and these softwares result lower power consumption than the most recent version of Prime95 LargeFFT torture test.

...

Regarding the CNB I haven't tested it at all, only on LN2.
For me it doesn't matter much as long as it is able to do syncronous with 2x8GB DDR-2400 memory.

Regarding Apm(Master): When enabled the Apm estimates the power consumption of the CPU and modulates the clocks in case the TDP limit is reached. Turbo requires Apm as otherwise it doesn't know the power consumption and when to engage or disengage the Turbo Boost. Since the highest boost is only available when half of the CUs are gated the C6 must be enabled in order to make Turbo Core boost work as it should.

You can actually have all of the above without having any TDP restrictions, however thats one of the tricks in my sleeve ("Infinite Turbo").
HPC is actually quite a bogus feature as it's only purpose is to fool to Apm / Turbo.

When the TDP limit is reached the Apm will shuffle down the PStates until the power consumption drops below the TDP limit.
Enabling HPC disables all of the PStates between the Pb0, Pb1, P0 and the throttling PState so that Apm cannot shuffle them down.
That way the CPU frequency stays at P0 (minimum) unless the thermal throttling limits are reached.

btw. The clocks for FX-8370 models:

FX-8370E = 3.3GHz (Base), 3.6GHz (Boost, 4CUs active), 4.3GHz (Boost, 2CUs active, 2CUs gated)
FX-8370 = 4.0GHz (Base), 4.1GHz (Boost, 4CUs active), 4.3GHz (Boost, 2CUs active, 2CUs gated)

Well that explains the TDP gap , thanks again.On the Prime95 normal/torture test, i really dont know, i only read that its essentially the same thing, but you're right, that needs to be tested.Another thing is i never actually destroyed anything using prime95, but the test needs to be supervised thats for sure, and beyond aircooling it certainly can be dangerous.
As for the APM, thats pretty much what i observed and was pretty disappointed :-/ and you come here and say there is a WAY to get the best of both worlds and wont tell anyone :P.I guess the best way would be to change the TDP limit ,that way power consumption would be configurable as just it is on newer AMD GFX cards.

[FlanK3r]

I dont much beleive PRIME from two experience.
1) with i7-2600K, PRIME 1h stable without error. But after I was in idle and got sometimes BSOD
2) FX-8150 time, I tried PRIME and got error after few seconds, wtf? Good board (C5F), good cooling (allinone liquid), awesome PSU (gold 1200W AX1200 Corsair) for it.

[vario]

I dont much beleive PRIME from two experience.
1) with i7-2600K, PRIME 1h stable without error. But after I was in idle and got sometimes BSOD
2) FX-8150 time, I tried PRIME and got error after few seconds, wtf? Good board (C5F), good cooling (allinone liquid), awesome PSU (gold 1200W AX1200 Corsair) for it.

Flanker, did you consider maybe that your FX8150 just wasnt stable at stock ?
Cause i know mine wasnt, at first when i had build this FX8320 it WAS NOT stable ta stock, as it turned out after a while, mainboard and its bios was to blame.It is stable now.
As for the idle not stable, i've encountered it also, on different setups, it has mostly to do with quick change of voltage and power states, its also mainly bios problem from what i gather...
Theres also 3) for you, it allows you to brag about higher clocks :-) .Stilt`s way of testing is certainly interesting, it is however complicated to a degree.Im probably gonna do a split testing now, run x264 at 2 modules and 2 with prime, then change the modules, it should still lower power consumption and heat somewhat but maybe i dont have to do the fan speed adjustments which just doesnt work on my board with 2 fan setup on radiator.

EDIT

Hey, STILT, is there a way of enforcing 9590 power consumption (220W)ceiling for a 120W cpu ?

[The Stilt]

Hey, STILT, is there a way of enforcing 9590 power consumption (220W)ceiling for a 120W cpu ?

A short answer would be no, however it can be done with a proper knowledge.
The value is fused so altering it requires using either JTAG or a password protected index register.

It is easier just to disable the TDP limit completely.
In practice you would never even reach the 220W limit on non 9590 grade CPU.
A normal chip would crash due overheating before that.

Frankly I can't see any reason for Turbo being enabled when you overclock.
For me the C6 requirement is the deal breaker as I cannot stand the coil-whine it causes.

[filmbot]

I don't think the normal use of Prime95 result as high power draw as the torture test, or does it?
AMD has some specially developed power viruses created to simulate worst case (unrealistic) workloads and these softwares result lower power consumption than the most recent version of Prime95 LargeFFT torture test.

The trouble with Prime95 is that the stress it creates can actually damage the CPU, the motherboard or even both.
The CPU package current rating is exceeded as are the ratings of all of the other VRM related components.
Even the EPS12V connector and cables, in case you don't have the ATX12V connector to share the load is at melting point.

Even if you don't manage to burn the VRM for example you will be putting out a lot of extra heat.
If your CPU pulls 220W during Prime95 the VRM alone is dissipating 55W of power.
In overload conditions the VRM efficiency drop to ~ 80% or even below.

So what do you use to define fully stability? Are you thoughts the same on other torture test software? (IE OCCT, LinPack, Intel Burn In and Aida)

[vario]

So what do you use to define fully stability? Are you thoughts the same on other torture test software? (IE OCCT, LinPack, Intel Burn In and Aida)

I guess he gave you his method earlier:

I usually set the voltage and clocks to the intended level an start encoding a MKV with X264 (which is the most stressful real world workload).
After around 15 minutes of encoding I check where the temperatures have peaked.
Once I have the temperature information available I start Prime95 Large FFT and set the thread count to four.
Before starting the torture test I set the affinity to the compute units I am going to test during the run (i.e. cores 0-3 = CU0 & CU1).
After starting the test I start reducing the fan rpm until the temperatures reach the same peak level they did during the X264 encode.

After you?re confident that the CPU is stable, repeat the procedure for CU2 & CU3 (cores 4-7).

It takes twice the time the normal method does, but it won't molest your motherboard or CPU beyond the potential breaking point.
Depending on the load line configuration I would recommend that you reduce the core voltage by 6.25 - 12.5mV during the test.
Thats just to compensate the difference in load line effect (droop) you will have from the reduced load from running two compute units instead of all four at once.

I guess the key here is to get the temperature of 8 core load without the 8 core load.

EDIT

I conected to PrimeNET as a worker ,so real world usage ,power consumption is the same as torture Large at least at my current settings (4.0/4.2T APM ON), 252W for whole system in both instances, 100W IDLE (not total idle tho, some tasks running in background as torrent sharing on AES drives).

[FlanK3r]

Reviews FX-8370/8370E around the web:
http://www.computerbase.de/2014-09/a...8370e-im-test/
-OC 5 GHz
http://www.anandtech.com/show/8427/a...ew-vishera-95w
-OC nedělali
http://www.guru3d.com/articles_pages..._review,1.html
-OC 4.7 GHz
http://ht4u.net/reviews/2014/amd_fx-8370e_im_test/
-no OC
http://hothardware.com/Reviews/AMD-F...e-CPU-Reviews/
-bez OC
http://techreport.com/review/26996/a...essor-reviewed
-no OC
http://www.pcper.com/reviews/Process...X-8370e-Review
- FX-9590 5.1 GHz, FX-8370 4.8 GHz with 0.07V offset, FX-8370E 4.5 GHz with 0.06V ofset
http://www.overclockersclub.com/revi...370__fx8370be/
-OC 4990 FX-8370E, 5015 MHz FX-8370
http://benchmarkreviews.com/19942/am...rmance-review/
-no OC

[filmbot]

I guess he gave you his method earlier:


I guess the key here is to get the temperature of 8 core load without the 8 core load.

EDIT

I conected to PrimeNET as a worker ,so real world usage ,power consumption is the same as torture Large at least at my current settings (4.0/4.2T APM ON), 252W for whole system in both instances, 100W IDLE (not total idle tho, some tasks running in background as torrent sharing on AES drives).

Thanks, I must of overlooked that post. Seems quite interesting way to test but does make some natural sense.

[The Stilt]

Linpack <> IBT, OCCT (Small Data Size) are around 90% of the stress that the most recent Prime95 V28.5 LargeFFT torture test generates.

Running any of these, including Prime torture test at stock or even with light overclock is just fine.
However the additional <20% of power consumption at high overclocks is the thing that is dangerous.

I personally don't feel comfortable running any of the software listed above (or any kind of burn-in tests) at settings which result current draw excess 140A.
The high current levels are fine as long as your CPU and motherboard (VRM) are properly cooled. The small stock heatsink installed on the VRM, especially without any proper and direct airflow just won't cut it at these current levels.
On heavily overclocked systems additional, direct airflow to the VRM is mandatory no matter what motherboard you are using. Especially the guys using AIOs should pay special attention to that.
Cooling down the VRM properly will not only increase realiablity and stability but also reduces the CPU temperature and total system power consumption by boosting the VRM efficiency.
The CPU is directly connected to the VRM with a very large amount of copper, which conducts the heat pretty damn well between the two. So improving the cooling on either will reduce the temperature on the other also.

For example the OnSemi 4955N mosfet (which is quite a common low-side fet on AM3r2 motherboards): At 25 degree C it is rated for 48A continuous current while at 130 degree C it is derated to 19A.

Basically your VRM requires (additional) 1/5th of the cooling capacity your CPU has.
i.e. At 200W CPU power draw the VRM cooling needs to handle 30-40W of power.

When I am talking about CPU power consumption I mean the actual amount of power the CPU draws.
The CPU VRM efficiency is around 80-85% and the PSU efficiency can vary from 80% to 94% or so.
So even in the most optimal case where the VRM efficiency would be 85% and the PSU efficiency would be 94%, the power consumption measured from the wall is 250W when the CPU itself draws 200W.

[The Stilt]

Did some additional test on one of the FX-8370E.
Obviously I have done some binning but these are the same chips that are shipping to OEMs and reviewers.
So the only sorting has been done by me.

This is also one of the highest leaking FX-8370E I've got, yet it has 40% lower leakage than an average older Vishera.

The clock scaling is following:

3.6GHz - 1.092V
4.2GHz - 1.212V (SW1)
4.4GHz - 1.260V
4.6GHz - 1.308V (SWC)
4.8GHz - 1.380V
4.9GHz - 1.410V
5.0GHz - 1.440V
5.1GHz - 1.472V

SW1 = Linear switching point 1 (uarch / process)
SWC = Linear switching point critical (uarch / process)

All of the frequencies above are initially Prime95 stable and have been tested on Cryorig R1 Ultimate, which is capable to ~ 0.189C/W performance at 210W.
Starting from 4.9GHz the CPU becomes temperature limited on air so Prime95 cannot be tested with the conventional method.
In this case I dropped the rpm of the fans until the temperatures peaked at 62 degree C.

At 5.1GHz / 1.472V the CPU consumes & dissipates 246.9W of power during Prime95 LFFT so either the cooler should be changed to one with <0.162C/W @ 250W performance or the ambient temperature should be lowered from 25 degree C to 18 degree C in order to keep the tCase temperature below 65 degree C.
So basically anything above 5GHz belong to custom water cooling. No AIO come close to the C/W requirement of 0.162 or below.

Did a quick test on the NB too.
With 1.275V the system boots right up at 2800MHz NCLK however the first L3 subcache seems to be worse than the rest of them.
It fails pretty quickly no matter how much voltage is being fed to it. All of them do 2700MHz at 1.25V thou (yes... Odd NCLK ratios are available too )
So there definitely seems to be some improvement on the NB too from the last time I tested.