OCNG5: OC firmware for Supermicro AMD G34 platforms

rvborgh

Weaksauce
Joined
Jul 1, 2014
Messages
108
6300 are newer with same or lower TDP I think. Then technology for transistors is lower size on newer ones 6300's, but even is with lower techology, current ones, at least in AMD side, have now TDP even up to 400W.
I barely thing that old, or even actual air coolers, will face that. Of course except servers air coolers which run on 16 000 rmps or above... .
It looks we are still doomed on workstation's CPU's to have either very big either insane noizy air coolears. Then liquid coolears need much more imprevements.

Just installed the 6380ES chips in my H8QGi-F tonight. Haven't really played with them much. Just set OCNG to the same settings for memory that i had with the 6328s. Temps stabilize at around 43C with psmax -1 set, with Cinebench R20. The 6328s would top out at around 27C. Ram bandwidth with RightMark Multi thread ram benchmark is down to 32 GB/s from 37 GB/s with the 6328s, but i've not upped the ref clock yet.

Power states for 6380ES:

Main processor is Family 15h (Bulldozer/Interlagos/Valencia) Processor
Family: 0xf Model: 0x2 Stepping: 0x0
Extended Family: 0x15 Extended Model: 0x2
Package Type: 0x3 BrandId: 0x0
Machine has 4 nodes
Processor has 8 cores
Processor has 7 p-states
Processor has 2 boost states

Power States table:
-- Node: 0 Core 0
core 0 pstate 0 (pb0) - En:1 VID:19 FID:18 DID:0.00 Freq:3400 VCore:1.3125
core 0 pstate 1 (pb1) - En:1 VID:32 FID:12 DID:0.00 Freq:2800 VCore:1.1500
core 0 pstate 2 (p0) - En:1 VID:42 FID:9 DID:0.00 Freq:2500 VCore:1.0250
core 0 pstate 3 (p1) - En:0 VID:42 FID:9 DID:0.00 Freq:2500 VCore:1.0250
core 0 pstate 4 (p2) - En:0 VID:42 FID:9 DID:0.00 Freq:2500 VCore:1.0250
core 0 pstate 5 (p3) - En:0 VID:42 FID:9 DID:0.00 Freq:2500 VCore:1.0250
core 0 pstate 6 (p4) - En:1 VID:51 FID:12 DID:1.00 Freq:1400 VCore:0.9125

Power states for 6328

Main processor is Family 15h (Bulldozer/Interlagos/Valencia) Processor
Family: 0xf Model: 0x2 Stepping: 0x0
Extended Family: 0x15 Extended Model: 0x2
Package Type: 0x3 BrandId: 0x0
Machine has 4 nodes
Processor has 4 cores
Processor has 7 p-states
Processor has 2 boost states

Power States table:
-- Node: 0 Core 0
core 0 pstate 0 (pb0) - En:1 VID:19 FID:22 DID:0.00 Freq:3800 VCore:1.3125
core 0 pstate 1 (pb1) - En:1 VID:27 FID:19 DID:0.00 Freq:3500 VCore:1.2125
core 0 pstate 2 (p0) - En:1 VID:35 FID:16 DID:0.00 Freq:3200 VCore:1.1125
core 0 pstate 3 (p1) - En:1 VID:38 FID:12 DID:0.00 Freq:2800 VCore:1.0750
core 0 pstate 4 (p2) - En:1 VID:43 FID:7 DID:0.00 Freq:2300 VCore:1.0125
core 0 pstate 5 (p3) - En:1 VID:46 FID:3 DID:0.00 Freq:1900 VCore:0.9750
core 0 pstate 6 (p4) - En:1 VID:51 FID:12 DID:1.00 Freq:1400 VCore:0.9125
 

rvborgh

Weaksauce
Joined
Jul 1, 2014
Messages
108
i thought i would add some initial impressions.

Initial attempts at overclocking with all 16 cores were not so promising. But that was compounded by the fact that i was switching to the wrong pstate prior to setting the processor voltages. This screwed up the windows timer causing all manner of strange issues, such as the driver blinking the screen on and off, various benchmarks measuring strangely, and my chrome browser all of a sudden going hyperactive. i tried various things to cure this before figuring out that it was my script and not a BIOS setting. Temps were also getting into the 50s as well. So what i did was artificially reduce the # of cores active by setting the "COMPUTE UNIT" in the "CPU Downcore Mode" BIOS setting. This effectively disables CMT, and uses a single core out of a cluster instead of both, halving the cores presented to the OS. This reduced the temps back down.

The nice thing about disabling CMT is that the shared decoder is no longer a possible slowdown factor. i was able to get 8 cores running an all core turbo of 3.5 GHz (pb1), and 4.0 GHz (pb0) with CPB on and PowerNow on. i didn't have to adjust the voltage up from the stock settings. i haven't put too much time into discovering the limits of each core for pb0, however core 0 runs at 4.5 GHz no problem as far as i can tell.

The interesting thing is that when running like this, it is very similar to the 6328s that i previously had in there, but actually faster for some reason. Some benchmarks are 20% faster when running the processors at the same speed. Not sure if this is a result of disabling CMT, or if there are larger caches or what.

One thing is for sure though. i don't think that these 63xx ES chips are anywhere as robust as the 61xx ES chips that i previously had. i do not know why it is, but i cannot even run a single iteration of IntelBurnTest at the stock settings, whereas the 61xx could run those for hours.

These things also seem to be on the ragged edge temps wise. A very fine balancing act between heat and voltage and frequency. Add just a little bit of voltage and the temps really climb. i have no measured power draw at the wall, but it seems these are very power hungry beasts.

I have gotten them running better than the 6328s though now as verified by playing some replays back of a few games. Once i know what the upper frequency limits of the cores are, i'm going to scale them back down and up the ref clock to get a boost on the CPU-NB/Northbridge/L3.
 

rvborgh

Weaksauce
Joined
Jul 1, 2014
Messages
108
Perhaps i should not be amazed by this, but to properly overclock this, i am having to *reduce* voltages. refclock of 210 would not work when running rightmark multi thread benchmark saturating all memory cores. The memory was not the problem, but i noticed when i reduced the CPU-NB voltage adjustment to +0.0 (in turionpowercontrol) that stability improved. Thinking it was the temps, i then reduced the voltage on p-state pb1 - it is the all core turbo that runs when RMMT is working the memory controllers over. This reduced the temps. It seems that once temps hit 33.8C it would give up the ghost. Things stabilized again when i reduced voltage from 1.1500v to 1.0750v which reduced temps by about 1.5C. i wonder how much extra voltage the factory added for the stock p-states over what was necessary.

Anyone else run into this? i'm new to OC for these Bulldozer/Piledriver cores.

Does anyone know whether it is possible to reprogram the TDP value that is used by the APM? Mine is just under 114.xxxx. Since this is an ES processor i wonder if there is a way to set it up to 140w like the 6386SE processors.
 
Last edited:

rvborgh

Weaksauce
Joined
Jul 1, 2014
Messages
108
i thought i would post some preliminary results. Using OCNG the processors seem to have hit a limit at current settings at ref clock 213. Beyond that is not stable at current settings (and i don't want to up voltage to CPU-NB due to increasing heat). RMMT really stresses out the CPUs more in terms of power than say CPU-Z stress or Cinebench. Package power is hitting 71 when running 8 threads hammering the mem controller with 128 bit SSE2 reads. Upping the ref clock seems to have also given me a jump in fps in one of the RTS games i usually play. Perhaps 8% or so.

i turned off HPC Mode in the BIOS and that helped with temps and increased stability.

This is where i am at for stability now with a 213 ref clock: Ends up being Pb0 - 3.621 GHz (single core turbo), and Pb1 - 2.982 GHz (all core turbo).
cd C:\tpc
TurionPowerControl -fo 2
TurionPowerControl -set node all core all pstate 0 freq 3400 vcore 1.2125
TurionPowerControl -set node all core all pstate 1 freq 2800 vcore 1.0500
timeout /t 2
TurionPowerControl -fo 0

i have a more aggressive file that also seems to work just fine (which is my goal - going to be starting with the above and working towards it to prove stability). That one is on the screen capture.
Ends up being Pb0 - 4.473 GHz (single core turbo), and Pb1 - 3.727 GHz (all core turbo). Cinebench 11.5 and R15 both run just fine. i include the CPU-Z benchmark.
This weekend i'll be experimenting using the Kill-a-Watt to see the power draw. i previously ran my Opteron 61xx for about 7 years at 1.2v, so ideally i would like to keep the all core turbo voltages at around that level.

i kind of like running these with the BIOS Downcore mode set to "Compute Unit", as it runs these almost like an FX-8350, but with a larger heat sink area for the same amount of cores.

1.jpg
 
Last edited:

rvborgh

Weaksauce
Joined
Jul 1, 2014
Messages
108
Update: With the processors still set at "Compute Unit" downcore mode i spent some time focusing on the voltage curve and pstate 1 (pb1) since that is where the CPUs spend most of their time.
3.9 GHz required 1.2125 volts
4.01 GHz required 1.2375 volts
i stopped at 4.0 GHz since i don't want to go much beyond 1.2 volts for daily use

i then focused on pstate 0 (pb0) - single core turbo
4.536 GHz requires 1.3125 volts - all of the cores can work at this frequency just fine.
i stopped here.

Settings are Powernow enabled, CPB enabled, HPC mode disabled. ref clock of 211 (i backed off from 213 for a margin of stability)

2.jpg
 

rvborgh

Weaksauce
Joined
Jul 1, 2014
Messages
108
[EDIT] i found out after a bunch of further stability testing, that the lowered speed HT setting just kept the stability from manifesting itself (ie it hid the instability for longer during my Prime95 runs). The problem actually was the Node Interleaving setting in the BIOS. For some reason setting Node interleaving to "disabled" caused the processors to reset under heavy load. Enabling the Node Interleaving setting (back to "Auto" in the BIOS) made the system completely stable. Not sure why node interleaving disabled does not function properly.

i thought i would post about a very strange instability i ran across with these 63xx processors (both the 6328 i previously had in the machine and the 6380ES that are currently in the machine). i did not experience this with the 61xx ES!

My overclocks up to 4.5 GHz (single core turbo) seem just fine on the 6380ES. i could run Prime95 just fine with the small FFT option. This pretty much keeps the benchmark running within the CPUs themselves without much traffic out to ram. No initial problems doing that.
However, i noticed that if i ran any sort of benchmark that saturated the ram. Rightmark is one, and Prime95 with the Blend mode is another, the machine would almost immediately black screen (and sometimes reboot). So what i did was run Prime95/Blend on progressively more cores. i could run 4 threads just fine on cores 0-3, or 4-7, etc. Essentially so long as i kept the computations confined to one of the cpu dies in the MCM package i was fine. if i tried to overlap the cores that is when i would run into issues. This pointed to a communication issue between the cpu dies in the MCM.

Just on that hunch, i downgraded the Hyper Transport speed in the BIOS to HT1. Viola, the instability went away.

Interestingly enough, the ram benchmarks do not see to be much slower if at all. i am going to have to do some comparisons.

But why does this issue manifest with these 63xx processors?

Has anyone done any Hyper transport tuning?

Here is what TurionPowercontrol is currently listing as the HT status:
TurionPowerControl 0.44-rc2 (tpc-0.44-rc2-r144)
Turion Power States Optimization and Control - by blackshard

Hypertransport Status:
Node 0 Link 0 Sublink 0 not connected
Node 0 Link 1 Sublink 0 Bits=16 Coh=0 SpeedReg=7 (1200MHz)
Node 0 Link 2 Sublink 0 Bits=16 Coh=1 SpeedReg=7 (1200MHz)
Node 0 Link 3 Sublink 0 Bits=8 Coh=1 SpeedReg=7 (1200MHz)
Node 0 Link 3 Sublink 1 not connected

Node 1 Link 0 Sublink 0 Bits=8 Coh=1 SpeedReg=7 (1200MHz)
Node 1 Link 0 Sublink 1 not connected
Node 1 Link 1 Sublink 0 Bits=16 Coh=1 SpeedReg=7 (1200MHz)
Node 1 Link 2 Sublink 0 not connected
Node 1 Link 3 Sublink 0 not connected

Node 2 Link 0 Sublink 0 Bits=8 Coh=1 SpeedReg=7 (1200MHz)
Node 2 Link 1 Sublink 0 Bits=16 Coh=0 SpeedReg=7 (1200MHz)
Node 2 Link 2 Sublink 0 Bits=16 Coh=1 SpeedReg=7 (1200MHz)
Node 2 Link 3 Sublink 0 not connected

Node 3 Link 0 Sublink 0 Bits=8 Coh=1 SpeedReg=7 (1200MHz)
Node 3 Link 0 Sublink 1 not connected
Node 3 Link 1 Sublink 0 Bits=16 Coh=1 SpeedReg=7 (1200MHz)
Node 3 Link 2 Sublink 0 not connected
Node 3 Link 3 Sublink 0 not connected


Done.


PS: i wanted to add that the option in OCGN to set the Northbridge multiplier does not seem to work with this 6380ES processor at least. i set it to 9x, but it had no effect.
 
Last edited:
Top