Overclock.net banner

1 - 20 of 178 Posts

·
Premium Member
Joined
·
2,753 Posts
Discussion Starter · #1 ·
Firstly: This fix is by no means an official one and you will be using it at your own risk. If you for any reason manage to fry your card while using it, you are not entitled to make any claims for warranty exchange (RMA).

As soon as the bios flashing software supporting Ellesmere becomes available, I'll be releasing a modified bios which makes the necessary changes automatically.

The procedure is rather simple, and can be done on any software which supports I2C control on Radeon graphics. Personally I recommend MSI Afterburner for the job.

- Download and install the newest MSI Afterburner
- Go to the installation folder of MSI Afterburner (where the MSIAfterburner.exe is located) and open a new command window ("Left Shift + Right Mouse Click" when no file is selected)
- Additional step: Verify that the I2C interface to the VRM controller is working by typing following command: "MSIAfterburner /ri06,08,0D" (without the quotes). If the response is ": 20" or ":44" you're good to go.
- The two commands you need to input are following: "MSIAfterburner /wi06,08,1E,BB" and "MSIAfterburner /wi06,08,1F,B0".

Once you have given these commands the power draw from the PCI-E slot has reduced and moved towards the PCI-E power connector. The recommended programming value is "B" (/wi06,08,1E,BB and /wi06,08,1F,B0), however you can go set it to anywhere between "B" and "F". Higher the value, more of the power draw will be directed to the PCI-E power connector.

Do note that this fix must be re-applied after every reboot.

EDIT: Added the I2C response for IR3567B controller (0x44).
IR3567A = 0x20.

The fix applies for both!


////////////////////////////////////

Update (alternative, more permanent way)

EDIT: Removed. Currently the Polaris bioses appear to be signature protected, which prevents any kind of modifcations to tables affecting the frequencies, voltages or power limits. The modified bioses will work otherwise normally, however the display driver will not load. The modified bioses must be sent to AMDs signature server, which recalculates the new security signature for the bios contents. A service which is not available for end-users, obviously.
 

·
*cough* Stock *cough*
Joined
·
2,786 Posts
This is very interesting. I do not have a 480 to test on.

Thanks for posting this.
thumb.gif
 

·
Premium Member
Joined
·
2,753 Posts
Discussion Starter · #5 ·
Quote:
Originally Posted by gupsterg View Post

+rep
thumb.gif
, sweet share for aiding the AMD community.
Thanks, there was indeed an error. Luckily only in the sample, not in the actual command
biggrin.gif
 

·
Meddling user
Joined
·
7,423 Posts

·
Registered
Joined
·
5 Posts

·
Premium Member
Joined
·
2,753 Posts
Discussion Starter · #9 ·
Quote:
Originally Posted by I-o-I View Post

Wow, thanks for that.

It would be nice if you could elaborate more on which component you talk to and what the values mean.

I assume you are talking to the PWM controller?
So these commands change the default configuration of the IR3567 VRM controller, which is used on all of the reference boards. There are six phases for the GPU power plane (VDDC) and half of these phases are fed from the PCI-E power connector and the other half from the PCI-E slot 12V source. At the default configuration the cards are shipped with, each of the VRM phases handle the same amount of load. The changes made by these commands affect how the load is distributed between the VRM phases.

With the configuration enabled by these commands, the half of the VRM which is fed from the PCI-E power connector will be taking care of a larger portion of the load than the other half. Since the total power draw of the GPU will remain intact, the power draw from the PCI-E slot will decrease and the power draw from the PCI-E power connector will increase.

This will increase the stress on the other half of the VRM, however the operating parameters will still remain safely below any risk limits. The per phase overcurrent limit on these cards is configured extremely conservatively (36A per phase) and even with the changed configuration there is no need to increase these safety limits from their default values. Also the over temperature protection (OTP) limits are left intact (115°C for forced throttling, 130°C for immediate shutdown).
 

·
Premium Member
Joined
·
2,753 Posts
Discussion Starter · #10 ·
Quote:
Originally Posted by rollingdice View Post

Aw, my card's response is :44. I guess it's unable to shift the load to the 6-pin connector?
+rep anyway for the tips
smile.gif
This is my bad actually!
IR3567A response is 20 and IR3567B is 44.
You're good to go.
 

·
Registered
Joined
·
5 Posts
thanks for the reply
thumb.gif
Quote:
Originally Posted by The Stilt View Post

So these commands change the default configuration of the IR3567 VRM controller, which is used on all of the reference boards. There are six phases for the GPU power plane (VDDC) and half of these phases are fed from the PCI-E power connector and the other half from the PCI-E slot 12V source. At the default configuration the cards are shipped with, each of the VRM phases handle the same amount of load. The changes made by these commands affect how the load is distributed between the VRM phases.
Yes, I understand those basics.
However, Igor from Tom's Hardware deduced that actually 4 phases are connected to the slot and only 2(+1) to the 6-pin connector.

source (sorry, German): http://www.tomshardware.de/amd-radeon-rx-480-grafikkarte-leistungsaufnahme-pci-sig,testberichte-242143.html

This makes more sense, too, imho, when you say "each of the VRM phases handle the same amount of load", looking at the measured total power consumption. How can you be sure the balance is actually 50/50? I'm not trying to take a swipe at you, just trying to understand.
Also, how did you validate those results? Do you have equipment to measure the draw from PEG/PCIe 6 Pin?
Quote:
Originally Posted by The Stilt View Post

With the configuration enabled by these commands, the half of the VRM which is fed from the PCI-E power connector will be taking care of a larger portion of the load than the other half. Since the total power draw of the GPU will remain intact, the power draw from the PCI-E slot will decrease and the power draw from the PCI-E power connector will increase.
Yes, I thought that. But what I was searching for (out of curiosity) is what do the exact parameters mean? And how is the exact load distribution between the phases when choosing values from B to F. Can you split the command value and explain the single bits?
 

·
Premium Member
Joined
·
2,753 Posts
Discussion Starter · #13 ·
Quote:
Originally Posted by I-o-I View Post

thanks for the reply
thumb.gif

Yes, I understand those basics.
However, Igor from Tom's Hardware deduced that actually 4 phases are connected to the slot and only 2(+1) to the 6-pin connector.

source (sorry, German): http://www.tomshardware.de/amd-radeon-rx-480-grafikkarte-leistungsaufnahme-pci-sig,testberichte-242143.html

This makes more sense, too, imho, when you say "each of the VRM phases handle the same amount of load", looking at the measured total power consumption. How can you be sure the balance is actually 50/50? I'm not trying to take a swipe at you, just trying to understand.
Also, how did you validate those results? Do you have equipment to measure the draw from PEG/PCIe 6 Pin?
Yes, I thought that. But what I was searching for (out of curiosity) is what do the exact parameters mean? And how is the exact load distribution between the phases when choosing values from B to F. Can you split the command value and explain the single bits?
I gave the instructions how to measure the VRM structure on RX 480 to a fellow Finnish hardware site author, who received the card for review. Based on the measurements he made based on my instructions the split is indeed 50/50 for the VDDC (GPU) power plane. The drains of the first three (of the green region) are connected together and connected to the PCI-E power connector. The drains of the lower three high-side fets are also connected together and connected to the PCI-E slot 12V pins.

The changes have been tested by w1zzard (TPU site author) who reported reduced PCI-E slot power draw with these settings.

The region highlighted in green are the high-side fets for the VDDC (GPU) power plane. The single high-side and low-side fet above these is for VDDCI (not affected). Also the picture you posted contains an error (besides the distribution). The phases go from 1-6 in descending order, not in ascending order.

 

·
Premium Member
Joined
·
2,753 Posts
Discussion Starter · #16 ·
Quote:
Originally Posted by I-o-I View Post

Yes, I thought that. But what I was searching for (out of curiosity) is what do the exact parameters mean? And how is the exact load distribution between the phases when choosing values from B to F. Can you split the command value and explain the single bits?
I forgot to answer (kind of) to this one.
The IR3567 is a 6+2 phase controller. There are eight control values (one for each phase) in total which determine the current imbalance between the phases. The second output (+2) is unused on RX 480 so that leaves us six effective control values. When the control values are all zero, each and every active phase will take care of an equal slice of the total load. That's the default configuration these cards are using. When any of these control values are programmed to any non zero (1-15) value, it means that the phase with this control value programmed will take care of a larger slice of the total load than the ones with the control value set to zero. In this case the control value has been programmed for phases 1, 2 and 3. Therefore phases 1 - 3 will do more work than phases 4 - 6. Since phases 1-3 are fed from the PCI-E power connector the load from the PCI-E slot will decrease.

Unfortunately I cannot disclose how the programming values are decoded, since that is IR confidential.
 

·
Registered
Joined
·
2,075 Posts
thank you Stilt for posting this
 
1 - 20 of 178 Posts
Top