Hashrate RX 470 - page 30. | Bitcointalksearch.org

Wolf0

member

Activity: 81

Merit: 1002

It was only the wind.

Quote from: B1tUnl0ck3r on August 27, 2016, 08:28:02 AM

Quote from: Longsnowsm on August 26, 2016, 04:33:22 PM

Quote from: B1tUnl0ck3r on August 26, 2016, 04:22:56 PM

Quote from: Longsnowsm on August 25, 2016, 04:02:54 PM

I downloaded the Trixx Sapphire app and tried it and it does not appear to give us any control over the 470 cards other than the power limit which we have under Wattman so no help at all from what I can see. It will let you slide the controls and looks like it is doing something, but you go to the hardware monitor and you will see the gpu and memory are still at stock settings. Only the power limit made any real change in the output of the hardware monitor page. So far Trixx is useless, we had this much control in Wattman.

Sad to read it. However what ever a software says, I think it's better to check with a watt-meter at the wall... it's always more accurate in my opinion.

I was hoping we would get control over gpu core and memory speeds. It does look like the undervolt setting does something in Trixx, but appears to be the extent of it so far. Not the kind of control I was hoping for.

A card without gpu voltage control (without bios flash or mod) is a clear no buy for me. I would even more like to be able to tweak memory voltage to get the most of the chips... be it in efficiency or raw power...

Easy in Linux. They have a sysfs entry that lets you control all of it through the PowerPlay table.

Here's a screen of a tool I wrote to do so (NSFW): https://ottrbutt.com/tmp/wolfamdctrl-priv.png

Needs more work and options, though.

zorvalth

full member

Activity: 223

Merit: 100

Can somebody give me the stock bios of RX 470 Nitro+ 4GB ?

niko2004x

member

Activity: 126

Merit: 10

Quote from: Grim on August 30, 2016, 04:59:38 AM

Another strange behaviour I noticed with a 470 Nitro 8GB:

(I assume this is a driver problem.) 16.7.3

If you set the cards above 2000 memory clock, or rather 2050 it seems,
the cards actually slow down due to some power limit?

eg. setting 2200 memory clock starts hashing fine but after a few seconds the cards slow down to approx. 85 Watt core voltage. Upping the tdp in bios and setting powerlimit to +50 doesn't help at all.

Is there a atikmdag.sys for the newest 16.8.2 driver?

Noticed something similar in Linux (for Saphire RX470 Nitro 8GB, 11256-02-20G).
Starting from around 65 degrees of core hashrate starts to gradually fall
and at 70+ degrees of core hashrate loses ~0.3 Mh/s.

However voltages and clocks does not change.

Grim

sr. member

Activity: 505

Merit: 252

Another strange behaviour I noticed with a 470 Nitro 8GB:

(I assume this is a driver problem.) 16.7.3

If you set the cards above 2000 memory clock, or rather 2050 it seems,
the cards actually slow down due to some power limit?

eg. setting 2200 memory clock starts hashing fine but after a few seconds the cards slow down to approx. 85 Watt core voltage. Upping the tdp in bios and setting powerlimit to +50 doesn't help at all.

Is there a atikmdag.sys for the newest 16.8.2 driver?

z0n0

legendary

Activity: 1274

Merit: 1006

Quote from: anatolikostis on August 30, 2016, 04:10:04 AM

470 @jet @29mh/s

Which card version is that?

anatolikostis

legendary

Activity: 2026

Merit: 1005

rx470 MSI 4Gb @jet @29mh/s

peg-slot @5A Shocked

GDDR5 voltage is 1500mV

fr4nkthetank

legendary

Activity: 2294

Merit: 1182

Now the money is free, and so the people will be

hey guys, so i've got 2 470s, nitros with the dual bios switch, I've been playing around to modify myself in polaris editor, timings, etc (win7). but for some god-forsaken reason I cant get the system to run the card with the modified bios. I've changed to ati system file to the modified one that was posted in safe mode, boot with option 7 no sig check, flashed the bios, used the tool to enable test mode and sign the system files, etc. im kinda stumped, can anybody offer some IT support lol

niko2004x

member

Activity: 126

Merit: 10

Quote from: Wolf0 on August 28, 2016, 07:16:55 PM

Quote from: niko2004x on August 28, 2016, 06:21:06 PM

Quote from: Eliovp on August 28, 2016, 05:47:32 PM

Quote from: Wolf0 on August 28, 2016, 02:18:49 PM

Quote from: Grim on August 28, 2016, 12:47:42 PM

afaik the rated voltage for eg. samsung memory at 2000mhz is 1.35 volt.

Older gddr5 chips were normally 1.5 or even 1.6 volt.

So setting it to 800mv without crashing is a sure sign it doesn't work.

Which is a pitty because most performance could probably be squeezed by upping mvdd. (not vddci)

Yeah, I tested myself and I do believe you're correct. It does seem to work on Fiji, IIRC, but probably because HBM.

Mentioned this a few times already, but i'll do it again Wink

Fiji and Those Lightning cards are the only ones capable of directly modifying the memory voltage (MVDDC or how it has been referred here "MVDD").

For every other card out there (i could have missed one or two) you can only raise or reduce the voltage given to the controller (VDDCI/AUX).
This isn't even stock, to be able to do this, you'll need to add the needed offsets in the rom, therefor also needing to remove extra whitespace in the rom to get the exact same size again. Not only that, the tables in the rom have to be modified as well. i2cdump ...

So basically, at this point, by opening polariseditor or wattman or whatever, and changing the voltage value next to the memory is not going to apply directly to the memory, but to the controller, so basically, giving that controller extra voltage does not affect your memory directly as that memory isn't getting extra voltage.

Wink

Could we just change vddci in Linux using pp_table?
After parsing it seems vddci is there:
Container:
ucRevId = 0
ucNumEntries = 2
ATOM_MCLK_ENTRY = [
Container:
ucVddcInd = 0
usVddci = 850
usVddgfxOffset = 0
usMvdd = 1000
ulMclk = 30000
usReserved = 0
Container:
ucVddcInd = 15
usVddci = 950
usVddgfxOffset = 0
usMvdd = 1000
ulMclk = 220000
usReserved = 0
]

This is only for the controller, and other shit (AUX voltage.)

Tested setting values for vddci other than 950 for mclk 2200MHz.
Immediate crash.

niko2004x

member

Activity: 126

Merit: 10

Quote from: Eliovp on August 28, 2016, 05:47:32 PM

Quote from: Wolf0 on August 28, 2016, 02:18:49 PM

Quote from: Grim on August 28, 2016, 12:47:42 PM

afaik the rated voltage for eg. samsung memory at 2000mhz is 1.35 volt.

Older gddr5 chips were normally 1.5 or even 1.6 volt.

So setting it to 800mv without crashing is a sure sign it doesn't work.

Which is a pitty because most performance could probably be squeezed by upping mvdd. (not vddci)

Yeah, I tested myself and I do believe you're correct. It does seem to work on Fiji, IIRC, but probably because HBM.

Mentioned this a few times already, but i'll do it again Wink

Fiji and Those Lightning cards are the only ones capable of directly modifying the memory voltage (MVDDC or how it has been referred here "MVDD").

For every other card out there (i could have missed one or two) you can only raise or reduce the voltage given to the controller (VDDCI/AUX).
This isn't even stock, to be able to do this, you'll need to add the needed offsets in the rom, therefor also needing to remove extra whitespace in the rom to get the exact same size again. Not only that, the tables in the rom have to be modified as well. i2cdump ...

So basically, at this point, by opening polariseditor or wattman or whatever, and changing the voltage value next to the memory is not going to apply directly to the memory, but to the controller, so basically, giving that controller extra voltage does not affect your memory directly as that memory isn't getting extra voltage.

Wink

Could we just change vddci in Linux using pp_table?
After parsing it seems vddci is there:
Container:
ucRevId = 0
ucNumEntries = 2
ATOM_MCLK_ENTRY = [
Container:
ucVddcInd = 0
usVddci = 850
usVddgfxOffset = 0
usMvdd = 1000
ulMclk = 30000
usReserved = 0
Container:
ucVddcInd = 15
usVddci = 950
usVddgfxOffset = 0
usMvdd = 1000
ulMclk = 220000
usReserved = 0
]

Wolf0

member

Activity: 81

Merit: 1002

It was only the wind.

Quote from: bobben2 on August 25, 2016, 03:58:25 PM

Quote from: Wolf0 on August 25, 2016, 03:45:32 PM

Quote from: bobben2 on August 25, 2016, 03:35:11 PM

Quote from: Wolf0 on August 25, 2016, 03:03:00 PM

Quote from: blue1984 on August 25, 2016, 02:38:09 PM

@wolf0
How can i set the offset in the bios ?

Hex editor.

Quote from: Grim on August 25, 2016, 02:32:11 PM

Quote from: Wolf0 on August 25, 2016, 02:22:55 PM

You can, but you need to offset the voltage, not use the DPM mV.

May I ask if that is true for all rx4.. cards? (reference as well as custom models)

Actually I would prefer to overwrite dpm states with true values ...

It's actually better to offset even on other GPUs - overwriting the DPM state's index into the voltage table is a heavy-handed way to do it, IMO - better to let the offset step down the voltage.

Hi Wolf,
It sounds like you have got this right..
I am trying as well... Could you pls shed some light on the undervolt detail?

If I look in the source code of the Polaris editor there is a data structure definition there called

unsafe struct ATOM_SCLK_ENTRY
   {
   public Byte ucVddInd;
   public UInt16 usVddcOffset;
   public UInt32 ulSclk;
   public UInt16 usEdcCurrent;
   public Byte ucReliabilityTemperature;
   public Byte ucCKSVOffsetandDisable;
   public UInt32 ulSclkOffset; // Polaris Only, remove for compatibility with Fiji
   };

So this is the table structure for the core clock values for each DPM state.
So I wrote my own utility since undervolt editing is missing in the editor itself.
I tried to assign -16 to usVddcOffset field in the DPM7 entry hoping to get a 100 mV undervolt, but it had no effect.
I am also curious about the ucCKSVOffsetandDisable field. It is zero. Should be non-zero to "enable" undervolting to work?

Finally! Someone willing to do some research! :3

It might have had no effect because it's defined as a uint32, but actually should be interpeted as a 32-bit SIGNED integer.

Since you're the first one to actually work at the suggestion, I'll tell you this: ucCKSVOffsetandDisable is a bitmask. Bits 0 - 6 are a voltage offset, and 7 is enable/disable.

Ok, So bits 0-6 is a 7-bit signed value? with range -64 -- +63
If not, then how would I get to a negative offset of say -100mV?

I believe it is - but I'm not quite sure what that voltage offset is for.

Eliovp

legendary

Activity: 1050

Merit: 1293

Huh?

Quote from: Wolf0 on August 28, 2016, 02:18:49 PM

Quote from: Grim on August 28, 2016, 12:47:42 PM

afaik the rated voltage for eg. samsung memory at 2000mhz is 1.35 volt.

Older gddr5 chips were normally 1.5 or even 1.6 volt.

So setting it to 800mv without crashing is a sure sign it doesn't work.

Which is a pitty because most performance could probably be squeezed by upping mvdd. (not vddci)

Yeah, I tested myself and I do believe you're correct. It does seem to work on Fiji, IIRC, but probably because HBM.

Mentioned this a few times already, but i'll do it again Wink

Fiji and Those Lightning cards are the only ones capable of directly modifying the memory voltage (MVDDC or how it has been referred here "MVDD").

For every other card out there (i could have missed one or two) you can only raise or reduce the voltage given to the controller (VDDCI/AUX).
This isn't even stock, to be able to do this, you'll need to add the needed offsets in the rom, therefor also needing to remove extra whitespace in the rom to get the exact same size again. Not only that, the tables in the rom have to be modified as well. i2cdump ...

So basically, at this point, by opening polariseditor or wattman or whatever, and changing the voltage value next to the memory is not going to apply directly to the memory, but to the controller, so basically, giving that controller extra voltage does not affect your memory directly as that memory isn't getting extra voltage.

Wink

Grim

sr. member

Activity: 505

Merit: 252

afaik the rated voltage for eg. samsung memory at 2000mhz is 1.35 volt.

Older gddr5 chips were normally 1.5 or even 1.6 volt.

So setting it to 800mv without crashing is a sure sign it doesn't work.

Which is a pitty because most performance could probably be squeezed by upping mvdd. (not vddci)

niko2004x

member

Activity: 126

Merit: 10

Quote from: Wolf0 on August 27, 2016, 08:57:39 PM

Quote from: niko2004x on August 27, 2016, 08:28:47 PM

Quote from: Wolf0 on August 27, 2016, 08:22:08 PM

Quote from: niko2004x on August 27, 2016, 08:10:34 PM

Quote from: Wolf0 on August 27, 2016, 09:53:34 AM

Quote from: B1tUnl0ck3r on August 27, 2016, 08:28:02 AM

Quote from: Longsnowsm on August 26, 2016, 04:33:22 PM

Quote from: B1tUnl0ck3r on August 26, 2016, 04:22:56 PM

Quote from: Longsnowsm on August 25, 2016, 04:02:54 PM

I downloaded the Trixx Sapphire app and tried it and it does not appear to give us any control over the 470 cards other than the power limit which we have under Wattman so no help at all from what I can see. It will let you slide the controls and looks like it is doing something, but you go to the hardware monitor and you will see the gpu and memory are still at stock settings. Only the power limit made any real change in the output of the hardware monitor page. So far Trixx is useless, we had this much control in Wattman.

Sad to read it. However what ever a software says, I think it's better to check with a watt-meter at the wall... it's always more accurate in my opinion.

I was hoping we would get control over gpu core and memory speeds. It does look like the undervolt setting does something in Trixx, but appears to be the extent of it so far. Not the kind of control I was hoping for.

A card without gpu voltage control (without bios flash or mod) is a clear no buy for me. I would even more like to be able to tweak memory voltage to get the most of the chips... be it in efficiency or raw power...

Easy in Linux. They have a sysfs entry that lets you control all of it through the PowerPlay table.

Here's a screen of a tool I wrote to do so (NSFW): https://ottrbutt.com/tmp/wolfamdctrl-priv.png

Needs more work and options, though.

Does it work?
I tried to change vddci and mvdd to values below 1000 (by setting new pp_table) and
according power meter and temperature measurements nothing happened.

You won't see those two as changing for two possible reasons - one, MVDD and VDDCI are small beans compared to core voltage, and two, you have to be running at least 4.8rc2 for it to apply.

Yes, running 4.8rc2.
The main reason I suspect it does not work is setting mvdd to <800 for 2200MHz memory does not result in crash.

Pretty sure it won't allow going THAT low - I raised MVDD and it seemed to make a high memclk more stable. I know for a fact the clock changing and DPM changes work, as I've observed them through heat and hashrate.

Yes, clock changing does work.
As for mvdd, tested by setting (through pp_table) values from 800 to 1000 step 10.
No visible changes at all except pp_table (read it back to check if it was changed; it was as expected).