Question XFX 5700 XT Cards

[Anthony]

Preface: @Mini_Me has demonstrated profound knowledge when modding the 5700 XT series of graphics cards, and for those looking for an awesome tutorial, please reference https://www.igorslab.de/community/threads/gigabyte-5700-xt-bios-mod-fails.3304/.

Story: At the present time, I own 40 XFX 5700 XT video cards; I have 40 more on order to make a total quantity of 80 cards. With this many cards, squeezing out an additional 1 MegaHash (MH) from each card would translate into +80MH. Out of the box, the stock bios on these XFX 5700 XT cards reflects about 51MH per card. With bios modifications and tweaks, it is possible for me to obtain between 54MH-56MH per card.

Configuration: I have 5 rigs in operation and soon I will have 5 more in operation. Each rig uses 8 cards. I am using HiveOS to manage my rigs.

The importance of bios modifications: Basically, using bios modifications allows each card to provide a better hashrate -- and as hashrate increases, potential ETH mined increases. (More Hash == More $$). Bios modifications also change how each card consumes electricity -- and aside from the initial cost of computer components, electricity is the second greatest cost when it comes to mining.

• Stock Bios: 80 Cards x 51MH from each card = 4.080GH
• Modified Bios: 80 Cards x (54MH || 56MH) = (4.320GH || 4.480GH).

The delicate balance in bios modification reflects a specific relationship between: 1.) Power Consumption, 2.) Temperature, 3.) Hashrate, and 4.) Stability vs. Instability. In my opinion, the goal of proper bios modification means that: (a.) each card consumes as little power as necessary, (b.) each card remains within "safe" temperatures to prevent card damage, (c.) each card produces the maximum hashrate, (d.) each card remains stable, which is to say that the card does not produce "invalid" shares or other errors that get in the way of making money.

The above mentioned thread link reflects a dialogue between two members, @Mini_Me and @AlleyCat. The thread is several pages long because it reflects a series of tests (trials, errors, and new trials with new errors, etc.) with different bios values and with different values in HiveOS. There are multiple variables in play (both on the bios tweaking side with MPT and RBE and on the software tweaking side in HiveOS), and changing one variable influences other variables. The thread is so long because many different settings were tried out in order to achieve the best possible result for specific card manufacturers, in this case manufacturer Sapphire and manufacturer Gigabyte.

This post in particular captured my attention: https://www.igorslab.de/community/threads/gigabyte-5700-xt-bios-mod-fails.3304/post-79315.

RBE modified values in use:

Option 1 (Preferable and Recommended), applying Apple Inc. vram timings straps linked below, once for MT61K256M32 Micron, and Save the vbios after that load the saved vbios and once again apply the straps for K4Z80325BC Samsung if it is existed and save the vbios again,
https://www.igorslab.de/community/a...-mt61k256m32_gddr6_optimized_timings-zip.6544

MPT modified values in use:

Features Tab:
PPTable Features -> Feature Control = Nothing done. Left at default.
Overdrive Features = All Boxes Checked.

Overdrive Limits Tab:
GFX Maximum Clock = 1550
Memory Maximum Cock = 1000
Power Limit Maximum = 0
Power Limit Minimum = 0
Memory Timing Control = 1
Fan RPM Maximum = 3500
Fan RPM Minimum = 1100
Fan Acoustic Limit RPM Maximum = 3500
Fan Acoustic Limit RPM Minimum = 1100
Zero RPM Control = 0

Power and Voltage Tab:
Maximum Voltage GFX = 900
Maximum Voltage SoC = 1150
Minimum Voltage GFX = 700
Minimum Voltage SoC = 750
Power Limit GPU = 140
TDC Limit GFX (A) = 140
TDC Limit SoC (A) = 14

Frequency Tab:
GFX Maximum = 1400
GFX Minimum = 300
SoC Maximum = 1267
SoC Minimum = 507
Memory DPM 0 = 100
Memory DPM 1 = 500
Memory DPM 2 = 625
Memory DPM 3 = 960

Curve Tab:
(Left Alone, no change).
AVFS (GHz->V)
Override box is not ticked.
a = 0.017810
b = -0.047280
c = 0.054020
StaticVoltageOffset (GHz->V) = 0.000000

Fan Tab:
PWM Minimum = 15
Fan Acoustic Limit RPM = 1550
Fan Throtteling RPM = 3200
Fan Maximum RPM = 3500
Fan Target Temperature = 85
Fan GFX Clock = 800
Zero RPM Enable Box is ticked.
Stop Temperature = 60
Start Temperature = 68

As you can see, in order to achieve stability (prevent invalid shares, rejected shares, and phoenixminer reboots), I have to run most cards at 1375 core. Other cards I had to reduce to 1325 core. Some cards run at memory 930, some at 905, some at 900, and some at 890.

I know my cards can give better performance. Even though I have reduced power values to reduce # of invalid-rejected shares, phoenixminer still reboots from time to time, and sometimes, I will have a card that will "stop" and disappear ... which causes the phoenixminer rebooter. For these cards, I reduce values.

Any advice from master @Mini_Me would be appreciated.

RIG #1


RIG #2


RIG #3


RIG #4


RIG #5

 

[w0rtez]

You can measure the consumption, at wall, of the complete rig (mining) and then remove the card and do the measurement again.

I know either way i have to stop the mining rig and start running it again for at least 30mins so it can measure it up well.... I just ordered a riser to connect that RAW II because i got one defective and i will have to shutdown the rig... I will measure it for you as soon as i get it so will let you know...

 

[aeix]

I have a feeling Thicc cards are different than the Triple Dissipation ones. I've done a lot of tests on this card and it definitely requires 150W to get 57. Also, you cannot trust what hiveOS says, that number can be off by up to 20%.

 

[aarakaki]

I know either way i have to stop the mining rig and start running it again for at least 30mins so it can measure it up well.... I just ordered a riser to connect that RAW II because i got one defective and i will have to shutdown the rig... I will measure it for you as soon as i get it so will let you know...

Thanks =)

Folks

I found something interesting, measured the Watts for the cards and checked the following:

Total rig consumption [at wall]	790,6W
Without XFX 5700XT Triple Dissipation (micron) [at wall]	632,0W
Without 5700XT Sapphire Pulse (micron) [at wall]	635,6W

Card	Watts on Hiveos	Watts at Wall
XFX 5700XT Tripple Dissipation (micron)	135W	158,6W
5700XT Sapphire Pulse (micron)	96W	155W

I think the way TDP, TBP and TGP is being reported is different on this card ...

I am curious because my measurements indicate that although on HiveOS it indicates the consumption of 96W for the 5700XT Sapphire Pulse, at wall corresponds to 155W.

While on the XFX 5700XT Triple Dissipation, HiveOS indicates 135W but at the wall it is 158W.

Image 1 - Rig with all Cards:





Image 2 - Rig without Sapphire:



Image 3 - Rig Without XFX:

 

[aeix]

Yes this is what I am talking about, you can't trust the HiveOS reported numbers. Although I wonder how the sapphire card is pulling 155 even thought people are setting TDP to 124 on MPT. Doesn't that limit the power consumption ? If you think about it however, 100W isn't very realistic, that's like a RX580 wattage.

 

[aarakaki]

Yes this is what I am talking about, you can't trust the HiveOS reported numbers. Although I wonder how the sapphire card is pulling 155 even thought people are setting TDP to 124 on MPT. Doesn't that limit the power consumption ? If you think about it however, 100W isn't very realistic, that's like a RX580 wattage.

Exactly, in the tests I am doing this version of the XFX Tripple Dissipation, it is measuring more correctly than the other AMD cards, but this influences the BIOS settings.

In this version we have of XFX the measurement on HiveOS is much closer to the real one, with a level of reliability similar to what NVidia cards indicate on HiveOS.

 

[aarakaki]

Yes this is what I am talking about, you can't trust the HiveOS reported numbers. Although I wonder how the sapphire card is pulling 155 even thought people are setting TDP to 124 on MPT. Doesn't that limit the power consumption ? If you think about it however, 100W isn't very realistic, that's like a RX580 wattage.

Perhaps, for some reason, this version of the XFX Tripple Dissipation has reformulated how TDP, TBP and TGP on this AMD cards are reported:

Power Consumption: TDP, TBP and TGP for Nvidia and AMD graphics cards calculated with destruction of a PR film | igorsLAB | igor´sLAB

What is really behind the information about Thermal Design Parameter (TDP) or GPU Power and Total Board Power (TBP) and Total Graphics Power (TGP), amD and Nvidia graphics cards, and who adheres to…

www.igorslab.de

 

[Mini_Me]

Yes, but if i follow the steps on page 32, the 124W on Power Limit GPU limit the eficient of this card... I think there is something wrong with the api of this card.. I only get the correct hashrate with 140W on Power Limit GPU, and at wall consumption is equivalent to Sapphire 5700xt :

Anhang anzeigen 9716

I would appreciate if you could re do the guide as DS_SOCCLK and DS_DECFCLK were removed as they do not work well with Micron memory.

 

[aeix]

Think it's safe to say that people who think they're getting 58mhs at 100W aren't actually. That number is way off. They are pulling at least 150 at the wall no matter what TDP you set them at in MPT.

 

[Mini_Me]

Think it's safe to say that people who think they're getting 58mhs at 100W aren't actually. That number is way off. They are pulling at least 150 at the wall no matter what TDP you set them at in MPT.

that what is expected,

150 TGP * 0.83 (17% losses) = 124.5 W
124.5 - 16 W memory - 6 W controllers - 8 W Fan = 94.5 W TDP

I added 30 W to compensate the losses.

 

[aarakaki]

I would appreciate if you could re do the guide as DS_SOCCLK and DS_DECFCLK were removed as they do not work well with Micron memory.

Same issue... The "Power Limit GPU: 124 W" on this card limits the ideal capacity.

that what is expected,

150 TGP * 0.83 (17% losses) = 124.5 W
124.5 - 16 W memory - 6 W controllers - 8 W Fan = 94.5 W TDP

I added 30 W to compensate the losses.

But using only this projection doesn't match the data I got (at wall) from my XFX, only apply for Sapphire model....

I think, for some reason, this XFX Triple Dissipation has a power consumption count much like NVidias TGP (closer "at wall"), it contrast from other AMD 5700XT cards, influencing the limit imposed by the bios to regulate the Power Limit....

 

[Mini_Me]

Same issue... The "Power Limit GPU: 124 W" on this card limits the ideal capacity.

Anhang anzeigen 9745




But using only this projection doesn't match the data I got (at wall) from my XFX, only apply for Sapphire model....

I think, for some reason, this XFX Triple Dissipation has a power consumption count much like NVidias TGP (closer "at wall"), it contrast from other AMD 5700XT cards, influencing the limit imposed by the bios to regulate the Power Limit....

I believe here they made the TDP as the TGP.

 

[aarakaki]

I believe here they made the TDP as the TGP.

It is a very plausible hypothesis ... would it be a design error eligible for recall?

 

[aeix]

Why would that be a design error? This is why we are asking for power draw at the wall using a power meter. This is the only way to measure power usage. Software based energy measurement is not accurate.

 

[aarakaki]

Why would that be a design error? This is why we are asking for power draw at the wall using a power meter. This is the only way to measure power usage. Software based energy measurement is not accurate.

You are right, I guess because it is different from other cards, many of them from the XFX itself for obtaining power consumption, I thought this was standardized by AMD for all AIBs to have a design standard ...

 

[Mini_Me]

It is a very plausible hypothesis ... would it be a design error eligible for recall?

I believe they intentionally did not want the card to go into deep sleep to erase any possible stuttering when running at high load and they implement this in the pcb design.

 

[Mini_Me]

Why would that be a design error? This is why we are asking for power draw at the wall using a power meter. This is the only way to measure power usage. Software based energy measurement is not accurate.

TDP purpose is as the minimum power limit, TGP is as the maximum power limit for the card, some XFX cards TDP = TGP for the reason mentioned before.

 

[dreptroman]

So bottom line is that SOME XFX Micron need 150w at the wall or more and that may not be fixable by bios update?

Also. If Micron memory, we need to do the guide without DS_SOCCLK and DS_DECFCLK manipulation. Correct?

I am still waiting for some other cards. I consider myself lucky to have gotten samsung memory on these ones..

 

[aeix]

from what it sounds like, other cards are likely pulling 150 from the wall too. Noone has 5700XTs pulling 100W from the wall. What hiveOS seems to be showing for those cards is the TDP which was calculated above from the TGP.

 

[Mini_Me]

from what it sounds like, other cards are likely pulling 150 from the wall too. Noone has 5700XTs pulling 100W from the wall. What hiveOS seems to be showing for those cards is the TDP which was calculated above from the TGP.

Yes, you are correct however HiveOS reading is not accurate and it is better to use other means to calculate the power draw.

 

[Mini_Me]

from what it sounds like, other cards are likely pulling 150 from the wall too. Noone has 5700XTs pulling 100W from the wall. What hiveOS seems to be showing for those cards is the TDP which was calculated above from the TGP.

In fact I like what XFX did by making the TDP = TGB as like this it will make the calculation more clearer and easier for the power limit.