Dell Precision 7670 & Dell Precision 7770 owner's thread

[Aaron44126]

1 minute ago, MyPC8MyBrain said:

Thank you Operator,
we learned the past few days that some of us received units with 125 some with 150 and some with 157 TDP,
one member notified me in private that he received his 7670 with 3080Ti at 157 TDP, 
the search is after the higher TDP vbios variants,
we should have a repo for these once located,

 

I am curious about this, it seems pretty odd.  I am wondering if maybe just an early/testing vBIOS was included on their card by mistake.

 

Did this user give the version number for the vBIOS in their system that was reporting 157W TDP?  It should be visible on the very same screen in NVIDIA control panel where you can see the card's max TDP.  NVIDIA is good about never reusing the same version number and kicking out a new one for each vBIOS variation that they produce.  (Just another thing that could maybe be used to quickly tell which vBIOS you have, or maybe even tease out which one is "newer" than the other.)

[MyPC8MyBrain]

9 minutes ago, Aaron44126 said:

Did this user give the version number for the vBIOS in their system

i have requested for his version and if possible a copy,
he has not yet visited back to read my respond and reply,

i also raised a technical ticket with Dell support requesting unrestricted vbios for the 3080,
the agent was very nice and cool about it, she was unable to locate a higher TDP vbios and also requested the version number so she can match to his version, but i am still waiting for his response to fill in the gaps,
 

[samatsu]

38 minutes ago, Aaron44126 said:

Yes, I got Delta fans,  I want to replace it. I call technical support, They ask me what is part number.🤣

 

Classic Precision issue.  Try going to BIOS setup and disabling C-states.  It should be under the "performance" section.

 

I just noticed that there is a setting there about C-states and discrete graphics.  I wonder if that "feature" is involved with the dGPU "stuttering" issue that I described above...  I will experiment with it.

 

 

Unfortunately, Dell doesn't have separate part numbers for variations of the same heatsink assembly.  (You'll get different part numbers for i.e. dGPU present or no dGPU present, but not for which type of fans are in there.)

 

Did you get a 7670 with Delta fans?

 

[Aaron44126]

15 minutes ago, samatsu said:

Yes, I got Delta fans,  I want to replace it. I call technical support, They ask me what is part number.🤣

 

You just want to have them send the "same part" you already have...  So far, I think that everyone who has tried has ended up receiving the Sunon part.

 

They should be able to figure out the part number (they shouldn't even be allowed to send you a part that isn't already in your system), but if they can't, you can find it yourself.  Go to dell.com/support, put in your service tag number, and then click "View product specs" on the next page and it will take you to the full parts list.  Look under the GPU section for one labeled "ASSY,HTSNK".

[operator]

39 minutes ago, MyPC8MyBrain said:

some of us received 3080 units with 125 some with 150 and some with 157 TDP

 

Except the one person with 157 TDP, who did report a 7670 with 150 TDP? Wouldn't that be just fine? I doubt this 7 W extra would be a huge factor..?

[MyPC8MyBrain]

1 hour ago, operator said:

who did report a 7670 with 150 TDP?

@Aaron44126 reported his 7770 with 150 TDP, he also posted his vbios few posts back,

unfortunately flash tool is having some issues cross flashing to this new 10DE:2420 chip variant,
150 TDP is better than 125, even 1w is a performance increase,
native turbo for these new 3080 Ti 16gb variant is 167w,
i see no reason why we should be limited to 125,
(regardless if we can fully utilize in this chassis)

 

some may perceive the following as nitpicking, i don't...

on 3DMARK our 125 TDP is scoring 50% lower than average same configuration out in the wild,
i tweaked and tuned everything i can but the gpu is scoring 9k max while others average 13.5k,
i was able to break top cpu score on 3DMARK for the 12950hx (top cpu score is 14.5k), but the gpu score drags the unit to 50% below average, cpu in these units now days is secondary to gpu importance, leaving over 50% gpu performance on the table is just wrong for precision class workstation,

[operator]

Yes I know. I did follow this thread. 😄

Thought you were talking about sb. with a 7670 reported 150W TDP.

 

I scored 9818 out of the box in TimeSpy in a 18°C (64°F) room.

But during that the GPU power only peaked at 102W.

 

I just upgraded 3DMARK to the latest version and during that the GPU Power in HWiNFO showed 751W peak on my GPU 🙈

[MyPC8MyBrain]

out of curiosity i had a feeling something is off with nvflash and sure enough there is,
i figured let me try to re-flash my own vbios, guess what...
got a mismatch error with my own vbios i dumped and tried to flash back,

 

run nvflash tool again with -6 switch, mismatch again only this time it asked me twice to override something, 
flash did complete this time around, so something in the tool needs updating to bypass NVidia new shenanigans,

 

 

[Aaron44126]

Your image has a mismatched subsystem ID.  This is pretty normal in laptops.  The subsystem ID is essentially assigned by the motherboard, not by the GPU.  This is why you need to mod the INF to get the NVIDIA driver to load if you, for example, take a MXM GPU card from a newer system and drop it into an older system.  Its subsystem ID actually changes depending on the system that it is in.

 

The value "burned in" to the BIOS might not match the value that is observed while the card is up and running.

 

We can even pick out the values from the nvflash output.

10DE = NVIDIA's vendor ID

2420 = NVIDIA's identifier for GeForce RTX 3080 Ti (laptop version)

0000 = A filler value...

1028 = Dell's vendor ID

0B2A = Dell's identifier for Precision 7670.

(You see these same values if you poke around NVIDIA's driver INF files.)

 

Anyway.  Comparing your image and my image, they mismatch on the subsystem ID and also the board ID.  You can override the subsystem ID mismatch (-6) but not the board ID mismatch.  nvflash has an option to override the board ID mismatch (-5), but, NVIDIA stopped allowing it with Turing cards.  So, you need a patched version of nvflash that ignores this mismatch in order to complete the cross-flash.

 

The problem is that the 3080Ti is "too new" and a patched version of nvflash that supports it has yet to surface.  I think that it will inevitably appear.  Maybe in just a few weeks/months when people want to start cross-flashing GeForce 4000 (desktop) cards and someone is motivated to mod a newer version of nvflash.

 

In the "old days" you could modify the vBIOS file itself and make all sorts of things work.  You could just change the IDs in the vBIOS file to match what you needed (as well as arbitrarily set the clock speeds, power limits, etc.).  My own flashing experience comes from messing with overclocking the old Quadro K5000M.  Anyone could produce vBIOS images to make it run faster than stock.  Since Pascal, NVIDIA GPUs are enforcing digital signatures on the vBIOS, so only NVIDIA can produce valid vBIOS images.

[1610ftw]

3 hours ago, MyPC8MyBrain said:

@Aaron44126 reported his 7770 with 150 TDP, he also posted his vbios few posts back,

unfortunately flash tool is having some issues cross flashing to this new 10DE:2420 chip variant,
150 TDP is better than 125, even 1w is a performance increase,
native turbo for these new 3080 Ti 16gb variant is 167w,
i see no reason why we should be limited to 125,
(regardless if we can fully utilize in this chassis)

 

some may perceive the following as nitpicking, i don't...

on 3DMARK our 125 TDP is scoring 50% lower than average same configuration out in the wild,
i tweaked and tuned everything i can but the gpu is scoring 9k max while others average 13.5k,
i was able to break top cpu score on 3DMARK for the 12950hx (top cpu score is 14.5k), but the gpu score drags the unit to 50% below average, cpu in these units now days is secondary to gpu importance, leaving over 50% gpu performance on the table is just wrong for precision class workstation,

 

Does that mean you are getting 125 + 25 vs 150 + 25 or is that 125 including turbo? Judging by your timespy benchmark it looks like 125 including turbo and I agree that it should be better, at least 12.5K should be possible even with a 150W TGP.

[Aaron44126]

15 minutes ago, 1610ftw said:

Does that mean you are getting 125 + 25 vs 150 + 25 or is that 125 including turbo? Judging by your timespy benchmark it looks like 125 including turbo and I agree that it should be better, at least 12.5K should be possible even with a 150W TGP.

 

HWiNFO reports a consistent 130W power draw despite NVIDIA control panel showing 150W TDP.  I believe 150W would include the boost TDP, but the boost TDP doesn't seem to kick in.

 

(This was also reported with last year's Precision 7760, 115W "actual" vs. 140W "max" TDP.)

[MyPC8MyBrain]

52 minutes ago, 1610ftw said:

is that 125 including turbo?

indeed, 

 

52 minutes ago, 1610ftw said:

it should be better, at least 12.5K

i under/overclock gpu, added 200 Mhz, but no matter what i did gpu wouldn't improve or get near average,
as if it was already pushing itself to its max, which i believe it is within its TDP limits atm,
these 42 off limit watts equal to 50%-60% inaccessible gpu performance,
which we all paid for and i think we all like to have available at our disposal,

 

37 minutes ago, Aaron44126 said:

but the boost TDP doesn't seem to kick in

it is possible that your cpu is overheating preventing dgpu from ramping up over 130W?
in your specific case you have all 4 m.2 populated with 128 CANN module,
maybe you are on the edge of power?
(in our private i forgot to include the monitor in the wattage count 😮)

 

I've come up with a theory, Dell might be segmenting gpu TDP based on the monitor of choice,
as if trying to balance power consumption to peripherals (in my case 4k oled touch),

all the while ignoring the basic scenario where end user do not use their screen as main output,

in that case the screen wattage math is out the window yet we still cannot regain gpu's full potential,

[Aaron44126]

29 minutes ago, MyPC8MyBrain said:

it is possible that your cpu is overheating preventing dgpu from ramping up over 130W?
in your specific case you have all 4 m.2 populated with 128 CANN module,
maybe you are on the edge of power?

I rather doubt it. The power draw cap is so specifically right at 130W, it has to be an intentional limit. Unlike the CPU, the dGPU has no problem staying cool. (It’s designed to be used at over 2x the power level in a desktop, after all…)

[brunooo84]

does the "new" heatsink (not delta) really makes a difference? How did you guys got dell to ship a new one? Just asked for it or did you say it doesn't work?

[win32asmguy]

1 hour ago, MyPC8MyBrain said:

it is possible that your cpu is overheating preventing dgpu from ramping up over 130W?
in your specific case you have all 4 m.2 populated with 128 CANN module,
maybe you are on the edge of power?
(in our private i forgot to include the monitor in the wattage count 😮)

 

I've come up with a theory, Dell might be segmenting gpu TDP based on the monitor of choice,
as if trying to balance power consumption to peripherals (in my case 4k oled touch),

all the while ignoring the basic scenario where end user do not use their screen as main output,

in that case the screen wattage math is out the window yet we still cannot regain gpu's full potential,

 

I do not think its related to installed components. Mine has FHD display, 32GB CAMM and a single SSD. Its 3080Ti is limited to 120W regardless of driver installed, CPU power consumption, or GPU temperature. If a fix is found it would be great to test it out.

[Aaron44126]

25 minutes ago, brunooo84 said:

does the "new" heatsink (not delta) really makes a difference?

 

See @win32asmguy's posts on page 5 to see what steps he went through to improve performance.  Then this one on page 9 where he discusses the new heatsink.  Basically, it can produce similar results without having to go through the hassle of repasting / repadding / etc.

[Aaron44126]

I was poking around in the BIOS to mess with C-states and I noticed another option in the performance section labeled "Intel maximum turbo boost 3.0" or something like that.  It says that it will enable the CPU to boost beyond the max turbo speed, but also disable NVIDIA Dynamic Boost 2.0.  It is disabled by default.  Has anyone messed with that ...?

[operator]

I played a bit around and as my best out of the box score was 17671 in CB23, I mostly got around 15k in all other runs. Today I was able to see the 17k again. But between all the benchmarks from 14k to 17k, nothing changed on my configuration side. So very odd to get so different values on the same setup.

 

Now I configured the AC/DC loadline fix and re-run the test. I had 18711 which is a 6% gain. I kept redoing the tests several times. I was able to achieve 21188 after some runs which is a 20% gain. But I had also all numbers in between (18k, 19k, 20k) as results and I changed nothing. Seeing CPU power peaked sometimes at 154W and sometimes at only 130W....

 

So I wonder why the system behaves so different.

[operator]

4 minutes ago, Aaron44126 said:

I was poking around in the BIOS to mess with C-states and I noticed another option in the performance section labeled "Intel maximum turbo boost 3.0" or something like that.  It says that it will enable the CPU to boost beyond the max turbo speed, but also disable NVIDIA Dynamic Boost 2.0.  It is disabled by default.  Has anyone messed with that ...?

 

Now you mentioned it - I had enabled this before my tests. But did not see any increase in turboboost. During benchmark it goes up to same values as before (4.7GHz) and stays then around 3.x GHz during the benchmark.

[Aaron44126]

8 minutes ago, operator said:

Now I configured the AC/DC loadline fix and re-run the test. I had 18711 which is a 6% gain. I kept redoing the tests several times. I was able to achieve 21188 after some runs which is a 20% gain.

 

No other tuning?  Man, why is 7670 faster than 7770...  😕

(15K out of box, 18K after repaste + AC/DC fix -- I know there is other stuff I could still do)

[operator]

Just now, Aaron44126 said:

 

No other tuning?  Man, why is 7670 faster than 7770...  😕

 

It is out of the box and only the loadline fix.

And of course it was only the single take run and not the 10 minutes run in CB23.

 

[Aaron44126]

Just now, operator said:

And of course it was only the single take run and not the 10 minutes run in CB23.

 

Oh nevermind then, I was thinking 10 minutes.  I've also gotten some higher single runs.

[operator]

25 minutes ago, Aaron44126 said:

 

Oh nevermind then, I was thinking 10 minutes.  I've also gotten some higher single runs.

 

10 minutes run in CB23 was 17154 which is a 13% gain from my out-of-the-box run.

CPU Load went up to 154W for short time and then it was stuck at 75W and 70-80°C (158-176°F).

 

3DMARK did only increase slightly. GPU was maxxed at 102W.

[MyPC8MyBrain]

here we go gents... regain ≥ 10% performance by enabling resizable bar.
3DMARK first test with these settings activated, no under or over volt (still with 125 TDP),
(my cpu is clearly not happy atm not being undervolted and is basically capped throughout the run)
 

setup_var CpuSetup 0x42B 0x1 # Enable PCIE Resizable BAR Support (default 0x0 = off)

 

 

 

[MyPC8MyBrain]

just heard back from Dell's engineer responding to my TDP support ticket,
he informed me that this specific gpu actually has a 175W TDP,
more to come...