Nvidia Tesla P4 on Lenovo M920q performance question

[axotopia]

I am a bit stumped on performance difference of a NVidia Tesla P4 on Lenovo M920q outperforming the P4 on my regular desktops.

I am putting a SFF Scanning Station for my large format scanner, using Topaz PhotoAI to do some on-the-spot image processing. Installed the Tesla P4 and noticed it is performing between the two RTX2080Super and GTX1080Ti installed on my desktops.

Now, compared to the Lenovo M920 (i9-9900ES), placing the same P4 into regular desktops ... image processing takes 3-3.5X longer on an i9-9900K and i9-13900K systems.

Using the i9-9900K desktop system as reference, both are on the same revision of Windows 11, iGPU driver, Nvidia Grid driver, WDDM driver. BIOS setting to no CSM, ASPM enabled, Above 4G enabled to match the 3 different systems. Just could figure why the lower performance on a higher spec desktop systems.

The BIOS is pretty basic on the M920q, is there something I am missing that is crippling the Tesla P4 on the desktop systems? Thanks.

 

[joeribl]

Cooling?

 

[piranha32]

The BIOS is pretty basic on the M920q, is there something I am missing that is crippling the Tesla P4 on the desktop systems? Thanks.

Did you look at the stats shown by nvidia-smi? Clock speeds, temperature, GPU utilization, and power draw should give you a good picture of how the card is performing, and maybe some ideas on what can be going wrong.

 

[axotopia]

Nvidia-smi stats are the same for clocks, utilization is similar , and cooling is actually better on the desktop vs the Lenovo. Really stumped... suspect it may be something with BIOS implement or something with how Windows handles the M920q.
The M920q actually is running on PCIe 16x on 8x electrical, instead of full 16x on desktops. Installing the P4 in a 8x PCIe slots on the desktops still didn't help.
Been scratching my head for over a week now, so need professional advise.

 

[CyklonDX]

care to supply some background?

Like OS, driver, full hardware configuration of each box, and nvidia-smi screencap / copy when your TopazAI is running, and at idle?
Also is your soft running on correct GPU? Could you also take a look at hwinfo, under load and at idle?

Note:
If your desktop has normal nv gpu's its likely its not running proper tesla driver.

 

[axotopia]

care to supply some background?

Like OS, driver, full hardware configuration of each box, and nvidia-smi screencap / copy when your TopazAI is running, and at idle?
Also is your soft running on correct GPU? Could you also take a look at hwinfo, under load and at idle?

Note:
If your desktop has normal nv gpu's its likely its not running proper tesla driver.

All 3 systems are on windows 11. Using the same Mvidia Grid driver 528.89, same WDDM 3.1 in windows.
Tried to narrow down to the 2 most similar systems both using i9-9900 K and ES. Nvidia-smi is reporting almost exactly the same stats with the exception of memory differerences.
Tried to narrow the BIOS setting to as similar as possible. The desktop being a gigabyte aorus elite z390, so a lot more settings vs the Lenovo M920q.
Ran GPUz to make sure things are similar.
So, even with everything i can observe looking almost exaxtly the same, including drivers.... the lower spec M920q is still running the tesla 3x faster on Topaz PhotoAI.

 

[piranha32]

All 3 systems are on windows 11. Using the same Mvidia Grid driver 528.89, same WDDM 3.1 in windows.
Tried to narrow down to the 2 most similar systems both using i9-9900 K and ES. Nvidia-smi is reporting almost exactly the same stats with the exception of memory differerences.
Tried to narrow the BIOS setting to as similar as possible. The desktop being a gigabyte aorus elite z390, so a lot more settings vs the Lenovo M920q.
Ran GPUz to make sure things are similar.
So, even with everything i can observe looking almost exaxtly the same, including drivers.... the lower spec M920q is still running the tesla 3x faster on Topaz PhotoAI.

Is it possible that the desktop is running low on memory and the OS is using the swap file? Or data I/O operations are slowing down the processing? How does GPU usage look like while you run the application?

 

[axotopia]

Ok. More update info testing out a new Nvidia Tesla P40 on the Z390 mobo. Results is almost the same as the GTX1080Ti on Topaz PhotoAI. The P40 performance is similar to the 1080Ti, but surprise here is it is still 3x slower than the Tesla P4 on the Lenovo M920q.
Still same drivers and OS.
Is there something in the P4 on M920q that is making it run way above spec? P4 is roughly 30-50% lower spec than the P40. Makes no sense.
I cannot run the P40 on the M920q unfortunately, Lenovo says not enough resource to use the P40.

 

[CyklonDX]

I recommend you DDU nv drivers on the desktop z390 and only install proper P40 drivers not the grid ones.
(i.e. Data Center Driver for Windows | 528.89 | Windows 10 64-bit, Windows 11 | NVIDIA )

You should also try to disable windows grabbing drivers on its own.
I presume its still using non-tesla libs/dlls from 1080 you had there before (a left over or something)

 

[axotopia]

Data Center Driver locks the Teslas in TCC mode, no WDDM The Tesla Grid drivers allows both TCC and WDDM., its Tesla and quadro specific driver not compatible with GTX and gaming RTX. Already uninstalled all 1080 components prior to installing the grid driver. No idea if Windows still keep the old library even if all nvidia components was uninstalled via the Windows unistall process.

System has 32gb and system is using 24gb during the processing, so no disk swapping factivities from what i can see on Task Manager.

The P40 is actually running to spec, about the same speed as the 1080ti, since it is physically almost identical board as the 1080ti.

Guess my confusion is that the P4 is running significant faster than its supposed to.... like roughly 4.5x faster than it technically supposed to do compared to the P40 tested performance. It is around +30% lower in spec compared to the P40, so should not be doing image processing 3-3.5x faster than the P40.

Just wondering what is going on with the P4 on M920q that is unleashing the extra processing capability on Topaz Photo AI.

I may do a clean install of Windows this weekend to see if its a driver issue.

 

[CyklonDX]

unless you do DDU, windows keeps everything.

 

[axotopia]

Done both DDU on existing OS, and a completely fresh Windows 11 install. Results still the same for P40. However, the 1080Ti saw a 15% but still 1/3 the speed of the P4 on the M920q. improvement.

Decided to put the P4 into the fresh Windows 11 install on the Z390 desktop.... it performed about like it should against the P40... 60% slower. 5X slower than when it is in the Lenovo M920q.

Placing the fresh Windows 11 SSD into the Lenovo M920q was not really successful as it ran unstable probably due to the chipset difference. For the brief moment I ran the benchmark, the P4 was slow like the Z390 desktop. Not sure I need to test a new fresh install on the M920q since just transferring the SSD was unstable with constant blank screens.

Not planning go any further at this point since I am suspecting it could be the copy of Windows 11 (Upgraded from 2019) that may not be following Nvidia's parameters and unlocking some crazy speed on the Tesla P4 (... or could be M920q hardware+BIOS).

Contacted Topaz Labs and they are looking into this abnormally.

 

[CyklonDX]

cool, post dets once you find something. I don't think its necessarily faster - its just your other setup is likely very slow.

Worth noting that P4 is bit different than P40;
Even tho they are from same Pascal generation P4 is more akin to P100 on pcie while P40 is more akin to Titan from Pascal series that is uncut. (but it depends on how topaz is utilizing their code compute pipe.) ~ P40 should be 1.8x P4 in everything.