Help running a Tesla P40

Emaa · December 19, 2025

Tesla P40 Performance Ceiling on HP Z620 Workstation

Hardware Configuration:

GPU (Compute): NVIDIA Tesla P40 24GB (Running Headless).
GPU (Display): NVIDIA Quadro P400 (Handling all display output).
Workstation: HP Z620 (Xeon E5-2620 v2 CPUs, PCIe 3.0 supported).
Memory: 128GB DDR3 ECC RAM (Functional and stable).
BIOS/Firmware: Tesla P40 flashed to a non-OEM, WDDM-compatible BIOS.
Drivers: Legacy NVIDIA drivers installed to enable WDDM mode support for the P40 (v. 30.0.15.1179)

Power Topology:

The GPU is powered via the official HP Z620 PSU using the following chain:

Primary Source: Dual HP PSU native PCIe 6-pin cables.
Adapters: Official 6-pin to 8-pin adapters → GPU Y-aggregator → Single 8-pin Tesla input.
Note: All workstation hardware (PSU, Board, Chassis) is original OEM equipment except for the RAM that I upgraded from 16 to 128.

Observed Technical Issues:

Despite the 250W power limit being correctly reported in software, the card remains locked in a low-performance state:

PCIe Link Speed: Hard-locked at Gen 1 (2.5GT/s). HWiNFO64 reports the link as PCIe v3.0 x16 (8.0 GT/s) @ x16 (2.5 GT/s)(HWinfo pic at the bottom).
Power Draw: Capped at a hard ceiling of 43W to 50W under full CUDA load (Blender/Octane).
Clocks: Memory clock remains stuck at the base strap 202MHz (in TCC mode) / 405MHz (in WDDM mode) and does not boost under load.
Throttle Reasons: nvidia-smi and diagnostic tools report "None" or "Not Active," yet the hardware refuse to scale.

Steps Taken & Results:

BIOS: Forced Gen 3 PCIe in the Z620 BIOS. → No change.
Registry: Forced MaxLinkSpeed to 3 (Gen 3) in the Windows Registry Class key. → Driver requests Gen 3, but link stays at Gen 1.**
Driver Mode: Switched between TCC and WDDM modes. → Behavior is identical* in both modes.
Physical Maintenance: Card reseated, PCIe interface cleaned with IPA, and air-dusted. → No change.
ECC Setting: Disabled ECC via Nvidia Control Panel. This resulted in a 5-minute reduction in render times and a slight increase in power draw (staying under 50W), confirming a reduction in memory latency overhead.

Being a datacenter card it doesn't have natively active cooling, so I strapped a 3D printed housing on the back of the card with a blower fan connected to a chassis fan.

Thermal's are not the issue here.

I need this card to be able to run in WDDM as well so if any of you know what's going on or have a rough idea of what I'm supposed to do it'll be super helpful.

* I already played around in the Registry to make the P40 work in WDDM and to make it appear as a performance rendering card.

**As previously mentioned the memory clocks varies as well as the performance (slightly better in TCC), but power draw, link speed and temps stay the same.