ESXi/vGPU: T4 and RTX PRO 6000 Blackwell in single host -> only 1 card recognized

[florian9000]

Hi there,

I have an ESXi host (VMware ESXi 8.0.3 build-25067014) with a T4 graphics card and the latest host drivers:

NVD-VMware_ESXi_8.0.0_Driver   580.105.06-1OEM.800.1.0.20613240      NVD     VMwareAccepted    2026-01-07    host
nvdgpumgmtdaemon               580.65.05-1OEM.700.1.0.15843807       NVD     VMwareAccepted    2026-01-07    host

Everything worked fine in terms of Shared Direct + Mixed Mode vGPU support during my initial tests.

I now added an additional RTX PRO 6000 Blackwell graphics card, but suddenly the T4 GPU is no longer detected by nvidia-smi:

nvidia-smi -L
GPU 0: NVIDIA RTX PRO 6000 Blackwell Server Edition (UUID: ...)

However, the host still seems to recognize the PCIe device:

lspci | grep 3D
0000:43:00.0 3D controller: NVIDIA Corporation Tesla T4
0000:9b:00.0 3D controller: NVIDIA Corporation NVIDIA RTX Pro 6000 Blackwell DC
0000:9b:00.2 3D controller: NVIDIA Corporation NVIDIA RTX Pro 6000 Blackwell DC
0000:9b:00.3 3D controller: NVIDIA Corporation NVIDIA RTX Pro 6000 Blackwell DC
0000:9b:00.4 3D controller: NVIDIA Corporation NVIDIA RTX Pro 6000 Blackwell DC
^... this repeats a ton of times; not sure if this is correct; not the case for the T4

According to the vGPU software, both the T4 and the RTX PRO 6000 Blackwell is supported.

Any ideas?

[sschaber]

You simply cannot mix these GPUs in the same host. They are using different technologies in terms of resource management.

[florian9000]

Thanks for confirming!

I was already expecting something like this since the RTX 6000 uses SRV-IO based vGPU mode while the T4 uses non SRV-IO.

[sschaber]

Main reason is the switch from software RM to the RISC processor that is now built on the newer generation GPUs and does the RM in hardware.
You can either run RM in software or RM in hardware so only GPU generations with the same “technology” can work in the same host

[florian9000]

Thanks for clarifying. One more question: What is the earliest architecture that also uses hardware RM?

[sschaber]

ADA and Hopper GPUs