Driver Incompatibility for Supporting Both Volta and Blackwell Simultaneously on Ubuntu

Accelerated Computing CUDA CUDA Setup and Installation
,
1

Hello NVIDIA team and community,

I’m currently facing a driver compatibility issue on Ubuntu when trying to support both Volta-based GPUs (Tesla V100) and the latest Blackwell GPUs (RTX 5090) in the same system(on Ubuntu LTS 24.04 and Supermicro H12ssl-i).

From my testing, V100s are supported by the proprietary NVIDIA drivers. However, my RTX5090 requires the open-source NVIDIA GPU kernel modules and is not supported by the proprietary drivers. This creates a dead end for my plan to combine V100’s FP64 jobs and RTX5090’s FP8 jobs in the same workstation.

Is there any release of a unified driver branch that supports both Volta and Blackwell architectures at the same time? If not, what is the recommended best practice for systems that require both GPUs to be operational simultaneously under Ubuntu/Linux? Will there be containerized or multi-driver solutions officially supported for such mixed setups?

Thank you for your time, and I look forward to any guidance or roadmap information.

Regards,

Zhiyuan

2

Perhaps this thread helps you out.

3

echo “=== BeccaLabs Dual GPU Status ===” && echo “” &&
echo “— 580 Stack (RTX 5080) —” && sudo cat /proc/driver/nvidia/gpus/0000:01:00.0/information && echo “” &&
echo “— 535 Stack (Tesla V100) —” && sudo cat /proc/driver/nvidia535/gpus/0000:11:00.0/information && echo “” &&
echo “— Device Nodes —” && ls -la /dev/nvidia[0-9] /dev/nvidia535-0 && echo “” &&
echo “— Kernel Modules —” && lsmod | grep nvidia
opened fd=3
=== BeccaLabs Dual GPU Status ===

-– 580 Stack (RTX 5080) —
Model: NVIDIA GeForce RTX 5080
IRQ: 140
GPU UUID: GPU-6378ff62-b578-e663-3dfc-1e95cb664f94
Video BIOS: 98.03.3b.c0.5a
Bus Type: PCIe
DMA Size: 52 bits
DMA Mask: 0xfffffffffffff
Bus Location: 0000:01:00.0
Device Minor: 0
GPU Firmware: 580.159.03
GPU Excluded: No

-– 535 Stack (Tesla V100) —
Model: Tesla V100-SXM2-16GB
IRQ: 149
GPU UUID: GPU-8cf39fff-f398-0757-ede3-165f2edbb382
Video BIOS: ??.??.??.??.??
Bus Type: PCIe
DMA Size: 47 bits
DMA Mask: 0x7fffffffffff
Bus Location: 0000:11:00.0
Device Minor: 0
GPU Excluded: No

-– Device Nodes —
crw-rw-rw- 1 root root 195, 0 May 21 17:40 /dev/nvidia0
crw-rw-rw- 1 root root 195, 1 May 21 17:40 /dev/nvidia1
crw-rw-rw- 1 root root 196, 0 May 21 17:41 /dev/nvidia535-0

-– Kernel Modules —
nvidia_uvm 2154496 0
nvidia_drm 139264 7
nvidia_modeset 1818624 7 nvidia_drm
nvidia 14413824 93 nvidia_uvm,nvidia_modeset
nvidia535 56803328 0
drm_ttm_helper 16384 3 amdgpu,nvidia_drm
video 77824 2 amdgpu,nvidia_modeset
(base) ileggal@pop-os:~$

4

Hi @dstover424,

I’m very interested in your May 21 post showing the simultaneous loading of NVIDIA 580 and 535 kernel modules for an RTX 5080 and Tesla V100.

I’m building a Linux workstation with an RTX 5060 Ti and two V100 SXM2 16GB GPUs connected via NVLink, and I’m investigating ways to use all three GPUs for LLM inference.

Your result is particularly interesting because it suggests a possible alternative to running separate VMs with PCI passthrough.

Could you clarify a few things?

  1. Were you able to create CUDA contexts and execute kernels on both GPUs, or does the current setup only achieve device initialization?

  2. How did you create the separate nvidia535 kernel module and /dev/nvidia535-0 device node? Did this require modifications beyond renaming the module?

  3. Have you published the driver patches or build instructions anywhere?

  4. If CUDA works on both GPUs, can they be used simultaneously from separate processes? Have you attempted to access both from a single CUDA process?

I also noticed the “BeccaLabs Dual GPU Status” heading in your output. Is there a public repository or technical documentation associated with that project?

Any information would be greatly appreciated. Your work could be very useful for people trying to combine Volta and Blackwell GPUs on Linux.

Thank you!