My attempts to set up a dual GPU system for inference utilizing dual Nvidia cards, even with just compute, have failed.
Both cards (RTX 4070 Ti 16 GB + RTX 3060 12 GB) work fine individually, but in a dual-GPU setup `nvidia-smi` will only ever show one card at a time. The second card consistently fails to initialize with “RmInitAdapter failed.” Clean OS, multiple boards tested, same result. 3060 VBIOS dump looks suspiciously small (~157 KB vs expected ~512 KB+).
System specs:
- CPU: Intel Core i9-13900
- RAM: 128 GB DDR5
- GPUs: NVIDIA RTX 4070 Ti Super 16 GB (primary), NVIDIA RTX 3060 12 GB (secondary), also tested with K80 in the past
- Motherboards tested: MSI Z790 Carbon Wi-Fi, Gigabyte Z790 Aorus Elite AX
- Storage: Multiple NVMe SSDs, plus 1 TB SATA SSD pulled from an old laptop for clean OS test
- OS: Linux Mint 22 XFCE (Ubuntu 24.04 base), also tested Ubuntu/Mint fresh installs
- Kernel: 6.8.x generic
- Display: Samsung Q90 OLED 65"
- PSU: Adequate for dual-GPU setup
- BIOS: IGFX set as primary display, Above 4G Decoding enabled, ReBAR enabled, PCIe slots forced to Gen4, ASPM disabled
What I’ve tried:
- Fresh OS installs on a completely clean SSD that never had NVIDIA before
- NVIDIA driver installs (mostly version 550 from Ubuntu repos)
- CUDA toolkit installs (12.1 and 12.5 from NVIDIA repos)
- Python 3.10.14 via pyenv and deadsnakes PPA for compatibility with AI repos
- Verified nouveau blacklisting and modeset settings
- Tried slot configurations (x16 + x8, swapped cards between slots)
- Each card works fine **on its own**: if only the 3060 is present, `nvidia-smi` sees it; if only the 4070 Ti is present, `nvidia-smi` sees it
- Problem is specifically **dual-GPU binding**: driver never brings both up at once
PCI devices (from lspci):
01:00.0 VGA compatible controller [0300]: NVIDIA Corporation GA104 [GeForce RTX 4070 Ti] [10de:2484]
02:00.0 VGA compatible controller [0300]: NVIDIA Corporation GA106 [GeForce RTX 3060] [10de:2504]
Error logs (dmesg/journalctl):
nvidia: module verification failed: signature and/or required key missing - tainting kernel
nvidia-drm[GPU ID 0x0000:02:00.0]: Failed to allocate NvKmsKapiDevice
nvidia-drm[GPU ID 0x0000:02:00.0]: Failed to register device
NVRM: GPU 0000:02:00.0: RmInitAdapter failed!
VBIOS notes:
- Dumped 3060 VBIOS from /sys/bus/pci (Linux method)
- File size was only ~157 KB (normal would be ~512 KB–1 MB)
- Could suggest VBIOS corruption or incomplete ROM
- However, since the 3060 works fine as the only card in the system, the card is not dead — the issue appears only when both GPUs are installed
Summary:
- Both GPUs visible on PCIe bus
- Each works perfectly by itself
- Dual-GPU setup: only one card ever initializes; the other fails with `RmInitAdapter failed`
- Tested clean SSD, clean OS, multiple boards, drivers, toolkits, same result
- Strong suspicion of either a VBIOS/firmware conflict on the 3060 or a resource binding issue with dual-GPU configs. At this point after several months of on and off attempts any advice would help. Thanks in advance. -rio
What’s more is that I’ve only been using Linux for maybe 3 years and I never looked back but I’ve never reached out on a forum for help with the stuff before. I usually just figured it out by exhausting web search but I’m at my wit’s end with his. So before I buy a bigger card or a WS mobo and tossing in the towel, I figured I’d come here as the last resort and maybe get some really good advice. I appreciate you guys.