How to set up a mining rig with mixed NVIDIA and AMD GPUs?

Hardware Compatibility and Power Requirements

1. Confirm motherboard PCIe lane allocation supports concurrent operation of both NVIDIA and AMD GPUs without bandwidth throttling.

2. Verify PSU wattage meets combined TDP: a dual-GPU rig with one RTX 4090 (450W) and one RX 7900 XTX (355W) demands at least 1200W 80+ Gold certified unit with dual 12VHPWR and 8-pin PCIe connectors.

3. Ensure physical clearance allows for proper airflow between dissimilar coolers—especially critical when stacking open-frame AMD blowers beside axial-cooled NVIDIA cards.

4. Use PCIe riser cables rated for Gen 3 or higher; avoid passive USB-based risers as they introduce latency and instability during DAG epoch transitions.

5. Install GPU drivers separately: NVIDIA Game Ready 551.86 or Data Center 535.129.03, and AMD Adrenalin 24.5.1 with ROCm 6.2 enabled for compute workloads.

Software Stack Configuration

1. Choose miner software supporting heterogeneous backends: T-Rex Miner v0.26.5 natively handles NVIDIA CUDA and AMD OpenCL simultaneously via unified config.json.

2. Assign devices explicitly using device index syntax: “–devices 0,2” targets GPU 0 (NVIDIA) and GPU 2 (AMD), bypassing automatic enumeration conflicts.

3. Disable Windows Hardware-accelerated GPU scheduling in Graphics Settings to prevent driver-level arbitration failures during ethash or kawpow kernel execution.

4. Set GPU clocks manually: lock NVIDIA core at 1500 MHz and memory at 2200 MHz; constrain AMD GPU core to 1800 MHz and VRAM to 2400 MHz using AMDGPU-PRO tools.

5. Configure DAG generation path on NVMe SSD mounted at /dagstore to avoid SATA bottleneck during Ethereum Classic or Ravencoin DAG swaps.

Thermal and Stability Management

1. Mount GPUs vertically using PCIe vertical riser brackets to ensure laminar airflow across both die types, reducing thermal crosstalk by up to 18°C under sustained load.

2. Deploy separate fan curves: NVIDIA fans ramp from 30% at 55°C to 100% at 82°C; AMD fans initiate at 45°C and hit max speed at 78°C to accommodate higher junction temperatures.

3. Monitor per-GPU metrics using HWiNFO64 with custom sensor logging—track not only temperature but also PCIe link width negotiation and memory controller retries.

4. Apply undervolting profiles: reduce NVIDIA VDDC by 85 mV and AMD VDD GFX by 120 mV to lower power draw without hash loss, validated over 72-hour stress runs.

5. Isolate memory timings: use NVIDIA’s nvidia-smi -r to reset ECC errors independently, while running amdmemtweak --reset on AMD side before each mining session restart.

Pool Integration and Protocol Handling

1. Select pools supporting multi-algo submission: Ethermine and 2Miners accept simultaneous ETHW and ETC shares from same IP if miner identifiers are prefixed with vendor tags like “nvidia_” and “amd_”.

2. Route traffic through dedicated VLANs: assign NVIDIA GPUs to VLAN 10 (192.168.10.0/24) and AMD GPUs to VLAN 20 (192.168.20.0/24) to eliminate UDP packet collision during stratum handshakes.

3. Configure failover endpoints per vendor: primary pool for NVIDIA is us1.ethermine.org:4444; for AMD, use eu1.2miners.com:6060 with distinct worker names ending in “-amd”.

4. Disable auto-algo switching in miner binaries—explicitly define –algo ethash for NVIDIA and –algo kawpow for AMD to prevent firmware-level instruction set mismatches.

5. Log share rejection reasons separately: NVIDIA rejections often cite “stale share due to block propagation delay”, whereas AMD logs show “invalid nonce format” when OpenCL compiler version mismatches occur.

Firmware and BIOS Tuning

1. Enter UEFI and disable CSM (Compatibility Support Module) to enforce UEFI-only boot paths required by modern AMD GPU initialization routines.

2. Enable Above 4G Decoding and Resizable BAR in BIOS—critical for AMD RDNA2 and NVIDIA Ampere cards to access full VRAM address space during DAG loading.

3. Set PCIe slot configuration to Gen 4 x16 for primary NVIDIA slot and Gen 3 x8 for secondary AMD slot to balance bandwidth without triggering PCIe training failures.

4. Flash GPU BIOSes individually: use NVFlash 5.322 for NVIDIA and AMDVBFlash 2.3.4 for AMD, ensuring no cross-vendor signature corruption occurs during write cycles.

5. Lock PCIe ASPM (Active State Power Management) to L0s only—disabling L1 prevents intermittent link drops observed when mixed vendors negotiate power states asynchronously.

Frequently Asked Questions

Q1: Can I use the same DAG file for both NVIDIA and AMD GPUs?No. NVIDIA requires little-endian DAG binary layout optimized for CUDA warp execution, while AMD expects big-endian OpenCL-compatible structures. Separate DAG directories must be maintained.

Q2: Why does my AMD GPU report “CL_OUT_OF_RESOURCES” during mining initialization?This error arises when ROCm runtime fails to allocate pinned memory due to NVIDIA driver reserving contiguous DMA buffers. Reboot with NVIDIA driver unloaded, initialize AMD first, then load NVIDIA modules.

Q3: Is it safe to run mixed GPUs on a single Windows installation?Yes, provided you install drivers sequentially—AMD first, then NVIDIA—and disable Windows Update automatic driver replacement to avoid forced rollback to generic display drivers.

Q4: Do PCIe bifurcation settings affect mixed-GPU mining stability?Yes. Configuring x8x8x0x0 bifurcation on a x16 slot forces both GPUs into x8 mode, eliminating bandwidth asymmetry that causes AMD kernel timeouts during high-hashrate epochs.