How to get an AMD Instinct MI60 running for AI workloads, including cooling solutions, BIOS settings, and fan control.

Introduction#

The AMD Instinct MI60 is a powerful server GPU featuring 32 GB of HBM2 VRAM and PCIe 3.0 connectivity. It remains a good choice for budget-conscious AI developers looking for high VRAM capacity at a fraction of modern GPU prices. With proper setup, the MI60 can handle local LLM inference, Whisper transcription, Stable Diffusion, and other workloads.

Parts List#

Item Details Notes
AMD Instinct MI60 GPU 32GB HBM2 VRAM Passive-cooled server GPU
92mm x 38mm Fan High static pressure, 12V, PWM Example: GDSTIME 92x38mm fan
Fan Mounting 3D printed shroud Single GPU STL
High Wattage PSU 600W+ with 2x 8-pin PCIe MI60 draws 300W under load

Tip: Look for “92mm x 38mm high static pressure fan” on Amazon or eBay. GDSTIME, Delta, and Sunon are common brands. Some models can be loud — use PWM or undervolting to reduce noise.

Photos#

MI60 with 92mm GDSTIME fan attached using a 3D-printed shroud. The housing also doubles as physical support for the GPU.

Dual MI60 setup using a shared dual-duct housing. STL available (fits one 92mm or one 120mm fan to cool both GPUs).

BIOS Settings#

To ensure the MI60 initializes properly:

Setting Value Reason
Above 4G Decoding Enabled Required for proper BAR mapping and ROCm
Resizable BAR Disabled ROCm doesn’t benefit, may break compatibility
PCIe Link Speed Auto or Gen 3 MI60 is PCIe 3.0
CSM Disabled Use UEFI boot mode
Integrated Graphics Disabled Frees resources, avoids conflicts
Primary GPU PCIe Slot / PEG Ensures MI60 initializes at boot

Tested platforms: X570, B550, TRX40.

Software Stack#

  • Operating System: Ubuntu 22.04 LTS or Linux Mint 21.x
  • Driver Stack: ROCm 5.6 (newer versions dropped MI60 support)

Verifying ROCm Installation#

Check if the GPU is detected:

rocminfo

Verify HIP support:

/opt/rocm/bin/hipInfo

Troubleshooting:

  • Kernel versions 5.15 or 6.2 are recommended
  • Ensure /dev/kfd exists and your user is in the video group
  • Stick to ROCm 5.6 for MI60 compatibility

Cooling and Fan Control#

Why This Matters#

The MI60 is a passive-cooled server GPU designed for datacenter airflow. Without proper cooling, junction temperatures can spike to 95-100°C during inference. Running GPUs at these temperatures dramatically shortens their lifespan.

The solution described here reduces junction temperatures from 96-97°C to 80°C - a 16-17°C improvement that significantly extends operational life.

Physical Setup#

  • Mount a 92mm x 38mm fan directly onto the MI60 heatsink using a 3D printed bracket
  • For dual setups, use the shared dual-duct housing
  • Connect the fan to a motherboard PWM header

Software Fan Control#

The GitHub repository includes two fan control options:

Data-Driven Controller (Recommended)#

Uses a preemptive, utilization-based approach. The key insight: by the time temperature spikes, it’s already too late. Instead, this controller uses a utilization→PWM curve learned from analyzing 300,000+ historical samples.

How it works:

  1. Preemptive Control: Fan speed based on GPU utilization, not temperature
  2. Learned Curve: PWM values derived from real operational data
  3. Smooth Interpolation: Linear interpolation prevents jarring speed changes
  4. Safety Backstop: Temperature still monitored as backup

Results:

  • Junction temperatures: 80°C max (down from 96-97°C)
  • Fan behavior: Smooth curves maxing at ~88%
  • Response time: Fan leads temperature changes

Simple Bash Script (Fallback)#

Basic utilization-based control without Python dependencies. Use as fallback if you can’t run Python.

Benchmarks#

Model Tokens/sec Temperature
qwen2.5-coder:32B 15.13 ~70°C
qwen2.5:8b 57.40 ~60°C
qwen3:8b 46.61 ~60°C
qwen3:30b 29.86 ~65°C
qwen3:32b 14.06 ~73°C

Known Issues#

  • ROCm 6.x+ drops MI60 support (stick with 5.6)
  • Docker containers require manual device mapping
  • MI60 lacks native FP16 acceleration; FP32 models perform better
  • Whisper.cpp falls back to CPU - GPU support may work with right OpenCL build
  • Stable Diffusion can cause GPU lockups in some configurations

Resources#