GPU Tuning - Wyatt's Notes

NVIDIA GPUs are organized into Streaming Multiprocessors (SMs), each containing a set of CUDA cores, Shared memory, register files, and scheduling units. The number of SMs and their configuration Defines the GPU”s compute capability.

AMD uses a similar architecture with Compute Units (CUs), each containing multiple Stream Processors (SP). RDNA 3 CUs contain 2 SIMD32 units (64 SPs per CU).

Memory bandwidth is often the bottleneck in GPU workloads. The theoretical bandwidth is:

The factor of 2 accounts for double data rate (GDDR6/GDDR6X) or multi-level signaling (HBM).

• ROPs (Render Output Units): Handle pixel output operations (blending, depth testing, anti-aliasing). Important for high-resolution gaming and rendering.
• TMUs (Texture Mapping Units): Handle texture sampling, filtering, and address calculations. Important for workloads with heavy texture use.

These are fixed-function units that cannot be overclocked independently. They scale with the GPU’s Base and boost clocks.

NVIDIA’s GPU Boost (version 3.0 and later on Maxwell+) is an autonomous frequency scaling algorithm That continuously adjusts the GPU clock speed based on:

1. Power consumption. Current draw vs. The configured power limit
2. Temperature. Current temperature vs. The thermal throttling threshold
3. Voltage. Current voltage vs. The maximum allowed voltage
4. Software limits. Any application-specific clock limits

The GPU operates at the highest frequency that satisfies all constraints. If temperature rises, the GPU reduces frequency to stay within the thermal limit. If power headroom exists, the GPU increases Frequency.

• Base Clock: The minimum guaranteed clock speed under typical gaming loads at the default power limit. This is the “worst case” frequency.
• Boost Clock: The maximum frequency the GPU can achieve under ideal conditions (adequate cooling, sufficient power, and a workload that hits the right utilization pattern).

In practice, most GPUs boost above the listed boost clock because the boost specification is based On a specific temperature and power envelope. If your cooling is better than the reference design, The GPU will boost higher.

NVIDIA GPUs operate along a voltage/frequency (V/F) curve. Each frequency point has a minimum Voltage required for stability. The GPU Boost algorithm selects the highest frequency point where The current conditions (temperature, power, voltage) allow operation.

The V/F curve is non-linear — higher frequencies require disproportionately more voltage. This is Because:

A small frequency increase at the top of the curve requires a larger voltage increase, which causes A quadratic increase in power consumption. This is the fundamental reason why undervolting works: You sacrifice a small amount of peak frequency for a large reduction in power consumption, which Allows the GPU to sustain higher average frequencies under thermal constraints.

AMD’s equivalent to GPU Boost is PowerPlay, which manages GPU frequency and voltage based on thermal And power constraints. The principles are similar but the implementation differs:

• AMD GPUs use a “power limit” rather than a hard frequency/voltage curve. The GPU boosts as high as possible within the power budget.
• The power limit is configurable via AMD Adrenalin or MSI Afterburner, up to +15–20% above the default TGP (Total Graphics Power).
• AMD’s automatic undervolting feature (in Adrenalin) is a simpler interface than NVIDIA’s curve editor but offers less fine-grained control.

The most effective GPU tuning technique is undervolting — reducing the voltage at which the GPU Operates while maintaining or only slightly reducing the clock frequency.

graph TD
    A[Open MSI Afterburner] --> B[Press Ctrl+F to open curve editor]
    B --> C[Identify target clock speed on V/F curve]
    C --> D[Hold Shift and click desired frequency point]
    D --> E[Drag the point down to lower voltage]
    E --> F[Apply Ctrl+F to flatten curve above target]
    F --> G[Test with 3DMark Time Spy or Superposition]
    G --> H{Stable?}
    H -->|Yes| I[Try lowering voltage another 10-25 mV]
    H -->|No| J[Raise voltage by one step]
    I --> G
    J --> K[Run 30-min FurMark to verify thermals]
    K --> L[Save profile]

1. Open MSI Afterburner and press Ctrl+F to open the voltage/frequency curve editor.
2. Observe the stock curve. Note the frequency at stock voltage (e.g., 1905 MHz at 1050 mV for an RTX 3070).
3. Hold Shift and click on the point corresponding to your target frequency (e.g., 1800 MHz).
4. Drag this point down to find the minimum stable voltage (e.g., 875 mV).
5. Hold Ctrl+F again and drag all points above your target down to the same voltage, creating a flat line. This prevents the GPU from ever exceeding your target voltage.
6. Click Apply (the checkmark button in MSI Afterburner).
7. Run a benchmark (3DMark Time Spy, Unigine Superposition).
8. If stable, try reducing voltage by another 10–25 mV.
9. When instability appears (artifacts, crashes, driver resets), raise voltage by one step.
10. Run a 30-minute FurMark session to verify thermals and sustained stability.

AMD GPUs can be undervolted using MSI Afterburner (same curve editor approach) or AMD Adrenalin:

Adrenalin Method:

1. Open AMD Adrenalin → Performance → Tuning.
2. Select “Manual Tuning” or “Automatic Undervolting.”
3. For automatic undervolting, move the slider to reduce the target voltage. Adrenalin will find a stable point.
4. For manual tuning, adjust the V/F curve similarly to the NVIDIA method.

MSI Afterburner Method:

Same process as NVIDIA. Hold Ctrl+F, select target frequency, drag voltage down. AMD GPUs Respond well to undervolting, with 50–80 mV reductions being achievable on most cards.

GPU memory (GDDR6, GDDR6X, or HBM) can be overclocked by increasing the memory clock frequency. This Increases memory bandwidth, which benefits workloads that are memory-bandwidth bound:

• High-resolution gaming (4K with high texture quality)
• GPU computing (machine learning, rendering)
• Cryptocurrency mining

1. In MSI Afterburner, increase the Memory Clock slider in increments of +100 MHz.
2. After each increment, run a memory-intensive benchmark:

• 3DMark Time Spy (for gaming workloads)
• AIDA64 GPGPU benchmark (for compute workloads)

1. Watch for visual artifacts. Colored squares, flashing textures, or screen corruption.
2. When artifacts appear, reduce the offset by 100 MHz.
3. Run a 30-minute stability test at the final setting.

• Cooling - How GPU thermal design and cooling solutions affect boost clock behavior
• Undervolting and Overclocking - How GPU undervolting reduces power consumption while maintaining clock speeds
• PSU and BIOS Guide - How PSU wattage and PCIe power delivery affect GPU stability