NVIDIA GPU Performance Limiters assist end-users in identifying the hardware-related constraints of GPU performance.
They are part of the GPU Boost 1.0, 2.0, 3.0, and 4.0 technology and I extensively cover using the GPU performance limiters in my RTX 3050 SkatterBencher guide.
Overview
Like any other semiconductor vendor, NVIDIA employs many tools to govern the health and performance of the GPU at runtime. That includes tools to capture metrics and prevent damage due to out-of-spec behavior. NVIDIA is pretty secretive about the performance limiters. However, they provide some performance limiter debug information via the NVAPI.
NVAPI is NVIDIA’s core software development kit that allows direct access to NVIDIA GPUs and drivers on all Microsoft Windows platforms. Applications such as EVGA Precision X1 or HWiNFO rely on this API to provide us, enthusiasts, with GPU performance information. Let’s open HWiNFO and look at the various Performance Limiters available.
- Power: The Power performance limiter is flagged if the graphics card is using the maximum allowed power consumption
- Input power: power that is pulled from the PSU
- Output power: power that is consumed by the GPU and memory
- Thermal: The Thermal performance limiter is flagged if the graphics card is hitting the maximum allowed temperature
- Reliability Voltage: The Reliability Voltage performance limiter is flagged if the GPU requests the Vrel voltage to the voltage controller.
- Max Operating Voltage: The Maximum Operating Voltage performance limiter is flagged if the graphics card allows GPU voltage beyond the reliability voltage and the GPU requests the Vmax voltage to the voltage controller.
- Utilization: The Utilization performance limiter is flagged when all GPU compute resources are currently in use.
- SLI GPUBoost Sync: The SLI GPUBoost Sync performance limiter is flagged when the GPU Boost frequency is limited by multi-GPU SLI frequency synchronization.
Note that many more undocumented performance limiters may impact the graphics card’s performance. We will get to that topic in due time. For now, suffice it to say that when we try to get the maximum performance, our goal is to have the maximum operating frequency when only the Utilization performance limiter is flagged.
Performance Limiter: Power
The power performance limiter can be triggered by multiple power measurements on the graphics card. I illustrate this in my RTX 3050 overclocking guide where we tried shunt-modding to eliminate the power performance limiter. However, the power performance limiter persists.
We can rely on the NVAPI and HWiNFO to investigate what’s happening. Let’s have a look at the GPU Power reporting. I won’t go over the details of each line item, but let’s compare all rails with NVVDD during the Furmark Stress Test.
As you can see from the comparison, shunt-modding (OC Strategy #2) effectively lowered the reported input power. However, it did not adjust the two output power metrics. So, we can theorize that the output power metrics are causing the power performance limiter to be flagged.
There’s not much information available on what this metric is. Still, we can speculate this is a calculated power (W=V*A) output based on the current GPU VID (voltage) and the workload characteristics (current).