Loading...
Frames Per Second (FPS) is the most familiar GPU performance metric, measuring how many complete frames your GPU renders each second. While seemingly straightforward, FPS interpretation requires understanding its limitations and what different values mean for real-world performance.
Average FPS is a simple metric but hides critical information about frame time consistency. Consider two scenarios:
Both scenarios have the same average FPS, but Scenario B will feel stuttery and inconsistent. This is why frame time analysis is crucial for understanding true performance quality.
We calculate FPS using a 60-frame rolling average, providing smoothed real-time feedback during the test. This approach filters out momentary spikes while remaining responsive to sustained performance changes, making it ideal for detecting thermal throttling over 10-30 minute stress tests.
Frame time (measured in milliseconds) is the inverse of FPS: how long it takes to render a single frame. While FPS gives overall throughput, frame times reveal consistency and worst-case performance.
P95 and P99 are percentile metrics that answer the question: "What's the frame time that 95% (or 99%) of frames are faster than?" These metrics focus on the worst-performing frames, which disproportionately impact perceived smoothness.
Human perception is highly sensitive to frame time spikes. A single 50ms frame (20 FPS) in the middle of otherwise smooth 16ms frames (60 FPS) is immediately noticeable as a "hitch" or stutter. P95/P99 metrics quantify these worst-case scenarios that determine perceived smoothness.
For competitive gaming or professional visualization, aim for P99 frame times under 20ms (50 FPS). For general use, P99 under 30ms (33 FPS) is acceptable if average FPS is significantly higher.
Stutter Frequency Density (SFD) measures what percentage of frames exhibit significant frame time increases compared to recent history. It quantifies how often your GPU produces "micro-stutters" that disrupt smooth motion.
A frame is considered a "stutter" if its frame time exceeds the recent average by a threshold (typically 150-200%). For example, if recent frames averaged 16ms and the current frame takes 26ms, that's a significant spike flagged as stutter.
Over a 10-second window at 60 FPS (600 frames):
Common culprits for elevated SFD scores:
The Sustained Load Score (SLS) measures GPU performance stability over time by comparing final performance to initial baseline. It's the single best metric for detecting thermal throttling and cooling efficiency.
SLS = (Final FPS / Initial FPS) × 100%Initial FPS: Average FPS during the first 60 seconds (after warmup)
Final FPS: Average FPS during the last 60 seconds of the test
Peak benchmark scores often measure performance in short bursts before thermal limits kick in. SLS reveals sustained performance during long gaming sessions, video rendering, or scientific simulations where thermal equilibrium matters more than cold-start speed.
A GPU with 10% lower peak performance but 95% SLS will outperform a faster GPU with 80% SLS in any workload lasting more than a few minutes. This is especially critical for laptops where cooling constraints dominate long-term performance.
Thermal throttling occurs when a GPU reduces its clock speeds to prevent overheating. Understanding throttling patterns helps diagnose cooling issues and optimize system performance.
Even without temperature monitoring, our benchmark reveals throttling through performance patterns:
Understanding your GPU's performance in context requires comparing against similar hardware and identifying whether your results align with expected capabilities.
Laptop GPUs with identical model numbers to desktop counterparts typically perform 20-40% slower due to power and thermal constraints. A laptop RTX 4070 operates at 80-115W, while the desktop version runs at 200W+. Adjust expectations accordingly when comparing results.
Our volume shader benchmark heavily utilizes floating-point arithmetic (FP32). GPUs optimized for different workloads (tensor cores for AI, RT cores for ray tracing) may perform differently than synthetic benchmarks like 3DMark predict. This provides valuable insight into shader compute performance specifically.
Your results warrant investigation if: