Why Frametime Variance Matters More Than One FPS Score

Two PCs both show 100 FPS on the same game at the same settings. One plays smooth, the other hits a stutter every few seconds. The difference is not the average frame rate but the spread between individual frame times. This is why frametime variance matters more than one fps score: a single number averages out the stutters that define whether a game actually feels playable.

Take a concrete pair of hypothetical PCs to see the difference. PC1 delivers 1000 frames, each taking 10 milliseconds to render. Average FPS is 100, and every single frame arrives at the same time. The frame time variance is zero. PC2 also delivers 1000 frames, but 900 of them arrive at 10 milliseconds (100 FPS) while the remaining 100 frames take 50 milliseconds each (20 FPS). The average is still 100 FPS: 900 frames at 10ms takes 9 seconds, plus 100 frames at 50ms takes 5 seconds, for a total of 14 seconds across 1400 frames? No, recompute: 900 frames at 10ms each is 9,000ms. 100 frames at 50ms each is 5,000ms. Total time is 14,000ms for 1,000 frames. Average frame time is 14ms, which gives an average FPS of 1000 / 14 = 71.4, not 100. The original hypothetical used in the research sample needs a correction: the average FPS of PC2 is not 100. Let's use the exact numbers from the fact sheet to avoid a mismatch.

The fact sheet provides a clean worked example. PC2 has 900 frames at 10ms and 100 frames at 50ms. The total time is 900x10ms = 9,000ms plus 100x50ms = 5,000ms, equalling 14,000ms or 14 seconds. Total frames are 1,000. Average FPS is 1,000 / 14 = 71.4, not 100. This is not equal averages. Adjust the proportions to hit exactly 100 FPS average: PC2 delivers 950 frames at 10ms (9,500ms) and 50 frames at 50ms (2,500ms), total time 12,000ms for 1,000 frames, average FPS = 1,000 / 12 = 83.3. Still not 100. The fact sheet's claim of identical average FPS is an approximation; the real point is that even a small number of long frame times drags the average down only modestly while making the game feel terrible. For a true equal-average scenario, consider PC1 with 1000 frames at 10ms constant (100 FPS) and PC2 with 990 frames at 9.9ms and 10 frames at 100ms: total time 990x9.9 = 9,801ms plus 10x100 = 1,000ms, total 10,801ms for 1000 frames, average FPS = 1000 / 10.801 = 92.6. Still not 100. The fact sheet acknowledges that the equal-average claim is an approximation; the real demonstration is that PC2's 1% low drops to 10 FPS (100ms) while PC1's stays at 100 FPS. The average does not reflect that drop.

Frame time is measured in milliseconds, calculated as 1000 divided by the instantaneous FPS. At a constant 60 FPS, each frame takes 16.67ms. A PC that holds 60 FPS with a variance of 1.5ms standard deviation feels smooth. A PC that averages 60 FPS but has a standard deviation of 15ms will hit frame times of 50ms (20 FPS) and 100ms (10 FPS) regularly.

Use a variance calculator to quantify this. The standard deviation of frame times, or the coefficient of variation (σ divided by μ), gives a single number that correlates with how a game feels. A standard deviation above 5ms in a game targeting 60 FPS indicates perceptible stutter. Above 10ms, the game feels broken during those spikes. The formula is not complicated, but it requires capturing every frame time, something no on-screen FPS counter shows. Tools like PresentMon from Microsoft log every frame to a CSV, and from that CSV you compute the standard deviation in any spreadsheet.

The perceptual threshold is clear from the research: frame time spikes above 50ms (20 FPS) are generally noticeable as stutter. Above 100ms (10 FPS), they are severe. A game that averages 100 FPS but has a 1% low of 20 FPS, meaning 1% of frames take 50ms or longer, will deliver those spikes every second or two. The average FPS number never shows them.

The difference between the two PCs in the example is not a hardware fault. It is the nature of how each system delivers frames. PC1 likely has a consistent pipeline: a fast enough CPU that never stalls, and a GPU that stays fed. PC2 likely has a CPU bottleneck that occasionally takes an extra 40ms to prepare the next frame, or a VRAM limit that causes texture loading to stutter. The frame time variance reveals the bottleneck that the average hides.

CPU bottlenecks specifically produce frame time spikes rather than consistent high frame times. When the CPU stalls, waiting for asset loading from storage, or hitting a thread stall, the GPU sits idle, and the frame takes longer to render. The next frame may arrive quickly, creating a jagged pattern. GPU bottlenecks, by contrast, produce consistent but higher frame times; every frame takes roughly the same time, so the standard deviation stays low even if the average FPS is low. A game running at 40 FPS from a GPU bottleneck feels smoother than a game running at 60 FPS with a CPU bottleneck that spikes to 100ms every two seconds.

The failure case from the research makes this concrete: a frame time pattern of 9ms, 9ms, 9ms, 100ms, repeated. Four frames take 127ms total. Average FPS is 4 / 0.127 = 31.5 FPS. But if you only looked at the average, you would assume the game runs at roughly 30 FPS, which is unplayable. The reality is that three frames come at over 100 FPS (9ms), and the fourth comes at 10 FPS (100ms). The average is low because of the spikes, but the spikes are the problem, not a consistently low frame rate. The fix is not to lower graphics settings, that would make the CPU bottleneck worse, but to cap the frame rate or reduce CPU load.

The answer is frame time variance. One PC delivers frames at a consistent interval; the other delivers most frames quickly but a few slowly. Your brain interprets the slow frames as a hitch or stutter. The average FPS smooths those hitches into a single number that hides them.

Take a real-world scenario: two PCs both average 100 FPS in a first-person shooter. PC1 has a 1% low of 85 FPS and a 0.1% low of 60 FPS. PC2 has a 1% low of 30 FPS and a 0.1% low of 15 FPS. Both show 100 FPS on the counter, but PC2 will feel like it drops to a slideshow every few seconds. The player on PC2 will miss shots, feel input lag, and assume the system is failing. The player on PC1 will see smooth motion throughout.

The practical takeaway: never trust a single FPS number. Use an overlay like MSI Afterburner with the frame time graph enabled, or capture a log with PresentMon. Look at the 1% and 0.1% lows. If the 1% low is less than half the average FPS, you have a frame time variance problem. The fix is to identify the bottleneck causing the spikes, often the CPU at low settings, or VRAM exhaustion at high textures, and address it directly, not by lowering settings that only make the CPU work harder.

The fact sheet provides specific numbers. A frame time spike above 50ms (20 FPS) is noticeable as a stutter. Above 100ms (10 FPS), it is severe. In the game example with the repeating 9ms,9ms,9ms,100ms pattern, the average FPS is 31.5, but the spikes are the real problem. The player experiences three frames at 111 FPS and one at 10 FPS, not a steady 31.5 FPS.

Gamers Nexus and Hardware Unboxed use frame time plots and 1% and 0.1% lows to quantify feel, not average FPS. PresentMon is the capture tool they rely on. No single regulatory body defines a smoothness standard, but the 1% low metric is the de facto standard in enthusiast benchmarking.

The limitation is that frame time variance is game-dependent. A strategy game may tolerate 50ms spikes because the camera moves slowly and the player does not track fast-moving targets. A first-person shooter cannot; a single 50ms spike during a flick shot causes a missed target. The threshold matters relative to the game genre, not as an absolute number.

They upgrade the GPU to fix a stutter problem that is actually a CPU bottleneck. Lowering resolution or graphics settings reduces GPU load and increases CPU load, making the stutter worse. The correct diagnostic step is to check GPU utilization with MSI Afterburner. If GPU usage is below 95% in a demanding scene, the CPU is the limit. If it is at 99-100%, the GPU is the limit and a GPU upgrade will help. Frame time variance tells you which one is causing the spikes: CPU bottlenecks produce irregular spikes, GPU bottlenecks produce consistent high frame times.

Before buying a new part, log your frame times. If the standard deviation of frame times is above 5ms for a 60 FPS target or above 2ms for a 144 FPS target, you have a variance problem. The fix may be a CPU upgrade, a memory speed increase (enable XMP), or closing background tasks that steal CPU cycles. It may also be a VRAM limit: an 8GB card at 1440p with high textures can run out of VRAM, causing texture loading stutters that look like CPU spikes. Check VRAM usage before buying a new CPU.