How to Read a Benchmark Without Being Misled

Benchmark charts are the main evidence people use to spend fifty thousand rupees, and most people read them wrong.

Average fps is the weakest number

An average tells you nothing about consistency, and consistency is what you feel.

Two cards both averaging 90fps: one delivers a steady 85–95, the other alternates between 130 and 45. Identical averages, completely different experiences. The second one feels broken.

1% and 0.1% lows are the real metric

The 1% low is the frame rate of the slowest 1% of frames. The 0.1% low is the worst tenth of a percent.

These capture stutter, and stutter is what makes a game feel bad.

The gap between average and 1% low is the number to look at. A card averaging 90 with 1% lows at 75 is smooth. A card averaging 100 with 1% lows at 40 is not, despite the better average.

If a review only publishes averages, it can’t answer the question you’re asking. Frame time graphs are better still — a flat line is good, spikes are stutter.

The test system matters enormously

Reviewers test GPUs on the fastest available CPU deliberately, to remove the CPU as a limit and isolate graphics performance.

That’s methodologically correct and it means the numbers don’t apply to you if your CPU is slower. A card showing 200fps on a top-tier CPU might show 130 on yours, because your CPU can’t feed it.

Always check what CPU, what memory speed and what resolution were used. The resolution matters most — GPU differences compress at 1080p (CPU-limited) and expand at 4K.

Settings disclosure

“Ultra settings” means different things in different games and different reviews. A chart without stated settings is not reproducible and not useful.

Watch particularly for:

• Whether upscaling was on, and at what preset
• Whether ray tracing was enabled
• Whether frame generation was included in the fps number

That last one is a real problem. A chart mixing native frame rates with frame-generated ones is comparing different things — see the upscaling explainer for why generated frames don’t reduce latency.

Chart tricks to watch for

Truncated Y-axis. Starting the axis at 60 instead of 0 makes a 10% difference look like a 300% one. Extremely common in manufacturer marketing.

Relative percentage charts with no absolute numbers. “40% faster” than what, at what frame rate? 40% faster than 20fps is still unplayable.

Cherry-picked titles. Every architecture has games it does unusually well in. A three-game sample proves nothing; look for ten or more.

Averages of averages. A single “average across all games” figure hides that one card is much better in the specific game you play.

Missing 1% lows — by now you know why.

Synthetic versus in-game

Synthetic benchmarks — 3DMark, Geekbench, Cinebench — are useful for one thing: comparing components on an identical, repeatable workload.

They’re poor predictors of game performance because games are not synthetic workloads. A CPU that scores brilliantly in a multi-core render test may be mediocre in a game that uses two threads heavily.

Use synthetics to rank components. Use game benchmarks to decide what to buy.

The PC builder shows both, deliberately — the published synthetic scores for each part, and separately a modelled game frame rate. They tell you different things.

The model versus measurement distinction

Worth being explicit, including about this site’s own tool: the frame rates in the PC builder are modelled from performance indices, not measured on each combination.

That’s useful for comparing tiers and finding bottlenecks. It is not review data, and anything claiming otherwise about any tool should be treated sceptically. Real measurement means someone assembled that exact configuration and ran it.

What to actually do before buying

1. Find two or three independent reviews of the exact model you’re considering
2. Check they tested at your resolution
3. Look at 1% lows, not averages
4. Check the specific games you play, not the aggregate
5. Note the test CPU and adjust expectations if yours is slower
6. Check whether upscaling or frame generation is in the numbers

Six checks, ten minutes, on a purchase you’ll live with for years.

---

Bottom line: read 1% lows rather than averages, check the test system and resolution match yours, be suspicious of truncated axes and small game samples, and treat any modelled figure — including this site’s — as a tier estimate rather than measurement.