How to test & benchmark your new PC. — Stable. Within spec. Worth what you paid.

Why test your new build

A new PC arrives, you turn it on, it boots, you celebrate. Most builds are fine. But "most" isn't all — components fail, fans don't spin up, RAM kits ship with a bad stick, GPUs have manufacturing defects. The first week of testing has one job: catch the 3-4% of builds that have a problem while you're still inside the return window.

Testing also gives you:

• Performance baselines for future comparison — if your PC feels slower in 12 months, you'll know whether that's drivers, software bloat, or actual hardware drift.
• Thermal headroom verification — if your cooler is mounted wrong or your case airflow is poor, stress testing reveals it before a real game session at midnight.
• Confidence in the build — running a heavy workload and seeing temps stay in spec is the difference between "I think it's fine" and "I know it's fine".

Total time investment: ~3-4 hours of testing spread across the first week. Cheap insurance.

Stability testing — Prime95, MemTest86, FurMark

Stability testing throws sustained heavy load at one component at a time to surface any hidden defect. The components fail under load in ways they don't fail at idle.

CPU — Prime95 small FFTs

Prime95 (free, prime95.com) runs intensive floating-point math that pushes the CPU to maximum sustained load. Use the Small FFTs stress test preset — this is the most CPU-intensive without involving RAM heavily.

Run for 30-60 minutes with HWMonitor or HWiNFO64 open in another window to watch temps. Healthy result:

• No errors reported by Prime95.
• CPU temp stays under 85°C sustained (95°C+ means thermal issues).
• Clock speeds stay near the all-core boost (small dips are normal).
• No system crashes, BSODs or reboots.

RAM — MemTest86

MemTest86 (free, memtest86.com) runs from a bootable USB stick, outside of Windows. This is the gold standard for memory testing because the OS doesn't interfere.

Create a bootable USB with the MemTest86 image, boot from it (set USB first in BIOS boot order), let it run. One full pass minimum — usually 60-90 minutes for 32GB at DDR5-6000. If you suspect issues, leave it running overnight (8+ hours, 8-10 passes).

Healthy result: zero errors after at least one full pass. A single error means the RAM, the motherboard memory slot, or the XMP/EXPO profile has an issue.

GPU — 3DMark Time Spy Stress Test or FurMark

3DMark Time Spy Stress Test (paid 3DMark, ~R450 on Steam) runs Time Spy 20 times in a row and reports a stability score. 97%+ is healthy.

FurMark (free, geeks3d.com) is the classic GPU torture test — pushes the GPU to its thermal and power limit. Run for 20-30 minutes; watch for:

• No visual artefacts (random pixels, lines, colour blocks).
• GPU temp under 85°C (above 90°C suggests airflow/cooler issues).
• No system crashes or driver timeouts.
• Frame rate stable (small variation OK; large dips suggest throttling).

Performance baselines — 3DMark, Cinebench, AIDA64

Stability passes. Now measure how fast the build actually is.

Recommended order:

• 1. 3DMark Time Spy — single most universal gaming benchmark. Get a Graphics score and CPU score; compare both to online median for your exact spec.
• 2. Cinebench R23 (or R24) — multi-core and single-core CPU scores. Quick (10 minutes) and matches well to real productivity workloads.
• 3. AIDA64 Cache & Memory — verify XMP/EXPO is delivering rated bandwidth. DDR5-6000 should show ~85-95 GB/s read; if you see 60-65 GB/s, XMP is off.
• 4. CrystalDiskMark — verify SSD speed. Gen 4 NVMe should hit 6,000-7,000 MB/s sequential read; Gen 5 NVMe 12,000-14,000 MB/s.
• 5. Geekbench 6 — useful for cross-platform comparison if you're researching upgrade impact later.

Game benchmarks

Synthetic scores prove components work to spec. Game benchmarks prove your actual experience matches review-site expectations.

• Cyberpunk 2077 — built-in benchmark, demanding on both CPU and GPU. Path tracing torture test for high-end builds.
• Forza Horizon 5 — built-in benchmark, well-optimised, useful for verifying expected FPS bands.
• F1 2024 — built-in benchmark, easy DLSS/FSR comparison.
• Black Myth Wukong — built-in benchmark tool, modern UE5 workload.
• Total War: Warhammer III — campaign benchmark, CPU-heavy.
• Counter-Strike 2 (CS2) — Workshop benchmark maps for FPS measurement at competitive settings.

Compare your results to YouTube benchmark videos for your exact GPU+CPU combination. If a reviewer with a similar spec reports 95 FPS at 1440p Ultra and you're seeing 70 FPS, dig into the difference: drivers, settings, thermal throttling, XMP, GPU PCIe speed.

Real-game soak test

Synthetics run for 20-30 minutes. Soak tests run for 1-2 hours of real gameplay. The two catch different problems.

Run an actual game session of 1-2 hours in something you enjoy. Watch for:

• Frame rate hitches, micro-stutters or sudden drops.
• Driver timeout errors (TDR) — Windows reports "Display driver has stopped responding".
• Thermal creep — temps slowly rising past 85°C as the case heats up.
• Fan noise escalating to unusable levels.
• Game crashes to desktop.
• Sudden reboots or BSODs.

A build that passes Prime95, MemTest, 3DMark and a 90-minute Cyberpunk session is genuinely healthy.

Where to compare scores

Your score is meaningless without context. These sources give you honest comparisons:

• 3DMark online database — official, automatically populated by users worldwide. Filter by your exact GPU and CPU to see percentile rank.
• PCPartPicker community benchmarks — users post their builds with scores; useful for full-system comparison.
• Geekbench Browser — every Geekbench run is publicly searchable. Compare to identical hardware.
• Cinebench R23 leaderboards — Maxon publishes single-core and multi-core leaderboards by CPU model.
• YouTube benchmark videos — channels like Hardware Unboxed, Gamers Nexus, Optimum and Daniel Owen publish per-game FPS data for nearly every CPU+GPU combo at multiple resolutions.

What good results look like

"Healthy" is a range, not a number. Hardware variance, BIOS revisions, ambient temperature, fan curves and background processes all contribute small differences. The right mental model:

Red flags vs healthy variance

Common causes of 10-20% under-spec performance:

• XMP/EXPO disabled. RAM running at JEDEC baseline. Single biggest performance hit on Ryzen especially. Fix in BIOS.
• GPU at PCIe x4 instead of x16. Check GPU-Z's bus interface readout. If it shows x4 or x8, you've probably plugged the GPU into the second PCIe slot. Move to the top slot.
• Resizable BAR disabled. 2-5% gaming hit. Fix in BIOS.
• Power plan set to Balanced or Power Saver. Should be Ultimate Performance or High Performance. Check in Windows Power Options.
• Cooler not properly mounted. CPU thermal throttling at 95°C+ caps performance. Re-seat cooler, fresh paste.
• Background processes. Antivirus scans, OneDrive sync, Windows Update can tank benchmark scores. Re-run benchmarks with these paused.
• Stale GPU drivers. Update to latest Nvidia / AMD / Intel Arc drivers before benchmarking.

When to RMA

After ruling out configuration issues, persistent under-performance or instability is a hardware problem. RMA candidates:

• Dead on arrival. Component shows no signs of life — CPU fan doesn't spin (CPU dead), no display output (GPU or motherboard), drive not appearing in BIOS (SSD/NVMe dead).
• MemTest86 errors after at least 2 retests at different speeds. One stick is bad.
• GPU artefacts in stress test. Random pixels, lines, colour blocks during FurMark or 3DMark = dying GPU memory or core.
• Fan stopped spinning on day 1. Case fan, AIO pump, GPU fan — any fan that doesn't spin from cold boot is faulty.
• CrystalDiskMark shows <20% of expected SSD speed. Verify the SSD is in the right M.2 slot (some boards have a Gen 3 slot mixed with Gen 4/5).
• Persistent instability after BIOS reset to defaults, XMP off, clean Windows install.

In SA, RMA the component to where you bought it within the return window (typically 7-14 days). Most SA retailers including Evetech, Rebel Tech, Wootware and Takealot have a no-quibble return policy for clearly defective parts.

Free SA-friendly benchmark tools

Everything in this guide is free or has a free mode that covers what you need. No download from suspicious "warez" sites required.

Common testing mistakes

Running stress tests without monitoring temps. Pointless. Always have HWiNFO64 or HWMonitor open to capture max temps during the run.

Benchmarking with antivirus / OneDrive / Discord overlay running. Background processes can drop 3DMark scores 5-10%. Close non-essential apps before testing.

Comparing to UserBenchmark. Use 3DMark, Geekbench Browser or Cinebench official leaderboards instead.

Treating a single benchmark run as definitive. Run twice; if there's >5% variance between runs, run a third. Background activity or thermal state can affect single runs.

Not updating GPU drivers before testing. Day-1 driver vs latest can differ 5-15% in benchmarks. Update first.

Testing on an unconfigured BIOS. XMP off, Resizable BAR off, CSM on — these all hurt scores. Set BIOS up first (see our BIOS setup guide), then benchmark.

Key takeaways

1. Stability first — Prime95 30-60 min, MemTest86 one full pass, GPU stress 20-30 min.
2. Performance baseline with 3DMark + Cinebench + AIDA64 + CrystalDiskMark.
3. Within 10% of online median = healthy. 20%+ below = something is wrong.
4. Real-game soak test 1-2 hours catches what synthetics miss (thermal creep, driver crashes).
5. Catch DOA components within 7-14 days while still in the return window — 94% catch rate when you test, 41% when you don't.

Frequently asked questions