A live ranking of UK graphics cards available right now — both new at retail and used on the secondary market — scored on 3DMark Time Spy ÷ price, then weighted to credit the CUDA tax, the VRAM ceiling, and the warranty haircut that comes with buying used.
UK pricing is verified April 2026; other regions extrapolated from typical regional multipliers (US ≈ 1.10×, EU ≈ 1.15×, Australia ≈ 1.65×, Asia ≈ 0.92×) plus card-level corrections where I have direct US data. Used market liquidity differs meaningfully — US used GTX 1080 Ti is ~$95, UK is ~£130. Treat extrapolated prices as ballpark.
Pure raster + ray tracing performance. Modulated by your resolution and refresh rate selections below.
Compute, codec hardware, VRAM. The 4:2:2 hardware decode checkbox below transforms this scoring dramatically.
AV1 encode quality + parallel encode sessions. Toggle the AV1 required and dual NVENC checkboxes for hard requirements.
VRAM dominates here. The Local LLM 24GB+ checkbox is the meaningful filter for serious AI work.
OptiX (RT cores) + VRAM. The CUDA-only software checkbox punishes AMD/Intel for Octane, Topaz, RED RAW workflows.
Cards near your target get a boost. Going under is fine (you save money) — gentle penalty. Going over is harsh — you literally can't afford it. Toggle off to disable budget weighting entirely.
Width of the budget curve. Strict: only cards within ~£100 of target rank well. Loose: cards within ~£300 each side score reasonably. The asymmetry persists — over-budget always falls off faster than under.
Stacks on top of task / capability logic. Default 60/40 is a mild Nvidia tilt — most task scoring already accounts for ecosystem. Slide left for vendor neutrality, right to make non-CUDA cards near-irrelevant.
At 50%, used cards take a 25% penalty — warranty loss, mining/age risk, older feature gens. CEX UK or eBay verified seller? Slide it down. Random unverified listing? Slide it up.
Switch to raw to see the unweighted 3DMark/£ ranking — sanity check against your active preferences.
Each GPU's raw value is its 3DMark Time Spy score divided by its lowest current UK price (new from retailers, used from eBay mid-market), normalised onto a 0–100 scale.
The weighted score stacks four layers on top of raw value:
Budget — skewed Gaussian. Cards near your target price get a boost; cards far off get penalised. The asymmetry matters: going under budget is gentle (you're saving money), going over is harsh (you literally can't afford it). The right tail of the curve falls off about 2.5× faster than the left. The flexibility slider widens or narrows the curve while preserving asymmetry.
Task priorities — normalised, not compounding. Sliders share a budget: setting all five to 100 is mathematically equivalent to setting all five to 20. Each task gets a proportional share of total intent. The strength of override is driven by your strongest single slider (the intensity meter), not the sum. This means you can express "I care about gaming and video equally" without worrying about whether to use 50/50 or 100/100.
Hard capabilities — checkboxes that transform scoring. Each active checkbox adds bonuses to capable cards and penalties to incapable ones, applied before the task averaging. Toggling "4:2:2 H.264 hwdec" pushes RTX 50 cards up and pushes everything else off the cliff for video work. Toggling "Local LLM 24GB+" reorders the entire AI ranking. These cross-weight automatically — a card without 4:2:2 doesn't get penalised in gaming, only video.
Context — dropdowns that nudge. Resolution, refresh rate, and upscaling preference modify gaming scoring specifically. 4K/8K reward VRAM and high-tier compute. 240Hz+ rewards DLSS 4 multi-frame-gen (RTX 50 only). "Native preferred" rewards strong-raster AMD cards. These are gentler than capability checkboxes — they shift the balance, not flip it.
NVIDIA bias is now a manual override on top of the above (default 60/40, was 75/25). Most ecosystem preference is already baked into task scoring — this slider exists for cases where you want to amplify or neutralise that further.
Used penalty at 50% knocks 25% off used-market cards. CEX UK or eBay verified seller? Slide it down. Random unverified listing? Slide it up.
41 cards across 5 regions: 17 available new (lowest current retail prices) and 24 used contenders spanning eight years of architectures from 2017's Pascal through 2026's Blackwell. The 30/40-series new market has largely seized — NVIDIA effectively EOL'd 16GB Blackwell production, the AI/GDDR memory crunch pushed retail prices above launch MSRP — so the secondary market is where most realistic value lives.
Regional pricing: UK prices are verified April 2026 from Overclockers / Scan / Amazon UK and eBay UK realised sales. US data is verified for current cards from bestvaluegpu.com US. EU, Australia, and Asia prices are extrapolated from typical regional multipliers (≈1.15× UK for EU, ≈1.65× for Australia due to distributor markup, ≈0.92× for Asia/Singapore/HK) with card-level corrections where direct data exists. Used market liquidity differs meaningfully — a US used GTX 1080 Ti is ~$95, UK is ~£130. Region auto-detected from your browser timezone; click any region button to override.
Standout raw value picks: the GTX 1080 Ti at ~£130 used for legacy 1440p gaming (raw value 75!), the RX 6800 XT at £260 used (75 raw), and the RTX 4070 Super used at £445. Standout capability picks: the RTX 5070 Ti as the cheapest card with both 4:2:2 hardware decode AND dual NVENC; the RTX 3090 / 4090 / 5090 for local AI where 24-32GB VRAM matters most.
Task scores reflect documented hardware support and ecosystem maturity, sourced primarily from Puget Systems' 2025 content creation roundups, NVIDIA's official NVENC/NVDEC matrix, and Lambda Labs ML benchmarks. Pascal/Polaris/Vega cards score very low across non-gaming tasks — they lack AV1 encode, RT cores, and modern tensor performance. They survive only as cheap 1080p gaming options, which the calculator correctly reflects.