Every current NVIDIA RTX 50, AMD RX 9000 & Intel Arc card — plus the last-gen parts still on shelves — scored across the jobs people actually buy them for: 4K raster, ray tracing, local AI/LLM, content creation and efficiency. The same card scores very differently per job, because the truth is in the benchmark numbers — independent FPS-per-5090, VRAM and perf-per-watt — not the vendor's TFLOPS slide.
Start with the job — it drives every chart below
Each bubble is a card. Right = more capable, left = cheaper — top-left is the value sweet spot. Color = generation. Capability is a transparent multi-domain index from real benchmarks (see Method). Re-scores when you pick a job above — pick Efficiency / value or Local AI / LLM and watch the order change.
Cards (warm) linked to the ecosystems they tap into (cyan) — DLSS / FSR / XeSS upscaling, frame-gen, AV1 encode, CUDA / ROCm, PCIe 5.0. Drag a node; hover to trace a stack. Built from each card's documented feature support.
Each card is scored on six jobs from real benchmark numbers, summed and normalized to 0–100. The numeric axes read measured values; the categorical ones read documented feature support:
raster 4K-FPS as % of the RTX 5090 → 5090-class=3.5 · 5080=3 · 5070 Ti-class=2.6 … entry=0.4 | ray tracing RT-% of 5090, same curve to 2.5 | LLM / AI by VRAM — 32GB=3 · 24=2.5 · 16=2 · 12=1.5 · 8=1 | upscaling DLSS 4+MFG=2 · DLSS 3=1.6 · FSR 4=1.5 · FSR 3=1.2 · XeSS 2=1.1 | content dual-AV1 encoders=2 · AV1+NVENC=1.6 · AV1=1.3 | efficiency raster-per-watt → champ=2.5 … floor=0.4
The headline number is a multi-domain average, so a one-trick flagship doesn't auto-win — the RTX 5090 tops five axes but sits mid-pack on efficiency, and the RX 9070 is the perf-per-watt champ. Pick a job above to rank for that domain. Measured beats claimed: independent 4K benchmarks (Tom's Hardware + TechPowerUp, shown as % of the 5090) sit next to the vendor spec and drive scoring.
What we left out, on purpose. We dropped a crypto-mining axis: GPU mining is effectively obsolete in 2026 — Ethereum moved to proof-of-stake in 2022 and Kaspa is now ASIC-dominated, so reviewers no longer benchmark it and the scraps of data weren't comparable (stock vs power-limited watts). Upscaling / frame-gen replaces it. Per-card tok/s is shown but not scored — it swings with model, quant and backend, so the LLM score is driven by VRAM, the real "what fits" constraint. The RTX 5050's raster figure is an interpolated estimate (flagged on its card); every other number is a verified benchmark. Accuracy is priority #1 — nothing here is fabricated.