GPU Tier List 2026: Graphics Cards by Gaming & Workload

What produces the grade. Raster performance index on our 1000-point scale (RTX 4070 = 1000, averaged over 3-5 game benchmarks), with VRAM gates and a cap for architectures that no longer get game-ready drivers.

How to use it. For a 60-144 Hz 1080p display, the cheapest A is usually the smart buy; S means headroom for high-refresh ultra and future titles. C and below means you will be turning settings down.

Easy to misread. Many cards reach S here — that is honest, not a bug: 1080p is an easy target in 2026. Use the Value column to separate them.

Rubric gpu-1080p-gaming-v1 · price is never an input to this grade.

What produces the grade. The same raster index, with a stricter ladder and hard VRAM gates — under 12GB caps at B, under 8GB at C, because texture budgets at 1440p exceed those buffers in current titles.

How to use it. Aim for A or better for a 1440p 144 Hz panel. If a card you like is capped at B, the tooltip names the reason — usually VRAM, which no price cut fixes.

Easy to misread. A high raster score does not override the VRAM cap. That is deliberate: running out of memory causes stutter, which an average-fps number hides.

Rubric gpu-1440p-gaming-v1 · price is never an input to this grade.

What produces the grade. Raster index against a much higher bar, plus a 16GB practical floor and a cap for cards without a modern upscaler. VR is treated as a 4K-equivalent load.

How to use it. Only A/S are comfortable native 4K. B/C are viable with quality upscaling. Below that, 4K is not the right target for the card.

Easy to misread. This column assumes native rendering. DLSS/FSR quality modes move real-world results up roughly one band.

Rubric gpu-4k-vr-v1 · price is never an input to this grade.

What produces the grade. MEASURED native fps from a ray-tracing benchmark suite (geometric mean of 8 RT titles, no upscaling, no frame generation) — not a synthetic index. Use the resolution buttons above the table to switch context; the tooltip shows all three at once.

How to use it. Check the resolution you actually use — a card can be S at 1080p and C at 4K. B or better is a solid base to add upscaling on top of; D/F means RT is not a realistic option at that resolution.

Easy to misread. These are NATIVE numbers, so they look harsh. DLSS/FSR Quality typically adds ~40-60%, which is exactly why we grade native: it is the honest floor before vendor-specific upscaling.

Thresholds: 1080p S ≥90, A ≥60, B ≥45, C ≥30, D ≥20 fps native · 1440p S ≥90, A ≥60, B ≥45, C ≥30, D ≥20 fps native · 4K S ≥90, A ≥60, B ≥45, C ≥30, D ≥20 fps native

Rubric gpu-raytracing-v2 · price is never an input to this grade.

What produces the grade. VRAM decides WHICH models fit at all; FP16 tensor throughput (ai_score) decides how fast they run. Both are gates — the tier needs the memory first, then the compute.

How to use it. Start from the model size you want: 24GB+ opens 30B-class models, 16GB is comfortable for 13B, 12GB handles 7-13B, 8GB is quantised-7B territory. Only then compare speed.

Easy to misread. A fast 8GB card loses to a slower 24GB card here, and that is correct — a model that does not fit does not run at any speed. N/A appears where we lack a tensor benchmark (Turing, pro and Arc cards), and N/A never means "bad".

Rubric gpu-ai-llm-v1 · price is never an input to this grade.

What produces the grade. Measured Blender Benchmark throughput (samples/min), NOT the AI index — path-tracing runs on RT cores via OptiX/HIP/oneAPI, while the AI column measures FP16 tensor work. VRAM is a hard gate because a scene that does not fit cannot render: under 12GB caps at A, under 8GB at C.

How to use it. Render time scales almost linearly with this number — an S card is roughly twice as fast as a B card on the same scene. Then check VRAM against your typical scene size; capacity decides what you can render at all, speed only decides how long it takes.

Easy to misread. AMD and Intel cards sit lower here than their gaming rank suggests. That is real, not a bias: Cycles is fastest on NVIDIA's OptiX backend. N/A means we have no render benchmark for that model, never that it is bad.

Rubric gpu-rendering-v1 · price is never an input to this grade.

What produces the grade. Board power (TDP) as the efficiency measure, plus a modern video-decode requirement (AV1) — because AV1 streams without hardware decode fall back to the CPU, which defeats the purpose.

How to use it. Read this column as fit, not ability: S/A are right-sized and quiet, C/D are oversized. For a quiet office box, buy the highest tier here that still meets your other needs.

Easy to misread. THE INVERTED ONE. A low grade does NOT mean the card cannot do the work — an RX 9070 XT runs a browser perfectly. It means the card is oversized: you pay for, and power, capability the task never uses. Every other column rewards more muscle; this one rewards the least muscle that suffices.

Rubric gpu-everyday-media-v2 · price is never an input to this grade.