Pick a baseline card and every other GPU is shown relative to it (baseline = 1.00×). Blender Cycles, render-farm cost (buy vs rent, 1–8 GPUs), local LLM inference and image/video generation.
Solid bar = the GPU count selected above; faint bar behind it = one card. Measured single-card figures are from Techgage's Blender 3.5 Scanlands chart (OptiX / HIP / oneAPI). Cards marked est are scaled from similar cards. Multi-GPU uses the scaling mode set on the Render farm tab. Click a GPU name to make it the baseline.
Runpod-rentable cards only. Hollow dot = one GPU, solid dot = the selected count. Lower-left is faster and cheaper. Both axes are logarithmic.
Job size is fixed so one RTX 5090 takes exactly 1 hour; other cards scale by their Scanlands time. Animation frames splits frames across GPUs, so it scales almost perfectly (98%). Single still uses Cycles multi-device rendering on one frame (about 85% efficiency per added card), and VRAM does not pool. Runpod $/h is the per-GPU price from a live catalog read on 15 Oct 2026; "max" is the most GPUs of that type Runpod offers in one pod on that tier. Power cost assumes 85% of board power per GPU plus 80 W per machine. Break-even = buy price ÷ (rent £/job − power £/job).
Big number = speed relative to the baseline (at the same GPU count). Small number = estimated tokens/s. With more than one GPU, VRAM pools via layer split: bigger models fit, but a single user's speed stays near one card's (about 8% split overhead). Token generation is modelled from memory bandwidth; prefill from matrix throughput. Offloaded layers run at ~60 GB/s system RAM. Qwen3.8 27B uses hybrid linear attention (only 16 of 64 layers keep a KV cache), so long contexts cost little memory. Click a column to select and sort.
Big number = speed relative to baseline. Second line = estimated wall time for the reference job; third line = Runpod rental cost for that job (selected tier, $). One job runs on one GPU; with N GPUs you get N jobs at once. MiniMax H3 (Hailuo 3) has no public weights, so it cannot run on any local card; it is shown at its API price (about $0.08 per output second at 768p, $0.13/s at 2K) for comparison with renting a GPU for Wan or HunyuanVideo. FP8 on RTX 40/50 and NVFP4 on RTX 50 can be faster than shown.
Prices are editable and saved in this browser. Defaults are rough UK street prices for October 2026, when RAM shortages pushed new high-end cards well above launch price (RTX 5090 listed from about £4,550 at UK retailers). Used prices are typical eBay / CEX levels. Check current listings before buying. FP16 matrix TFLOPS are dense, FP32 accumulate; GTX 900/10-series show FP32 shader TFLOPS. Pro/datacenter cards use effective throughput normalised to real-world results.