中
GPU LADDER · MINIMAX H3

MiniMax H3 GPU ladder

Every GPU runs MiniMax H3 with the same models on the same reference PC, ranked by generation time with the RTX 4090 at 100. Prices are 2026-10-03 references: new for cards in production, used for discontinued ones, market price for modded cards.

# GPU Score 864×480·5s 1344×768·5s Price pts/$100
1 RTX PRO 6000 Blackwell 96GB 96GB · 1792 GB/s 149 1 min 2 s 2 min 1 s $16,000 new 0.9 2 RTX 5090 32GB 32GB · 1792 GB/s 139 1 min 6 s 2 min 11 s $4,899 new 2.8 3 RTX 5090 D 32GB 32GB · 1792 GB/s 121 1 min 14 s 2 min 32 s — — 4 RTX 5090 D V2 24GB 24GB · 1344 GB/s 108 1 min 23 s 2 min 52 s — — 5 RTX 4090 48GB (modded) 48GB · 1008 GB/s modded · at your own risk 100 1 min 29 s 3 min 6 s — — 6 RTX 4090 24GB 24GB · 1008 GB/s baseline 100 1 min 29 s 3 min 6 s $2,800 used 3.6 7 RTX 4090 D 24GB 24GB · 1008 GB/s 95 1 min 33 s 3 min 16 s — — 8 RTX 5080 16GB 16GB · 960 GB/s 83 1 min 45 s 3 min 46 s $1,599 new 5.2 9 RTX 5070 Ti 16GB 16GB · 896 GB/s 73 2 min 4 min 22 s $1,119 new 6.5 10 RTX 4080 Super 32GB (modded) 32GB · 736 GB/s modded · at your own risk 72 2 min 1 s 4 min 25 s — — 11 RTX 4080 Super 16GB 16GB · 736 GB/s 72 2 min 1 s 4 min 25 s $1,099 used 6.6 12 RTX 4080 32GB (modded) 32GB · 717 GB/s modded · at your own risk 69 2 min 6 s 4 min 37 s — — 13 RTX 4080 16GB 16GB · 717 GB/s 69 2 min 6 s 4 min 37 s $1,084 used 6.4 14 RTX 3090 24GB 24GB · 936 GB/s 67 2 min 9 s 4 min 43 s $1,199 used 5.6 15 RTX 3080 Ti 12GB 12GB · 912 GB/s 65 2 min 12 s 4 min 52 s $549 used 11.8 16 RTX 4070 Ti Super 16GB 16GB · 672 GB/s 64 2 min 15 s 4 min 59 s $939 used 6.8 17 RTX 3080 12GB 12GB · 912 GB/s 62 2 min 18 s 5 min 7 s ≈$500 used 12.4 18 RTX 3080 20GB (modded) 20GB · 760 GB/s modded · at your own risk 57 2 min 32 s 5 min 39 s — — 19 RTX 3080 10GB 10GB · 760 GB/s 57 2 min 32 s 5 min 39 s $350 used 16.3 20 RTX 4070 Ti 12GB 12GB · 504 GB/s 54 2 min 39 s 5 min 58 s $579 used 9.3 21 RTX 5070 12GB 12GB · 672 GB/s 54 2 min 38 s 5 min 53 s $799 new 6.8 22 RTX 4070 Super 12GB 12GB · 504 GB/s 51 2 min 48 s 6 min 19 s $675 used 7.6 23 RTX 4070 12GB 12GB · 504 GB/s 47 3 min 3 s 6 min 56 s $590 used 8.0 24 RTX 3070 Ti 8GB 8GB · 608 GB/s 44 3 min 12 s 7 min 17 s $289 used 15.2 25 RTX 5060 Ti 16GB 16GB · 448 GB/s 40 3 min 32 s 8 min 6 s $779 new 5.1 26 RTX 5060 Ti 8GB 8GB · 448 GB/s 40 3 min 32 s 8 min 6 s $499 new 8.0 27 RTX 3070 16GB (modded) 16GB · 448 GB/s modded · at your own risk 37 3 min 48 s 8 min 43 s — — 28 RTX 3070 8GB 8GB · 448 GB/s 37 3 min 48 s 8 min 43 s $230 used 16.1 29 RTX 2080 Ti 22GB (modded) 22GB · 616 GB/s modded · at your own risk 36 3 min 58 s 9 min 7 s — — 30 RTX 2080 Ti 11GB 11GB · 616 GB/s 36 3 min 58 s 9 min 7 s $255 used 14.1 31 RTX 5060 8GB 8GB · 448 GB/s 36 3 min 54 s 8 min 58 s $379 new 9.5 32 RTX 3060 Ti 8GB 8GB · 448 GB/s 33 4 min 13 s 9 min 44 s $248 used 13.3 33 RTX 4060 Ti 16GB 16GB · 288 GB/s 31 4 min 29 s 10 min 22 s ≈$790 used 3.9 34 RTX 4060 Ti 8GB 8GB · 288 GB/s 31 4 min 29 s 10 min 22 s $330 used 9.4 35 RTX 3060 12GB 12GB · 360 GB/s 27 5 min 13 s 12 min 9 s $280 used 9.6 36 RTX 4060 8GB 8GB · 272 GB/s 26 5 min 29 s 12 min 49 s $297 used 8.8 37 RTX 5050 8GB 8GB · 320 GB/s 26 5 min 25 s 12 min 40 s $309 new 8.4 38 RTX 2070 Super 8GB 8GB · 448 GB/s 25 5 min 31 s 12 min 53 s $189 used 13.2 39 RTX 2060 Super 8GB 8GB · 448 GB/s 23 6 min 9 s 14 min 24 s $175 used 13.1 40 RTX 3060 8GB 8GB · 240 GB/s 22 6 min 18 s 14 min 42 s — — 41 RTX 2060 12GB 12GB · 336 GB/s 20 6 min 54 s 16 min 11 s — — 42 RTX 3050 8GB 8GB · 224 GB/s 19 7 min 16 s 17 min 21 s $165 used 11.5 43 RTX 2060 6GB 6GB · 336 GB/s 19 7 min 8 s 16 min 55 s $135 used 14.1 44 GTX 1080 Ti 11GB 11GB · 484 GB/s 13 10 min 28 s 27 min 52 s $100 used 13.0 45 GTX 1080 8GB 8GB · 320 GB/s 10 13 min 27 s 37 min 39 s $90 used 11.1 46 GTX 1070 8GB 8GB · 256 GB/s 8 16 min 44 s 48 min 26 s $69 used 11.6 47 GTX 1060 6GB 6GB · 192 GB/s 6 22 min 21 s 1.1 h $55 used 10.9 — Apple M4 Max 128GB (unified memory) 96GB · 546 GB/s MLX path · not ranked — — ≈$3,499 new — — GTX 1050 Ti 4GB 4GB · 112 GB/s Won't run — — $65 used —

⚠ = not enough VRAM at this tier, weights get offloaded; ≈ = small sample, stale or converted price (hover for range and source). Click a row for the per-card page.

How it is measured

  • Reference PC: 64GB DDR5-6000 dual channel, NVMe Gen4, --fast-disk on, Windows.
  • Same models for all: pruned INT8 DiT + NVFP4 AWQ text encoder + LightX2V 4-step Turbo LoRA, no attention speed-up. Cards that cannot fit the weights offload them, and the slowdown shows in the score.
  • Score = geometric mean of speed vs RTX 4090 at 864×480·5s and 1344×768·5s (official max), × 100. The baseline card takes about 1 min 29 s and 3 min 6 s.
  • Speed model: H3 INT8 inference is partly bandwidth-bound and partly tensor-compute-bound, so each card’s effective speed is the geometric mean of memory bandwidth and dense INT8 tensor throughput (FP32 for Pascal, with a penalty for Turing’s weaker INT8 kernels). All bandwidth and compute figures come from NVIDIA specs. The calculator, per-card pages and workflow pages use exactly the same model. Its constants are fitted to all community benchmarks; error is about ±25%.
  • Apple M4 Max runs through MLX and is not comparable with ComfyUI numbers, so it is listed but not ranked.
Calculate for my setup →

About the prices

GPU prices swung hard in 2026 with memory costs; these are snapshots, so check the live listing before buying. In-production cards use a typical non-premium new price, discontinued cards use used-market prices, and modded cards carry no warranty.

Price sources (38)