MiniMax video benchmarks on AMD GPUs

24 reported configurations · 4 sources · 76 measurements

Documented results

24 selected configurations

Selected across sources, not a performance ranking. Match model weights, resolution, steps, software and timing scope before comparing. Playback FPS describes the output video.

Source	Model	GPUs	Output	Measurements
View ReportComfyUI	MiniMax H3 FL2VAINT8 ConvRotModel reference ↗Text → video	Radeon AI PRO R97001 GPU Configuration Engine ComfyUI 0.30.0 Backend ROCm / HIP OS Linux Parallelization Single GPU Memory 32 GB per GPU Reported 2026-09-08 Full record →	864 × 480Batch 1124 frames20 stepsBatch 1	262.34 sGeneration time · reportedCold startup and model loading excluded. ≈ 240.6 sSampling time · reportedSource says about; component sampling time. Whole workflow · Warm
View ReportPaiton	MiniMax H3 Ref2VA Base50W4A8Tested weights ↗Image → video	Radeon AI PRO R97001 GPU Configuration Engine Paiton Paiton 1.1.0 Backend ROCm / HIP OS Linux Parallelization Single GPU Memory 32 GB per GPU Reported 2026-10-19 Full record →	864 × 480≈ 5.17 s clip · Batch 1124 frames24 FPS playback50 stepsBatch 1	393.7 sGeneration time · reportedComplete playable MP4 including conditioning, H.264 encode and AAC mux; model download/build/import separate. 360 sSampling time · reportedSampling component; already included in workflow time. 7.2 sStep time · reported 30.55 GiBPeak VRAM · reportedSampled driver peak; not Torch allocation; sampling may miss short peaks. Whole workflow · Timing details missing
View ReportComfyUI	MiniMax H3 FL2VAFP8 scaledTested weights ↗Text → video	Radeon AI PRO R97001 GPU Configuration Engine ComfyUI Backend ROCm / HIP OS Not reported Parallelization Single GPU Memory 32 GB per GPU Reported 2026-08-18 Full record →	864 × 480≈ 7.29 s clip · Batch 1175 frames24 FPS playback20 stepsBatch 1	465.94 sGeneration time · reportedComfyUI /history execution_start to execution_success; not the separate server_reported_execution_s value. Whole workflow · Timing details missing
View ReportComfyUI	MiniMax H3 FL2VAINT8 ConvRotModel reference ↗Image → video	Radeon RX 7900 XT1 GPU Configuration Engine ComfyUI Backend ROCm / HIP OS Ubuntu 26.04 Parallelization Single GPU Memory 20 GB per GPU Reported Not reported Full record →	544 × 800Batch 1Frame count not reported20 stepsBatch 1	426.3 sGeneration time · reported 330.8 sSampling time · reported 16.54 sStep time · reported Whole workflow · Warm
View ReportComfyUI	MiniMax H3 FL2VAINT8 ConvRotModel reference ↗Text → video	Radeon AI PRO R97002 GPUs Configuration Engine ComfyUI 0.30.0 Backend ROCm / HIP OS Linux Parallelization sequence parallel Memory 32 GB per GPU Reported 2026-09-08 Full record →	864 × 480Batch 1124 frames20 stepsBatch 1	149.16 sGeneration time · medianCold startup and model loading excluded. ≈ 129 sSampling time · reportedSource says about; component sampling time. Whole workflow · Warm · median
View ReportPaiton	MiniMax H3 Ref2VA Base50W4A8Tested weights ↗Image → video	Radeon AI PRO R97001 GPU Configuration Engine Paiton Paiton 1.1.0 Backend ROCm / HIP OS Linux Parallelization Single GPU Memory 32 GB per GPU Reported 2026-10-19 Full record →	864 × 480≈ 5.17 s clip · Batch 1124 frames24 FPS playback50 stepsBatch 1	417.2 sGeneration time · reportedComplete playable MP4 including conditioning, H.264 encode and AAC mux; model download/build/import separate. 388.7 sSampling time · reportedSampling component; already included in workflow time. 7.77 sStep time · reported 30.56 GiBPeak VRAM · reportedSampled driver peak; not Torch allocation; sampling may miss short peaks. Whole workflow · Timing details missing
View ReportComfyUI	MiniMax H3 FL2VAFP8 scaledTested weights ↗Image → video	Radeon AI PRO R97001 GPU Configuration Engine ComfyUI Backend ROCm / HIP OS Not reported Parallelization Single GPU Memory 32 GB per GPU Reported 2026-08-18 Full record →	864 × 480≈ 7.29 s clip · Batch 1175 frames24 FPS playback20 stepsBatch 1	513.89 sGeneration time · reportedComfyUI /history execution_start to execution_success; not the separate server_reported_execution_s value. Whole workflow · Timing details missing
View ReportComfyUI	MiniMax H3 FL2VAINT8 ConvRotModel reference ↗Image → video	Radeon RX 7900 XT1 GPU Configuration Engine ComfyUI Backend ROCm / HIP OS Ubuntu 26.04 Parallelization Single GPU Memory 20 GB per GPU Reported Not reported Full record →	544 × 800Batch 1Frame count not reported20 stepsBatch 1	657 sGeneration time · reported 561.2 sSampling time · reported 28.06 sStep time · reported Whole workflow · Warm
View ReportPaiton	MiniMax H3 Ref2VA Base50W4A8Tested weights ↗Image → video	Radeon AI PRO R97001 GPU Configuration Engine Paiton Paiton 1.1.0 Backend ROCm / HIP OS Linux Parallelization Single GPU Memory 32 GB per GPU Reported 2026-10-19 Full record →	864 × 480≈ 5.17 s clip · Batch 1124 frames24 FPS playback50 stepsBatch 1	595.2 sGeneration time · reportedComplete playable MP4 including conditioning, H.264 encode and AAC mux; model download/build/import separate. 565.6 sSampling time · reportedSampling component; already included in workflow time. 11.31 sStep time · reported 30.57 GiBPeak VRAM · reportedSampled driver peak; not Torch allocation; sampling may miss short peaks. Whole workflow · Timing details missing
View ReportComfyUI	MiniMax H3 Ref2VAFP8 scaledTested weights ↗Image → video	Radeon AI PRO R97001 GPU Configuration Engine ComfyUI Backend ROCm / HIP OS Not reported Parallelization Single GPU Memory 32 GB per GPU Reported 2026-08-18 Full record →	864 × 480≈ 7.29 s clip · Batch 1175 frames24 FPS playback20 stepsBatch 1	561.8 sGeneration time · reportedComfyUI /history execution_start to execution_success; not the separate server_reported_execution_s value. Whole workflow · Timing details missing
View ReportComfyUI	MiniMax H3 FL2VAINT8 ConvRotModel reference ↗Image → video	Radeon RX 7900 XT1 GPU Configuration Engine ComfyUI Backend ROCm / HIP OS Ubuntu 26.04 Parallelization Single GPU Memory 20 GB per GPU Reported Not reported Full record →	544 × 800Batch 1Frame count not reported20 stepsBatch 1	449.9 sGeneration time · reported 351.4 sSampling time · reported 17.57 sStep time · reported Whole workflow · Warm
View ReportPaiton	MiniMax H3 Ref2VA Turbo4W4A8Tested weights ↗Image → video	Radeon AI PRO R97001 GPU Configuration Engine Paiton Paiton 1.1.0 Backend ROCm / HIP OS Linux Parallelization Single GPU Memory 32 GB per GPU Reported 2026-10-19 Full record →	864 × 480≈ 5.17 s clip · Batch 1124 frames24 FPS playback4 stepsBatch 1	50.6 sGeneration time · reportedComplete playable MP4 including conditioning, H.264 encode and AAC mux; model download/build/import separate. 22 sSampling time · reportedSampling component; already included in workflow time. 5.5 sStep time · reported 30.6 GiBPeak VRAM · reportedSampled driver peak; not Torch allocation; sampling may miss short peaks. Whole workflow · Timing details missing
View ReportComfyUI	MiniMax H3 FL2VAINT8 ConvRotModel reference ↗Image → video	Radeon AI PRO R97001 GPU Configuration Engine ComfyUI Backend ROCm / HIP OS Ubuntu 26.04 Parallelization Single GPU Memory 32 GB per GPU Reported Not reported Full record →	544 × 800Batch 1Frame count not reported20 stepsBatch 1	293.6 sGeneration time · reported 269.6 sSampling time · reported 13.48 sStep time · reported Whole workflow · Warm
View ReportPaiton	MiniMax H3 Ref2VA Turbo4W4A8Tested weights ↗Image → video	Radeon AI PRO R97001 GPU Configuration Engine Paiton Paiton 1.1.0 Backend ROCm / HIP OS Linux Parallelization Single GPU Memory 32 GB per GPU Reported 2026-10-19 Full record →	864 × 480≈ 5.17 s clip · Batch 1124 frames24 FPS playback4 stepsBatch 1	85.3 sGeneration time · reportedComplete playable MP4 including conditioning, H.264 encode and AAC mux; model download/build/import separate. 25.2 sSampling time · reportedSampling component; already included in workflow time. 6.31 sStep time · reported 30.84 GiBPeak VRAM · reportedSampled driver peak; not Torch allocation; sampling may miss short peaks. Whole workflow · Timing details missing
View ReportComfyUI	MiniMax H3 FL2VAINT8 ConvRotModel reference ↗Image → video	Radeon AI PRO R97001 GPU Configuration Engine ComfyUI Backend ROCm / HIP OS Ubuntu 26.04 Parallelization Single GPU Memory 32 GB per GPU Reported Not reported Full record →	544 × 800Batch 1Frame count not reported20 stepsBatch 1	291.2 sGeneration time · reported 268.8 sSampling time · reported 13.44 sStep time · reported Whole workflow · Warm
View ReportPaiton	MiniMax H3 Ref2VA Turbo4W4A8Tested weights ↗Image → video	Radeon AI PRO R97001 GPU Configuration Engine Paiton Paiton 1.1.0 Backend ROCm / HIP OS Linux Parallelization Single GPU Memory 32 GB per GPU Reported 2026-10-19 Full record →	864 × 480≈ 5.17 s clip · Batch 1124 frames24 FPS playback4 stepsBatch 1	178.7 sGeneration time · reportedComplete playable MP4 including conditioning, H.264 encode and AAC mux; model download/build/import separate. 45.3 sSampling time · reportedSampling component; already included in workflow time. 11.32 sStep time · reported 30.6 GiBPeak VRAM · reportedSampled driver peak; not Torch allocation; sampling may miss short peaks. Whole workflow · Timing details missing
View ReportComfyUI	MiniMax H3 FL2VAINT8 ConvRotModel reference ↗Image → video	Radeon AI PRO R97001 GPU Configuration Engine ComfyUI Backend ROCm / HIP OS Ubuntu 26.04 Parallelization Single GPU Memory 32 GB per GPU Reported Not reported Full record →	544 × 800Batch 1Frame count not reported20 stepsBatch 1	224.5 sGeneration time · reported 202.8 sSampling time · reported 10.14 sStep time · reported Whole workflow · Warm
View ReportPaiton	MiniMax H3 Ref2VA Turbo8W4A8Tested weights ↗Image → video	Radeon AI PRO R97001 GPU Configuration Engine Paiton Paiton 1.1.0 Backend ROCm / HIP OS Linux Parallelization Single GPU Memory 32 GB per GPU Reported 2026-10-19 Full record →	864 × 480≈ 5.17 s clip · Batch 1124 frames24 FPS playback8 stepsBatch 1	109.2 sGeneration time · reportedComplete playable MP4 including conditioning, H.264 encode and AAC mux; model download/build/import separate. 56.9 sSampling time · reportedSampling component; already included in workflow time. 7.11 sStep time · reported 30.58 GiBPeak VRAM · reportedSampled driver peak; not Torch allocation; sampling may miss short peaks. Whole workflow · Timing details missing
View ReportComfyUI	MiniMax H3 FL2VAINT8 ConvRotModel reference ↗Image → video	Radeon AI PRO R97001 GPU Configuration Engine ComfyUI Backend ROCm / HIP OS Ubuntu 26.04 Parallelization Single GPU Memory 32 GB per GPU Reported Not reported Full record →	544 × 800Batch 1Frame count not reported20 stepsBatch 1	222.4 sGeneration time · reported 201.2 sSampling time · reported 10.06 sStep time · reported Whole workflow · Warm
View ReportPaiton	MiniMax H3 Ref2VA Turbo8W4A8Tested weights ↗Image → video	Radeon AI PRO R97001 GPU Configuration Engine Paiton Paiton 1.1.0 Backend ROCm / HIP OS Linux Parallelization Single GPU Memory 32 GB per GPU Reported 2026-10-19 Full record →	864 × 480≈ 5.17 s clip · Batch 1124 frames24 FPS playback8 stepsBatch 1	119.2 sGeneration time · reportedComplete playable MP4 including conditioning, H.264 encode and AAC mux; model download/build/import separate. 58.1 sSampling time · reportedSampling component; already included in workflow time. 7.26 sStep time · reported 30.66 GiBPeak VRAM · reportedSampled driver peak; not Torch allocation; sampling may miss short peaks. Whole workflow · Timing details missing
View ReportPaiton	MiniMax H3 Ref2VA Turbo8W4A8Tested weights ↗Image → video	Radeon AI PRO R97001 GPU Configuration Engine Paiton Paiton 1.1.0 Backend ROCm / HIP OS Linux Parallelization Single GPU Memory 32 GB per GPU Reported 2026-10-19 Full record →	864 × 480≈ 5.17 s clip · Batch 1124 frames24 FPS playback8 stepsBatch 1	221.5 sGeneration time · reportedComplete playable MP4 including conditioning, H.264 encode and AAC mux; model download/build/import separate. 95.7 sSampling time · reportedSampling component; already included in workflow time. 11.96 sStep time · reported 30.67 GiBPeak VRAM · reportedSampled driver peak; not Torch allocation; sampling may miss short peaks. Whole workflow · Timing details missing
View ReportPaiton	MiniMax H3 Ref2VA Turbo4W4A8Tested weights ↗Image → video	Radeon AI PRO R97002 GPUs Configuration Engine Paiton Paiton 1.1.0 Backend ROCm / HIP OS Linux Parallelization Independent replicas Memory 32 GB per GPU Reported 2026-10-19 Full record →	864 × 480≈ 5.17 s clip · Batch 1124 frames24 FPS playback4 stepsBatch 1	51.7 sGeneration time · p50Per-request latency within this queue campaign; each request uses only one GPU. 52.52 sGeneration time · p95Source percentile over eight total requests; small sample. 0.038 clips/sClip rate · reportedDerived unit conversion from source requests/hour; aggregate for two independent workers. 208.16 scampaign makespan s · reportedEight jobs across two independent workers; not per-video latency. Whole workflow · Warm · p50
View ReportPaiton	MiniMax H3 Ref2VA Base50W4A8Tested weights ↗Image → video	Radeon AI PRO R97001 GPU Configuration Engine Paiton Paiton 1.1.0 Backend ROCm / HIP OS Linux Parallelization Single GPU Memory 32 GB per GPU Reported 2026-10-19 Full record →	1344 × 768Batch 1Frame count not reported24 FPS playback50 stepsBatch 1	1,753.15 sGeneration time · reportedComplete playable MP4 including conditioning, H.264 encode and AAC mux; model download/build/import separate. 1,602.6 sSampling time · reportedSampling component; already included in workflow time. 32.05 sStep time · reported 30.39 GiBPeak VRAM · reportedSampled driver peak; not Torch allocation; sampling may miss short peaks. Whole workflow · Warm
View ReportPaiton	MiniMax H3 Ref2VA Turbo4W4A8Tested weights ↗Image → video	Radeon AI PRO R97001 GPU Configuration Engine Paiton Paiton 1.1.0 Backend ROCm / HIP OS Linux Parallelization Single GPU Memory 32 GB per GPU Reported 2026-10-19 Full record →	1344 × 768Batch 1Frame count not reported24 FPS playback4 stepsBatch 1	258.43 sGeneration time · reportedComplete playable MP4 including conditioning, H.264 encode and AAC mux; model download/build/import separate. 118 sSampling time · reportedSampling component; already included in workflow time. 29.5 sStep time · reported 30.46 GiBPeak VRAM · reportedSampled driver peak; not Torch allocation; sampling may miss short peaks. Whole workflow · Warm

Explore all 24 configurations →

All video benchmarks · Methodology and source review