Rows are hardware profiles
Each row names the card class and the environment that was actually used. The owned RTX 3060 row is a record, not a promise about every RTX 3060.
How to read
Read the matrix without guessing
Columns are task tracks
Each column combines a model or precision track with a task. A measured cell links to its test_id; it is a projection of that run record.
Untested stays explicit
not tested means this site has not run the combination. It does not mean “unsupported,” “impossible,” or “will fail.”
No cross-row ranking
The rows do not share one machine, workflow, or memory budget. We place facts beside one another and do not calculate a speed ranking across them.
Hardware × task
The measured matrix
Measured cells show the first formal run in the series and link to the full record below. Repeats remain separate records; the matrix is not a second set of measurements.
| GPU profile | REF2VA INT8 · R2V | FL2VA · T2V | FL2VA · I2V | Turbo LoRA |
|---|---|---|---|---|
| RTX 3060 12GB owned bench · GPU 0 |
2,581.8 s11,649 MiB · 94.8% 1344×768 · 124 frames |
2,175.5 s11,023 MiB · 89.7% 1344×768 · 124 frames |
2,373.7 s11,125 MiB · 90.5% 1344×768 · 124 frames |
528.6 smedian · 4-step Turbo 1344×768 · 124 frames |
| 8GB rental class not tested · no support verdict |
not testedRental-class run queued under Requirements §7.2 item 6. | not testedRental-class run queued under Requirements §7.2 item 6. | not testedRental-class run queued under Requirements §7.2 item 6. | not testedNo 8GB Turbo A/B; no support verdict. |
The 12GB row is site-measured on one owned bench. The 8GB row is a queued inventory item, not an inference from the 12GB measurements.
Full ledger · 26 records
Every retained site run
Formal runs, smoke checks, shared baselines, cache-node tests and failures live together here so a result cannot disappear merely because it was inconvenient. Each anchor is the stable test_id.
R2V · REF2VA INT8
These records use the site benchmark track and the official Ref2VA INT8 workflow family. The GPU 1 incident remains a first-class failure record; the two GPU 0 formal runs are independent repeats.
SITE-3060-R2V-B1-SMOKE
completedGPU 1 smoke check · 512×288 / 22 frames · comparison telemetry before the formal attempt.
55.3 swall time · ComfyUI11,578 MiBpeak VRAM · 94.2%43,154 MiBpeak system RAM
- Task / track
- R2V ·
site_benchmark - GPU / source
- RTX 3060 12GB · physical GPU 1 · owned bench
- Workload
- 512×288 · 22 frames · 24 fps · 20 steps
- Output
- Generated successfully · 0.9167 s · audio off
Full provenance
- Workflow
SITE-3060-R2V-B1-SMOKE-api.json; derived smoke workflow, hash retained in the pinned R2V baseline report.- Weights
- Official Ref2VA INT8 diffusion model; Qwen3-VL encoder; video VAE; audio VAE available but audio path off. Manifest recorded in the baseline report.
- Sampler / scheduler / denoise
res_multistep/simple/ 1.0- Seed / optimization
- Fixed seed in
submit.json; SageAttention absent; Turbo LoRA absent; cache none. - Host
- Intel Core i9-10850K · VM 16 vCPU · 47.05 GiB RAM · no swap · Ubuntu 24.04.4 LTS · driver 580.173.02 · CUDA 13.0
- Software
- Python 3.12.3 · PyTorch 2.13.0+cu130 · Triton installed · ComfyUI commit
bf4c9a08fc854df6d3b2bef1b92b509e2ef2d2c9· custom nodes N/A - Cache / contamination
- Cold after container restart · other GPU load unchanged at 104 MiB ·
contaminated=false - Temperature / notes
- GPU 1 peak 79 °C. This smoke is a monitoring and comparison run, not a formal ≥30-minute baseline.
SITE-3060-R2V-B1
failedGPU 1 formal attempt · retained hardware incident · VM froze after roughly nine minutes.
N/Awall time · no completionN/Apeak VRAM · no final sampleN/Apeak RAM · no final sample
- Task / track
- R2V ·
site_benchmark - GPU / source
- RTX 3060 12GB · physical GPU 1 · owned bench
- Workload
- 1344×768 · 124 frames · 24 fps · 20 steps · audio off
- Failure state
VM_FROZEN· no output · no valid wall time
Full provenance
- Workflow / weights
- Official R2V-derived workflow with Ref2VA INT8; the workflow and model manifest are pinned in the R2V baseline report.
- Sampler / scheduler / denoise
res_multistep/simple/ 1.0- Seed / optimization
- Fixed seed recorded in the run submission; SageAttention absent; Turbo LoRA absent; cache warm after smoke.
- Host
- Intel Core i9-10850K · VM 16 vCPU · 47.05 GiB RAM · no swap · Ubuntu 24.04.4 LTS · driver 580.173.02 · CUDA 13.0
- Software
- Python 3.12.3 · PyTorch 2.13.0+cu130 · ComfyUI commit
bf4c9a08fc854df6d3b2bef1b92b509e2ef2d2c9· custom nodes N/A - Last telemetry
- Peak temperature sample 87–88 °C; last VRAM samples 7.7–8.4 GiB; MemAvailable plateau approximately 4.9 GB. These are last samples, not final peaks.
- Failure note
- The guest stopped responding and PVE reported an internal error. No guest OOM log was found. The report records a GPU/PCIe hardware incident diagnosis; it does not turn the incident into a model-support verdict.
SITE-3060-R2V-GPU0-SMOKE
completedGPU 0 smoke comparison · 512×288 / 22 frames · same workload family as the GPU 1 smoke.
55.6 swall time · ComfyUI11,591 MiBpeak VRAM · 94.3%43,176 MiBpeak system RAM
- Task / track
- R2V ·
site_benchmark - GPU / source
- RTX 3060 12GB · physical GPU 0 · owned bench
- Workload
- 512×288 · 22 frames · 24 fps · 20 steps
- Output
- Generated successfully · 0.9167 s · audio off
Full provenance
- Workflow / weights
- R2V smoke derivative; official Ref2VA INT8 family; exact file hashes retained in the R2V baseline report.
- Sampler / scheduler / denoise
res_multistep/simple/ 1.0- Seed / optimization
- Fixed seed in the run submission; SageAttention absent; Turbo LoRA absent; cache none.
- Host / software
- Intel Core i9-10850K · 16 vCPU · 47.05 GiB RAM · no swap · Ubuntu 24.04.4 · driver 580.173.02 · CUDA 13.0 · Python 3.12.3 · PyTorch 2.13.0+cu130 · ComfyUI
bf4c9a08fc854df6d3b2bef1b92b509e2ef2d2c9 - Cache / contamination
- Cold after GPU binding change · other GPU load unchanged at 104 MiB ·
contaminated=false - Temperature / notes
- GPU 0 peak 60 °C. This is a smoke and hardware comparison record, not a formal baseline.
SITE-3060-R2V-GPU0-B1
completedGPU 0 formal run 1 · 1344×768 / 124 frames · the first formal site R2V record.
2,581.8 swall time · ComfyUI11,649 MiBpeak VRAM · 94.8%43,587 MiBpeak system RAM
- Task / track
- R2V ·
site_benchmark - GPU / source
- RTX 3060 12GB · physical GPU 0 · owned bench
- Workload
- 1344×768 · 124 frames · 24 fps · 20 steps · audio off
- Output
- 1344×768 · 124 frames · 5.166667 s · no audio ·
cc2a6f75…be6060f
Full provenance
- Workflow / SHA-256
SITE-3060-R2V-B1-api.json·f52a1b919c2fb29acfbe9affb7fa82ceef1ad2a57f2d68427383d25bc8257790- Weights
minimax_h3_ref2va_pruned_int8_convrot.safetensors· Ref2VA SHA-2569255f52b6677845ad238f20dfaafa94727053694127ab7f255c048f0f9365779; Qwen3-VL encoder and VAEs from the pinned manifest.- Sampler / scheduler / denoise
res_multistep/simple/ 1.0- Seed / optimization
- Fixed seed recorded in
submit.json; SageAttention absent; Turbo LoRA absent; cache none. - Host
- Intel Core i9-10850K · VM 16 vCPU · 47.05 GiB RAM · no swap · Ubuntu 24.04.4 LTS · driver 580.173.02 · CUDA 13.0
- Software
- Python 3.12.3 · PyTorch 2.13.0+cu130 · Triton installed · ComfyUI commit
bf4c9a08fc854df6d3b2bef1b92b509e2ef2d2c9· custom nodes N/A - Cache / contamination
- Warm after GPU 0 smoke · GPU 1 stayed at 104 MiB and GPU 0 resident load stayed at 104 MiB ·
contaminated=false - Repeat / notes
- Formal run ≥30 minutes; protocol requires an independent repeat. Repeat is SITE-3060-R2V-GPU0-B1-RUN2.
SITE-3060-R2V-GPU0-B1-RUN2
completedGPU 0 formal run 2 · independent repeat of SITE-3060-R2V-GPU0-B1.
2,579.8 swall time · ComfyUI11,649 MiBpeak VRAM · 94.8%43,907 MiBpeak system RAM
- Task / track
- R2V ·
site_benchmark - GPU / source
- RTX 3060 12GB · physical GPU 0 · owned bench
- Workload
- 1344×768 · 124 frames · 24 fps · 20 steps · audio off
- Output
- 1344×768 · 124 frames · 5.166667 s · no audio · same SHA-256
cc2a6f75…be6060f
Full provenance
- Workflow / SHA-256
SITE-3060-R2V-B1-api.json·f52a1b919c2fb29acfbe9affb7fa82ceef1ad2a57f2d68427383d25bc8257790- Weights
- Same pinned Ref2VA INT8, Qwen3-VL encoder and VAE manifest as run 1; no re-download or manifest change.
- Sampler / scheduler / denoise
res_multistep/simple/ 1.0- Seed / optimization
- Fixed seed
20260812; SageAttention absent; Turbo LoRA absent; cache none. - Host / software
- Same Intel Core i9-10850K · 16 vCPU · 47.05 GiB RAM · no swap · Ubuntu 24.04.4 · driver 580.173.02 · CUDA 13.0 · Python 3.12.3 · PyTorch 2.13.0+cu130 · ComfyUI
bf4c9a08fc854df6d3b2bef1b92b509e2ef2d2c9 - Cache / contamination
- Warm; GPU 1 stayed at 113 MiB and
contaminated=false. GPU 0 peak temperature 74 °C. - Repeat / notes
repeat_of=SITE-3060-R2V-GPU0-B1. The two raw values are intentionally shown separately, never collapsed into a single figure.
FL2VA INT8 · T2V and I2V
These six records use the pinned Comfy-Org template revision and the site benchmark track. Each formal task crossed the 30-minute threshold and therefore has two independent raw runs. The two smoke records verify the template, input and audio path; they are not formal baseline statistics.
SITE-3060-FL2VA-T2V-GPU0-SMOKE
completedT2V smoke · 512×288 / 22 frames · native audio on.
51.8 swall time · ComfyUI11,625 MiBpeak VRAM · 94.6%42,511 MiBpeak system RAM
- Task / track
- FL2VA T2V ·
site_benchmark - GPU / source
- RTX 3060 12GB · physical GPU 0 · owned bench
- Workload
- 512×288 · 22 frames · 24 fps · 20 steps · audio on
- Output
- Success; smoke output retained in the run archive
Full provenance
- Workflow
- T2V smoke derivative of the pinned Comfy-Org template; smoke hash retained in the FL2VA baseline report.
- Weights
- Official FL2VA INT8 ConvRot; Qwen3-VL-32B encoder; video VAE; audio VAE.
- Sampler / scheduler / denoise
res_multistep/simple/ 1.0 · fixed seed20260820- Host / software
- Intel Core i9-10850K · 16 vCPU · 47 GiB guest RAM · no swap · Ubuntu 24.04 · driver 580.173.02 · CUDA 13.0 · Python 3.12.3 · PyTorch 2.13.0+cu130 · ComfyUI 0.31.0, commit
bf4c9a08fc854df6d3b2bef1b92b509e2ef2d2c9 - Optimization / contamination
- Turbo LoRA off · cache none · SageAttention absent · Torch compile off · GPU 1 113→113 MiB ·
contaminated=false - Notes
- Smoke only; it does not enter the formal ≥30-minute baseline statistics.
SITE-3060-FL2VA-T2V-GPU0-B1
completedT2V formal run 1 · pinned official template · native audio on.
2,175.5 swall time · ComfyUI11,023 MiBpeak VRAM · 89.7%43,607 MiBpeak system RAM
- Task / track
- FL2VA T2V ·
site_benchmark - GPU / source
- RTX 3060 12GB · physical GPU 0 · owned bench
- Workload
- 1344×768 · 124 frames · 24 fps · 20 steps · audio on
- Output
- H.264 + AAC MP4 · 5.166667 s · SHA-256
3d582cbcf88b3aa8d60e93cea4c65bba445ab475c8c3243591107c7dab4ee802
Full provenance
- Workflow / SHA-256
- Official-derived T2V API workflow ·
1b083dfa1a4540b3805b998a66f5585f4337d5fbe14b441ca7ac0a2e1bf545a2 - Template revision
Comfy-Org/workflow_templates@7837633a31c1aa40495ed7f9171e9c97358a1d93- Weights
minimax_h3_fl2va_pruned_int8_convrot.safetensorsSHA-256e889202c41dafb67b10d67b97f0d8541508036a6090af23425a5c2615d03c47a; Qwen3-VL encoder and video/audio VAEs match the pinned manifest.- Sampler / scheduler / denoise
res_multistep/simple/ 1.0 · fixed seed20260820- Host / software
- Intel Core i9-10850K · 16 vCPU · 48,153 MiB VM RAM · no swap · Ubuntu 24.04 · driver 580.173.02 · CUDA 13.0 · Python 3.12.3 · PyTorch 2.13.0+cu130 · Triton 3.7.1 · ComfyUI 0.31.0, commit
bf4c9a08fc854df6d3b2bef1b92b509e2ef2d2c9 - Optimization / contamination
- Turbo LoRA off · cache none · SageAttention absent · Torch compile off · GPU 1 113→113 MiB ·
contaminated=false· cache state mixed - Repeat / notes
- Formal run ≥30 minutes; repeat is SITE-3060-FL2VA-T2V-GPU0-B1-RUN2. No cross-task causal claim is made from T2V versus I2V time.
SITE-3060-FL2VA-T2V-GPU0-B1-RUN2
completedT2V formal run 2 · independent repeat of SITE-3060-FL2VA-T2V-GPU0-B1.
2,176.2 swall time · ComfyUI10,863 MiBpeak VRAM · 88.4%43,880 MiBpeak system RAM
- Task / track
- FL2VA T2V ·
site_benchmark - GPU / source
- RTX 3060 12GB · physical GPU 0 · owned bench
- Workload
- 1344×768 · 124 frames · 24 fps · 20 steps · audio on
- Output
- Byte-identical to run 1 · SHA-256
3d582cbcf88b3aa8d60e93cea4c65bba445ab475c8c3243591107c7dab4ee802
Full provenance
- Workflow / SHA-256
- Same byte-frozen T2V API workflow ·
1b083dfa1a4540b3805b998a66f5585f4337d5fbe14b441ca7ac0a2e1bf545a2 - Weights / settings
- Same FL2VA INT8, Qwen3-VL, VAE files, template revision, prompt, seed
20260820, 1344×768, 124 frames, 20 steps, audio on. - Host / software
- Same RTX 3060 GPU0 · Intel Core i9-10850K · 16 vCPU · 48,153 MiB VM RAM · no swap · Ubuntu 24.04 · driver 580.173.02 · CUDA 13.0 · Python 3.12.3 · PyTorch 2.13.0+cu130 · ComfyUI
bf4c9a08fc854df6d3b2bef1b92b509e2ef2d2c9 - Optimization / contamination
- Turbo LoRA off · cache none · SageAttention absent · GPU 1 113→113 MiB ·
contaminated=false· cache state mixed - Repeat / notes
repeat_of=SITE-3060-FL2VA-T2V-GPU0-B1. The 0.7-second difference is shown as raw values, not collapsed into a mean.
SITE-3060-FL2VA-I2V-GPU0-SMOKE
completedI2V smoke · official input image · 512×288 / 22 frames · native audio on.
54.1 swall time · ComfyUI11,689 MiBpeak VRAM · 95.1%42,348 MiBpeak system RAM
- Task / track
- FL2VA I2V ·
site_benchmark - GPU / source
- RTX 3060 12GB · physical GPU 0 · owned bench
- Workload
- 512×288 · 22 frames · 24 fps · 20 steps · audio on
- Input / output
transparent_rgb_gaming_mouse.pnginput; smoke output retained
Full provenance
- Workflow
- I2V smoke derivative of the pinned Comfy-Org template; smoke hash retained in the FL2VA baseline report.
- Weights / input
- Same four FL2VA track weights as the formal run; input SHA-256
49696748d2fff0e8c9b63c7173c6d6282b70eac0195def5b39402f9564410e75. - Sampler / scheduler / seed
res_multistep/simple/ 1.0 · fixed seed20260820- Host / software
- Intel Core i9-10850K · 16 vCPU · 47 GiB guest RAM · no swap · Ubuntu 24.04 · driver 580.173.02 · CUDA 13.0 · Python 3.12.3 · PyTorch 2.13.0+cu130 · ComfyUI
bf4c9a08fc854df6d3b2bef1b92b509e2ef2d2c9 - Optimization / contamination
- Turbo LoRA off · cache none · SageAttention absent · GPU 1 113→113 MiB ·
contaminated=false - Notes
- Smoke only; not part of the formal I2V baseline statistic.
SITE-3060-FL2VA-I2V-GPU0-B1
completedI2V formal run 1 · official first-frame input · native audio on.
2,373.7 swall time · ComfyUI11,125 MiBpeak VRAM · 90.5%43,910 MiBpeak system RAM
- Task / track
- FL2VA I2V ·
site_benchmark - GPU / source
- RTX 3060 12GB · physical GPU 0 · owned bench
- Workload
- 1344×768 · 124 frames · 24 fps · 20 steps · audio on
- Output
- H.264 + AAC MP4 · 5.166667 s · SHA-256
8681605254c00719f7af488c13adb0124136de177c0dde02e906b198cc5cb47f
Full provenance
- Workflow / SHA-256
- Official-derived I2V API workflow ·
22648d0a8d42cae3207e9d5b0e9def8e55afd2ae4f7d6a2b809271141de30e40 - Template / input
Comfy-Org/workflow_templates@7837633a31c1aa40495ed7f9171e9c97358a1d93·transparent_rgb_gaming_mouse.png, SHA-25649696748d2fff0e8c9b63c7173c6d6282b70eac0195def5b39402f9564410e75- Weights / settings
- FL2VA INT8 ConvRot, Qwen3-VL encoder, video VAE and audio VAE;
res_multistep/simple/ 1.0; fixed seed20260820. - Host / software
- Intel Core i9-10850K · 16 vCPU · 48,153 MiB VM RAM · no swap · Ubuntu 24.04 · driver 580.173.02 · CUDA 13.0 · Python 3.12.3 · PyTorch 2.13.0+cu130 · Triton 3.7.1 · ComfyUI
bf4c9a08fc854df6d3b2bef1b92b509e2ef2d2c9 - Optimization / contamination
- Turbo LoRA off · cache none · SageAttention absent · Torch compile off · GPU 1 113→113 MiB ·
contaminated=false· cache state mixed - Repeat / notes
- Formal run ≥30 minutes; repeat is SITE-3060-FL2VA-I2V-GPU0-B1-RUN2. The T2V/I2V records are separate task cards, not a causal comparison.
SITE-3060-FL2VA-I2V-GPU0-B1-RUN2
completedI2V formal run 2 · independent repeat of SITE-3060-FL2VA-I2V-GPU0-B1.
2,378.1 swall time · ComfyUI11,637 MiBpeak VRAM · 94.7%43,907 MiBpeak system RAM
- Task / track
- FL2VA I2V ·
site_benchmark - GPU / source
- RTX 3060 12GB · physical GPU 0 · owned bench
- Workload
- 1344×768 · 124 frames · 24 fps · 20 steps · audio on
- Output
- Byte-identical to run 1 · SHA-256
8681605254c00719f7af488c13adb0124136de177c0dde02e906b198cc5cb47f
Full provenance
- Workflow / SHA-256
- Same byte-frozen I2V API workflow ·
22648d0a8d42cae3207e9d5b0e9def8e55afd2ae4f7d6a2b809271141de30e40 - Weights / input
- Same FL2VA INT8, Qwen3-VL, video/audio VAE files and official input image as run 1; fixed seed
20260820. - Host / software
- Same RTX 3060 GPU0 · Intel Core i9-10850K · 16 vCPU · 48,153 MiB VM RAM · no swap · Ubuntu 24.04 · driver 580.173.02 · CUDA 13.0 · Python 3.12.3 · PyTorch 2.13.0+cu130 · ComfyUI
bf4c9a08fc854df6d3b2bef1b92b509e2ef2d2c9 - Optimization / contamination
- Turbo LoRA off · cache none · SageAttention absent · GPU 1 113→113 MiB ·
contaminated=false· cache state mixed - Repeat / notes
repeat_of=SITE-3060-FL2VA-I2V-GPU0-B1. The 4.4-second difference is retained as a raw observation, not a mean.
FL2VA T2V · Turbo LoRA A/B
These five records freeze the FL2VA T2V baseline as the A side and add the third-party 4-step Turbo LoRA as the B side on the same RTX 3060 GPU 0. Only the LoRA, steps (20 → 4) and two shift schedules (video 6 / audio 3) changed; prompt, seed, canvas, frames and audio were held identical. B-side wall times were under 30 minutes, so the protocol's three-run branch applies and a median plus range is published. This is a site_benchmark pair on one 12GB card, not a cross-environment comparison and not a reproduction of any community claim.
SITE-3060-FL2VA-T2V-TURBO-GPU0-SMOKE
completedB-side smoke · same LoRA, 4 steps and shift 6/3 as the formal runs · 512×288 / 22 frames · not part of formal statistics.
32.5 swall time · ComfyUI11,625 MiBpeak VRAM43,645 MiBpeak system RAM
- Task / track
- FL2VA T2V Turbo smoke ·
site_benchmark - GPU / source
- RTX 3060 12GB · physical GPU 0 · owned bench
- Workload
- 512×288 · 22 frames · 4 steps · shift 6/3 · audio on
- Output
- Success; output SHA-256
d701cee5…7193· 134,730 bytes
Full provenance
- Workflow / SHA-256
- Smoke derivative of the formal B-side workflow ·
fbd544c2…bfded1; canvas and frames only changed from the formal file. - Weights / settings
- Same FL2VA INT8, Qwen3-VL, video/audio VAE files as A side; LoRA
minimax_h3_fl2v_turbo_4step_v1.0_768p_comfyui_bf16.safetensorsat strength 1.0 (site value; upstream gives no recommendation); 4 steps; shift video 6 / audio 3. - Host / software
- RTX 3060 GPU0 · Intel Core i9-10850K · 16 vCPU · 48,153 MiB VM RAM · no swap · Ubuntu 24.04.4 · driver 580.173.02 · CUDA 13.0 · Python 3.12.3 · PyTorch 2.13.0+cu130 · ComfyUI
bf4c9a08fc854df6d3b2bef1b92b509e2ef2d2c9 - Optimization / contamination
comfy-aimdo0.4.10 · cache node none · SageAttention absent · Torch compile off · GPU 1 113→113 MiB ·contaminated=false- Repeat / notes
- Smoke gate passed before the formal B-side runs; not included in the three-run formal statistics.
SITE-3060-FL2VA-T2V-TURBO-GPU0-B1
completedB-side formal run 1 of 3 · 4-step Turbo LoRA · same T2V workload as the A-side baseline.
526.7 swall time · ComfyUI11,185 MiBpeak VRAM43,596 MiBpeak system RAM
- Task / track
- FL2VA T2V Turbo ·
site_benchmark - GPU / source
- RTX 3060 12GB · physical GPU 0 · owned bench
- Workload
- 1344×768 · 124 frames · 24 fps · 4 steps · shift 6/3 · audio on
- Output
- Success; output SHA-256
e3f19c0d…60bc2· 2,650,811 bytes
Full provenance
- Workflow / SHA-256
- Formal B-side API workflow ·
735972ca…d10a; reused byte-identical for all three formal runs. - Weights / settings
- Same FL2VA INT8, Qwen3-VL, video/audio VAE files as A side; LoRA
minimax_h3_fl2v_turbo_4step_v1.0_768p_comfyui_bf16.safetensorsat strength 1.0 (site value; upstream gives no recommendation); 4 steps; shift video 6 / audio 3;res_multistep/simple; fixed seed20260820. - Host / software
- RTX 3060 GPU0 · Intel Core i9-10850K · 16 vCPU · 48,153 MiB VM RAM · no swap · Ubuntu 24.04.4 · driver 580.173.02 · CUDA 13.0 · Python 3.12.3 · PyTorch 2.13.0+cu130 · ComfyUI
bf4c9a08fc854df6d3b2bef1b92b509e2ef2d2c9 - Optimization / contamination
comfy-aimdo0.4.10 · cache node none · SageAttention absent · Torch compile off · GPU 1 113→113 MiB ·contaminated=false· cache state mixed- Repeat / notes
- Run 1 of the three-run formal branch; B-side median 528.6 s over range 526.7–528.9 s. A-side baseline wall times remain 2,175.5 s / 2,176.2 s; the same-environment pairing gives 4.113×–4.132×, published as a range with both sides' raw values, not as a mean.
SITE-3060-FL2VA-T2V-TURBO-GPU0-B1-RUN2
completedB-side formal run 2 of 3 · independent repeat of SITE-3060-FL2VA-T2V-TURBO-GPU0-B1.
528.9 swall time · ComfyUI10,479 MiBpeak VRAM43,893 MiBpeak system RAM
- Task / track
- FL2VA T2V Turbo ·
site_benchmark - GPU / source
- RTX 3060 12GB · physical GPU 0 · owned bench
- Workload
- 1344×768 · 124 frames · 24 fps · 4 steps · shift 6/3 · audio on
- Output
- Byte-identical to run 1 · SHA-256
e3f19c0d…60bc2· 2,650,811 bytes
Full provenance
- Workflow / SHA-256
- Same byte-frozen formal B-side API workflow ·
735972ca…d10a. - Weights / settings
- Same files, LoRA strength 1.0, 4 steps, shift 6/3 and seed
20260820as run 1. - Host / software
- Same RTX 3060 GPU0 · Intel Core i9-10850K · 16 vCPU · 48,153 MiB VM RAM · no swap · Ubuntu 24.04.4 · driver 580.173.02 · CUDA 13.0 · Python 3.12.3 · PyTorch 2.13.0+cu130 · ComfyUI
bf4c9a08fc854df6d3b2bef1b92b509e2ef2d2c9 - Optimization / contamination
comfy-aimdo0.4.10 · cache node none · SageAttention absent · Torch compile off · GPU 1 113→113 MiB ·contaminated=false· cache state mixed- Repeat / notes
- Run 2 of 3; raw values retained separately, no mean.
SITE-3060-FL2VA-T2V-TURBO-GPU0-B1-RUN3
failedRetained orchestration failure · never entered inference · FileNotFoundError on a host-path workflow before prompt submission.
- Task / track
- FL2VA T2V Turbo orchestration ·
site_benchmark - GPU / source
- RTX 3060 12GB · physical GPU 0 · owned bench
- Failure
- Client passed the host path
/opt/minimax-h3/data/user/bench/…, which does not exist inside the container;FileNotFoundError. Queue stayed empty; GPU0/GPU1 peaks 331/113 MiB; no inference started. - Correction
- Retried with the container-internal path
/data/user/bench/…as test_id B1-RUN3-RETRY1; workflow bytes and workload parameters unchanged.
Full provenance
- Failure summary
- Structured failure summary SHA-256
11e3bf6a…09a2749retained;node_errorsnot applicable because the prompt was never submitted. - Statistical treatment
- Failure directory and summary retained in the run archive; not counted in the three formal successful runs.
SITE-3060-FL2VA-T2V-TURBO-GPU0-B1-RUN3-RETRY1
completedB-side formal run 3 of 3 · retry of the retained RUN3 failure with the corrected container path.
528.6 swall time · ComfyUI10,895 MiBpeak VRAM43,706 MiBpeak system RAM
- Task / track
- FL2VA T2V Turbo ·
site_benchmark - GPU / source
- RTX 3060 12GB · physical GPU 0 · owned bench
- Workload
- 1344×768 · 124 frames · 24 fps · 4 steps · shift 6/3 · audio on
- Output
- Byte-identical to runs 1 and 2 · SHA-256
e3f19c0d…60bc2· 2,650,811 bytes
Full provenance
- Workflow / SHA-256
- Same byte-frozen formal B-side API workflow ·
735972ca…d10a; workload parameters identical to the failed RUN3. - Weights / settings
- Same files, LoRA strength 1.0, 4 steps, shift 6/3 and seed
20260820as runs 1 and 2. - Host / software
- Same RTX 3060 GPU0 · Intel Core i9-10850K · 16 vCPU · 48,153 MiB VM RAM · no swap · Ubuntu 24.04.4 · driver 580.173.02 · CUDA 13.0 · Python 3.12.3 · PyTorch 2.13.0+cu130 · ComfyUI
bf4c9a08fc854df6d3b2bef1b92b509e2ef2d2c9 - Optimization / contamination
comfy-aimdo0.4.10 · cache node none · SageAttention absent · Torch compile off · GPU 1 113→113 MiB ·contaminated=false· cache state mixed- Repeat / notes
- Run 3 of 3; median 528.6 s over range 526.7–528.9 s published per protocol.
Cache-node baseline and Phase B
These records share the 864×480 / 124-frame baseline family. The baseline has no cache node; Phase B records test individual third-party nodes. A node's author claim is not converted into a site claim, and the failed C5 run remains in the ledger.
BASE-864-A
completedShared cache-node baseline A · cache none · 864×480 / 124 frames.
596.9 swall time · ComfyUI11,433 MiBpeak VRAM43,488 MiBpeak system RAM
- Task / track
- R2V-derived cache baseline ·
site_benchmark - GPU / source
- RTX 3060 12GB · physical GPU 0 · owned bench
- Workload
- 864×480 · 124 frames · 24 fps · 20 steps · audio on
- Output
- Baseline MP4 retained; reference
mdatSHA-25662392754…85ddb
Full provenance
- Workflow / SHA-256
BASE-864-A-api.json·565dfdbf326bf226b701b6fbd27eb47fe0e17ce175d1537606252222d29f0676- Weights / settings
- Official FL2VA INT8 track, Qwen3-VL encoder, video/audio VAE ·
res_multistep/simple/ 1.0 · seed20260814· cache none. - Host / software
- RTX 3060 GPU0 · Intel Core i9-10850K · 16 vCPU · 47.05 GiB RAM · no swap · Ubuntu 24.04.4 · driver 580.173.02 · CUDA 13.0 · Python 3.12.3 · PyTorch 2.13.0+cu130 · ComfyUI
bf4c9a08fc854df6d3b2bef1b92b509e2ef2d2c9 - Contamination / notes
- GPU1 113→113 MiB ·
contaminated=false· reference run for Phase B speedup calculations; no cross-device ranking.
BASE-864-B
completedShared cache-node baseline B · independent baseline observation.
597.0 swall time · ComfyUI11,679 MiBpeak VRAM43,546 MiBpeak system RAM
- Task / track
- R2V-derived cache baseline ·
site_benchmark - GPU / source
- RTX 3060 12GB · physical GPU 0 · owned bench
- Workload
- 864×480 · 124 frames · 24 fps · 20 steps · audio on
- Output
- Baseline MP4 retained; reference
mdatSHA-25662392754…85ddb
Full provenance
- Workflow / SHA-256
- Same
BASE-864-A-api.json·565dfdbf326bf226b701b6fbd27eb47fe0e17ce175d1537606252222d29f0676 - Weights / settings
- Same official FL2VA INT8 track, seed
20260814, samplerres_multistep, schedulersimple, denoise 1.0, cache none. - Host / software
- Same RTX 3060 GPU0 · Intel Core i9-10850K · 16 vCPU · 47.05 GiB RAM · no swap · Ubuntu 24.04.4 · driver 580.173.02 · CUDA 13.0 · Python 3.12.3 · PyTorch 2.13.0+cu130 · ComfyUI
bf4c9a08fc854df6d3b2bef1b92b509e2ef2d2c9 - Contamination / notes
- GPU1 113→113 MiB ·
contaminated=false. Kept as a separate raw baseline observation; baselines are not collapsed into a single figure.
BASE-864-C-POSTCOMPOSE
completedPost-compose baseline check · verifies the deployment change did not alter the output payload.
597.2 swall time · ComfyUI11,305 MiBpeak VRAM43,626 MiBpeak system RAM
- Task / track
- R2V-derived cache baseline ·
site_benchmark - GPU / source
- RTX 3060 12GB · physical GPU 0 · owned bench
- Workload
- 864×480 · 124 frames · 24 fps · 20 steps · audio on
- Output
mdatbox SHA-25662392754f4b6e0ae232a68e1b86b375f20d7e4dfa681c06e1a04ccad6fd85ddb
Full provenance
- Workflow / SHA-256
BASE-864-A-api.json·565dfdbf326bf226b701b6fbd27eb47fe0e17ce175d1537606252222d29f0676- Weights / settings
- Same official FL2VA INT8 track, seed
20260814,res_multistep/simple/ 1.0, cache none. - Host / software
- Same RTX 3060 GPU0 · Intel Core i9-10850K · 16 vCPU · 47.05 GiB RAM · no swap · Ubuntu 24.04.4 · driver 580.173.02 · CUDA 13.0 · Python 3.12.3 · PyTorch 2.13.0+cu130 · ComfyUI
bf4c9a08fc854df6d3b2bef1b92b509e2ef2d2c9 - Contamination / notes
- GPU1 113→113 MiB ·
contaminated=false. Full mdat box matched the baseline; this record is not a new cache-node result.
PB-C1
completedTE-Speed-MiniMaxH3-OSS · completed with the required core patch inactive.
597.5 swall time · ComfyUI11,337 MiBpeak VRAM43,523 MiBpeak system RAM
- Task / track
- Cache Phase B ·
site_benchmark - Node / commit
- TE-Speed-MiniMaxH3-OSS ·
c1dacf47bc02cb9326f7b93c69280529b93d391b - Result
ACTIVE_NO_GAIN· 0.999× against the 597.0 s baseline; no acceleration conclusion- Output
- Generated;
mdatchanged from baseline · output SHA retained in the cache report
Full provenance
- Workflow / SHA-256
PHASEB-C1-api.json·95cb19d724fa65fe333792603f0e3bdb1cd6ffc7ede1ca52e066b71f2e4c9c5f- Weights / workload
- Shared 864×480 · 124 frames · 24 fps · 20 steps · seed
20260814· official FL2VA track · audio on. - Node state
- The required
patch_model.pywrite failed on the read-only ComfyUI core. The node log said it returned the model unpatched, so this is not evidence of a working speed patch. - Host / software
- RTX 3060 GPU0 · Intel Core i9-10850K · 16 vCPU · 47.05 GiB RAM · no swap · Ubuntu 24.04.4 · driver 580.173.02 · CUDA 13.0 · Python 3.12.3 · PyTorch 2.13.0+cu130 · ComfyUI
bf4c9a08fc854df6d3b2bef1b92b509e2ef2d2c9 - Contamination / notes
- GPU1 113→113 MiB ·
contaminated=false. C1 is both a compatibility and execution record; the patch failure is deliberately visible.
PB-C2
completedFirstBlockCache · active cache run.
363.2 swall time · ComfyUI11,601 MiBpeak VRAM43,474 MiBpeak system RAM1.644×site baseline comparison
- Task / track
- Cache Phase B ·
site_benchmark - Node / commit
- FirstBlockCache ·
725973c3bfd9de6dce249bc93dc5fe27f820df31 - Result
ACTIVE· 1.644× against the fixed 597.0 s comparison value- Output
- Generated; output and mdat hashes retained in the cache report
Full provenance
- Workflow / SHA-256
PHASEB-C2-api.json·e268ada3c0b55e18c500c41533da4c43335af1f46a7ec0fc521edaf40cccdf05- Weights / workload
- Shared 864×480 · 124 frames · 24 fps · 20 steps · seed
20260814· official FL2VA track · audio on. - Node state
- API node
ApplyMiniMaxH3FirstBlockCache; measured presetH3 Fast — 0.10 / max 2; console recorded cached 10/20 steps. - Host / software
- RTX 3060 GPU0 · Intel Core i9-10850K · 16 vCPU · 47.05 GiB RAM · no swap · Ubuntu 24.04.4 · driver 580.173.02 · CUDA 13.0 · Python 3.12.3 · PyTorch 2.13.0+cu130 · ComfyUI
bf4c9a08fc854df6d3b2bef1b92b509e2ef2d2c9 - Contamination / notes
- GPU1 113→113 MiB ·
contaminated=false. This is a site result under this workflow, not a claim about the repository author's advertised speed.
PB-C4-RUN1
completedTeaCache · independent run 1.
308.2 swall time · ComfyUI11,647 MiBpeak VRAM43,466 MiBpeak system RAM1.937×site baseline comparison
- Task / track
- Cache Phase B ·
site_benchmark - Node / commit
- TeaCache ·
4cbb50d69c73a19a5d6ec42c5aec1989d5a04b6f - Result
ACTIVE· 1.937× against the fixed 597.0 s comparison value- Output
- Generated; output SHA-256 and mdat SHA-256 retained; mdat differs from baseline
Full provenance
- Workflow / SHA-256
PHASEB-C4-RUN1-api.json·01a093b0858f085485d43158da9dd46e08a996e9dde93e2ea91f090e6b327bd0- Weights / workload
- Shared 864×480 · 124 frames · 24 fps · 20 steps · seed
20260814· official FL2VA track · audio on. - Node settings
rel_l1_thresh=0.15·start_step=2·end_step=-2·total_steps=20; API nodeMiniMaxH3TeaCache.- Host / software
- RTX 3060 GPU0 · Intel Core i9-10850K · 16 vCPU · 47.05 GiB RAM · no swap · Ubuntu 24.04.4 · driver 580.173.02 · CUDA 13.0 · Python 3.12.3 · PyTorch 2.13.0+cu130 · ComfyUI
bf4c9a08fc854df6d3b2bef1b92b509e2ef2d2c9 - Contamination / notes
- GPU1 113→113 MiB ·
contaminated=false. The result is a local site measurement, not a confirmation of a third-party claim.
PB-C4-RUN2
completedTeaCache · independent run 2; raw repeat shown beside run 1.
308.6 swall time · ComfyUI11,647 MiBpeak VRAM43,557 MiBpeak system RAM1.935×site baseline comparison
- Task / track
- Cache Phase B ·
site_benchmark - Node / commit
- TeaCache ·
4cbb50d69c73a19a5d6ec42c5aec1989d5a04b6f - Result
ACTIVE· 1.935× against the fixed 597.0 s comparison value- Output
- Byte-identical to PB-C4-RUN1 · same mdat SHA-256
Full provenance
- Workflow / SHA-256
- Same
PHASEB-C4-RUN1-api.json·01a093b0858f085485d43158da9dd46e08a996e9dde93e2ea91f090e6b327bd0 - Weights / settings
- Same TeaCache settings, FL2VA track, 864×480 workload, seed
20260814, 20 steps and audio path as PB-C4-RUN1. - Host / software
- Same RTX 3060 GPU0 · Intel Core i9-10850K · 16 vCPU · 47.05 GiB RAM · no swap · Ubuntu 24.04.4 · driver 580.173.02 · CUDA 13.0 · Python 3.12.3 · PyTorch 2.13.0+cu130 · ComfyUI
bf4c9a08fc854df6d3b2bef1b92b509e2ef2d2c9 - Contamination / notes
- GPU1 113→113 MiB ·
contaminated=false·repeat_of=PB-C4-RUN1. The page retains 1.937× and 1.935× as two separate observations.
PB-C5
failedcomfyui-speed-minimaxH3 · compatibility failure on ComfyUI 0.31.0.
50.0 sclient wall · failed execution11,163 MiBlast recorded peak VRAM41,338 MiBlast recorded peak RAM
- Task / track
- Cache Phase B ·
site_benchmark - Node / commit
- comfyui-speed-minimaxH3 ·
2f507d687cd6767212ae003d272042bf886a3cd3 - Result
FAILED· no successful ComfyUI wall time · no output- Failure
AttributeError: ... time_shift_sloperuntime hook missing
Full provenance
- Workflow / SHA-256
PHASEB-C5-api.json·b17a050fd4f6d4459f7d0c2b4be50d7feb53beb0d3ef40685fd4d85cc1fd766c- Weights / workload
- Shared 864×480 · 124 frames · 24 fps · 20 steps · seed
20260814· official FL2VA track · audio on. - Failure log
MiniMax H3 Speed Cache ... estimated 1.00x; thenAttributeError: module 'comfy.ldm.minimax.model' has no attribute 'time_shift_slope'. No core source was patched and no retry altered the environment.- Host / software
- RTX 3060 GPU0 · Intel Core i9-10850K · 16 vCPU · 47.05 GiB RAM · no swap · Ubuntu 24.04.4 · driver 580.173.02 · CUDA 13.0 · Python 3.12.3 · PyTorch 2.13.0+cu130 · ComfyUI 0.31.0, commit
bf4c9a08fc854df6d3b2bef1b92b509e2ef2d2c9 - Contamination / notes
- GPU1 113→113 MiB ·
contaminated=false. Client telemetry is retained, but no speedup or output hash is computed from a failed execution.
PB-C6
completedFBcache-shendumao · cache profile measured without SageAttention or Sol-Attn.
447.0 swall time · ComfyUI11,773 MiBpeak VRAM42,060 MiBpeak system RAM1.336×site baseline comparison
- Task / track
- Cache Phase B ·
site_benchmark - Node / commit
- FBcache-shendumao ·
a19b53c4507192a9cb164263335b238a454c9690 - Result
ACTIVE· 1.336× against the fixed 597.0 s comparison value- Output
- Generated; output and mdat hashes retained in the cache report
Full provenance
- Workflow / SHA-256
PHASEB-C6-api.json·6afd00e8095eee95c8749f0c5718b159f0d22456db329c0255a8d3a9cc7bd893- Weights / workload
- Shared 864×480 · 124 frames · 24 fps · 20 steps · seed
20260814· official FL2VA track · audio on. - Node state
- API node
H3CombinedAcceleration; SageDisabled, cache profileBalanced, Sol-AttnDisabled; console recorded 14 full steps and 6 cache skips. - Host / software
- RTX 3060 GPU0 · Intel Core i9-10850K · 16 vCPU · 47.05 GiB RAM · no swap · Ubuntu 24.04.4 · driver 580.173.02 · CUDA 13.0 · Python 3.12.3 · PyTorch 2.13.0+cu130 · ComfyUI
bf4c9a08fc854df6d3b2bef1b92b509e2ef2d2c9 - Contamination / notes
- GPU1 113→113 MiB ·
contaminated=false. Local result only; no author-number comparison is calculated.
PB-C7
completedAdaptiveCache · balanced preset · no Turbo LoRA or Sol-Attn.
377.7 swall time · ComfyUI11,201 MiBpeak VRAM42,642 MiBpeak system RAM1.581×site baseline comparison
- Task / track
- Cache Phase B ·
site_benchmark - Node / commit
- AdaptiveCache ·
dc35065676617055aa90c4acd8b926cb265873de - Result
ACTIVE· 1.581× against the fixed 597.0 s comparison value- Output
- Generated; output and mdat hashes retained in the cache report
Full provenance
- Workflow / SHA-256
PHASEB-C7-api.json·42a8a8495823d54f6f1ad8bd9ddd60ce881af0e3a97106560fe62621b4f4f3bf- Weights / workload
- Shared 864×480 · 124 frames · 24 fps · 20 steps · seed
20260814· official FL2VA track · audio on. - Node state
- API node
MiniMaxH3AdaptiveCache; measured presetbalanced;cache_device=auto; advanced node not used. - Host / software
- RTX 3060 GPU0 · Intel Core i9-10850K · 16 vCPU · 47.05 GiB RAM · no swap · Ubuntu 24.04.4 · driver 580.173.02 · CUDA 13.0 · Python 3.12.3 · PyTorch 2.13.0+cu130 · ComfyUI
bf4c9a08fc854df6d3b2bef1b92b509e2ef2d2c9 - Contamination / notes
- GPU1 113→113 MiB ·
contaminated=false. Local result only; no cross-GPU ranking.
Cache-node compatibility · checked 2026-08-14
Installation and loading are separate from performance
C1–C8 are compatibility-screening objects, not eight invented performance records. Where a Phase B run exists, its test_id links to the ledger. “Not measured” means no performance run was generated.
| Node | Commit | Install | Load / visible | Phase B record | Observed note |
|---|---|---|---|---|---|
| C1 · TE-Speed-MiniMaxH3-OSS | c1dacf47…d391b | Partial | Yes / yes | PB-C1 | Core patch failed on read-only filesystem; completed run returned to stock speed. |
| C2 · FirstBlockCache | 725973c3…0df31 | Yes | Yes / yes | PB-C2 | API node visible; active result recorded. |
| C3 · MiniMaxH3-Cache | 8a45e096…73819 | Yes | Yes / yes | Not measured | Dynamic in-memory patch observed; README/documentation boundary kept it out of Phase B. |
| C4 · TeaCache | 4cbb50d6…04b6f | Yes | Yes / yes | PB-C4-RUN1 + RUN2 | Two raw active results; 1.937× and 1.935× are both retained. |
| C5 · comfyui-speed-minimaxH3 | 2f507d68…a3cd3 | Yes | Yes / yes | PB-C5 failed | Runtime hook raised missing time_shift_slope on ComfyUI 0.31.0. |
| C6 · FBcache-shendumao | a19b53c4…c9690 | Yes | Yes / yes | PB-C6 | Balanced cache profile; Sage and Sol-Attn disabled. |
| C7 · AdaptiveCache | dc350656…873de | Yes | Yes / yes | PB-C7 | Balanced preset; advanced node not used. |
| C8 · TE-Speed-MiniMaxH3 | e8b6549b…61b135 | Yes | No / no | Not measured | ModuleNotFoundError for the repository's Windows binary on this Linux environment. |
Compatibility source: cache-node benchmark report. Compatibility labels describe this checked environment; they are not universal support claims.
What these runs establish
Three facts from this ledger
Each statement is tied to records above and carries its conditions. None is a cross-GPU ranking or a community-result verdict.
Workload scaling did not move VRAM much
In the R2V GPU 0 smoke-to-formal pair, the frame-pixel workload is approximately 40×. Peak VRAM moved from 11,591 to 11,649 MiB (+0.5%), while peak RAM moved from 43,176 to 43,587 MiB (+1.0%). The smoke was cold and the formal run warm, so this is an observation about this pair, not a scaling law.
Cache Phase B telemetry stayed in a narrow VRAM band
The seven Phase B execution records, including the failed C5 telemetry record, span 11,163–11,773 MiB of recorded peak VRAM. C5 has no successful output, and C3/C8 have no Phase B performance record; the range is not evidence that all eight nodes work.
R2V repeated output bytes matched
The two formal GPU 0 R2V runs each produced the same output SHA-256 prefix recorded in the cards: cc2a6f75…be6060f. Under this fixed seed, workflow and hardware setup, the outputs were byte-identical. The wall times remain 2,581.8 s and 2,579.8 s as separate observations.
Protocol in one screen
What makes a row a site benchmark
- Track: every row here is
site_benchmark, not a community reproduction. - Freeze: record the workflow revision and SHA-256, model files, input files, prompt, seed, canvas, frames, sampler, scheduler and audio state.
- Measure: report wall time, peak VRAM in MiB, peak system RAM, temperature, GPU binding and contamination before/after.
- Repeat: a formal site benchmark lasting at least 30 minutes runs twice; both test IDs and raw values remain visible. No average stands in for them.
- Retain failures: a failed execution or hardware incident is a record, not a footnote to delete.
- Separate tracks: smoke checks validate a chain; cache-node compatibility is not performance; rental hardware gets its own row; GPU 0 was used for the owned-bench heavy runs after the GPU 1 incident.
Full protocol: site-reproduction-protocol.md, version 0.2-draft.
Boundaries
What is not here
No second-hand numbers in the matrix
Community reports remain community-reported. Their setup, timing and claims are not inserted into a site-measured cell.
No cross-card speed ranking
Different cards, environments and workflows cannot be reduced to one leaderboard. The 8GB row is not a support or failure verdict.
Turbo A/B stays on this 12GB card
The Turbo LoRA A/B ran once on this owned RTX 3060 12GB bench with the same prompt, seed, canvas, frames and workflow as the FL2VA T2V baseline, changing only the LoRA, steps and shift schedules. The same-environment pairing gives a 4.113×–4.132× wall-time range; it is not a cross-environment comparison and it is not placed beside any community "5×" claim. All three B-side runs peaked above 8,192 MiB, so this page makes no 8GB support or failure claim.
No averaged formal result
Repeat values are separate elements with separate IDs. If a future run changes, the raw ledger can show what changed.
FAQ
GPU benchmark questions
What does “not tested” mean in the GPU matrix?
It means MiniMax H3 Tutorial has not run that GPU, precision and task combination. It is not a statement that the combination is unsupported or impossible.
Why does the GPU benchmark page show two raw values?
The site protocol requires an independent repeat for a formal benchmark lasting at least 30 minutes. Both test IDs and both original values stay visible; this page does not replace them with an average.
Did this site reproduce the community RTX 3060 timing claim?
No. These records use the site_benchmark track and their own pinned workflows, settings and hardware conditions. They are not presented as a reproduction, validation or confirmation of a community result.
Can the RTX 3060 results be applied to an 8GB GPU?
No conclusion is made here. The 8GB rental-class row is explicitly not tested, and a result on the owned RTX 3060 12GB bench is not an extrapolation to another card, memory size, precision or task.
What did the site measure for the Turbo LoRA A/B?
On one owned RTX 3060 12GB bench, the same FL2VA T2V workload with the 4-step Turbo LoRA took a median 528.6 seconds (range 526.7 to 528.9 seconds) against an A-side baseline of 2,175.5 and 2,176.2 seconds — a same-environment 4.113× to 4.132× wall-time range, not an average. All three Turbo runs peaked above 8,192 MiB VRAM, so the page makes no 8GB support or failure claim.
Next checks