Hugging Face's logo

Datasets:
FlowForgeLabAi
/
minimax-h3-8gb-bench

Tags:
video-generation
minimax-h3
comfyui
low-vram
benchmark
windows
License:
Dataset card Data Studio Files
xet
 Community
Dataset Preview
Duplicate
The full dataset viewer is not available (click to read why). Only showing a preview of the rows.
The dataset generation failed because of a cast error
Error code:   DatasetGenerationCastError
Exception:    DatasetGenerationCastError
Message:      An error occurred while generating the dataset

All the data files must have the same columns, but at some point there are 33 new columns ({'chunkff', 'model', 'process_crash', 'lora', 'peak_vram_mib', 'lora_strength', 'label', 'sec_post', 'quality_note', 'wall_sec', 'prompt_id', 'sec_total', 'date', 'seed', 'sec_load', 'attn_backend', 'peak_power_w', 'peak_pagefile_gb', 'peak_ram_used_gb', 'sec_denoise', 'peak_temp_c', 'blocks_logged', 'seconds_per_block', 'steps', 'completed', 'crashed', 'opt_nodes', 'scheduler', 'tespeed', 'shift', 'sampler', 'text_encoder', 'head_chunks'}) and 13 missing columns ({'ram_total_gb', 'mem_clock_mhz', 'temp_c', 'sm_clock_mhz', 'timestamp', 'pagefile_avail_gb', 'vram_used_mib', 'vram_total_mib', 'gpu_util_pct', 'gpu_mem_util_pct', 'power_w', 'ram_load_pct', 'ram_avail_gb'}).

This happened while the csv dataset builder was generating data using

hf://datasets/FlowForgeLabAi/minimax-h3-8gb-bench/results.csv (at revision 210fe4cb0a38cdbb34522b6af1626b6be2fa2bc8), ['hf://datasets/FlowForgeLabAi/minimax-h3-8gb-bench@210fe4cb0a38cdbb34522b6af1626b6be2fa2bc8/h3_monitor_log.csv', 'hf://datasets/FlowForgeLabAi/minimax-h3-8gb-bench@210fe4cb0a38cdbb34522b6af1626b6be2fa2bc8/results.csv']

Please either edit the data files to have matching columns, or separate them into different configurations (see docs at https://hf.co/docs/hub/datasets-manual-configuration#multiple-configurations)
Traceback:    Traceback (most recent call last):
                File "/usr/local/lib/python3.14/site-packages/datasets/builder.py", line 1848, in _prepare_split_single
                  writer.write_table(table)
                  ~~~~~~~~~~~~~~~~~~^^^^^^^
                File "/usr/local/lib/python3.14/site-packages/datasets/arrow_writer.py", line 765, in write_table
                  self._write_table(pa_table, writer_batch_size=writer_batch_size)
                  ~~~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
                File "/usr/local/lib/python3.14/site-packages/datasets/arrow_writer.py", line 773, in _write_table
                  pa_table = table_cast(pa_table, self._schema)
                File "/usr/local/lib/python3.14/site-packages/datasets/table.py", line 2378, in table_cast
                  return cast_table_to_schema(table, schema)
                File "/usr/local/lib/python3.14/site-packages/datasets/table.py", line 2306, in cast_table_to_schema
                  raise CastError(
                  ...<3 lines>...
                  )
              datasets.table.CastError: Couldn't cast
              date: string
              label: string
              prompt_id: string
              model: string
              text_encoder: string
              lora: string
              lora_strength: double
              attn_backend: string
              opt_nodes: string
              shift: string
              head_chunks: int64
              chunkff: string
              steps: int64
              sampler: string
              scheduler: string
              seed: int64
              peak_vram_mib: int64
              peak_ram_used_gb: double
              peak_pagefile_gb: double
              peak_power_w: double
              peak_temp_c: int64
              wall_sec: double
              sec_load: double
              sec_denoise: double
              sec_post: double
              sec_total: double
              blocks_logged: int64
              seconds_per_block: double
              tespeed: string
              completed: bool
              crashed: bool
              process_crash: bool
              quality_note: string
              -- schema metadata --
              pandas: '{"index_columns": [{"kind": "range", "name": null, "start": 0, "' + 4177
              to
              {'timestamp': Value('string'), 'ram_avail_gb': Value('float64'), 'ram_total_gb': Value('float64'), 'ram_load_pct': Value('float64'), 'pagefile_avail_gb': Value('float64'), 'gpu_util_pct': Value('float64'), 'gpu_mem_util_pct': Value('float64'), 'sm_clock_mhz': Value('float64'), 'mem_clock_mhz': Value('float64'), 'temp_c': Value('float64'), 'power_w': Value('float64'), 'vram_used_mib': Value('float64'), 'vram_total_mib': Value('float64')}
              because column names don't match
              
              During handling of the above exception, another exception occurred:
              
              Traceback (most recent call last):
                File "/src/services/worker/src/worker/job_runners/config/parquet_and_info.py", line 1369, in compute_config_parquet_and_info_response
                  parquet_operations, partial, estimated_dataset_info = stream_convert_to_parquet(
                                                                        ~~~~~~~~~~~~~~~~~~~~~~~~~^
                      builder, max_dataset_size_bytes=max_dataset_size_bytes
                      ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
                  )
                  ^
                File "/src/services/worker/src/worker/job_runners/config/parquet_and_info.py", line 948, in stream_convert_to_parquet
                  builder._prepare_split(split_generator=splits_generators[split], file_format="parquet")
                  ~~~~~~~~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
                File "/usr/local/lib/python3.14/site-packages/datasets/builder.py", line 1694, in _prepare_split
                  for job_id, done, content in self._prepare_split_single(
                                               ~~~~~~~~~~~~~~~~~~~~~~~~~~^
                      gen_kwargs=gen_kwargs, job_id=job_id, **_prepare_split_args
                      ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
                  ):
                  ^
                File "/usr/local/lib/python3.14/site-packages/datasets/builder.py", line 1850, in _prepare_split_single
                  raise DatasetGenerationCastError.from_cast_error(
                  ...<4 lines>...
                  )
              datasets.exceptions.DatasetGenerationCastError: An error occurred while generating the dataset
              
              All the data files must have the same columns, but at some point there are 33 new columns ({'chunkff', 'model', 'process_crash', 'lora', 'peak_vram_mib', 'lora_strength', 'label', 'sec_post', 'quality_note', 'wall_sec', 'prompt_id', 'sec_total', 'date', 'seed', 'sec_load', 'attn_backend', 'peak_power_w', 'peak_pagefile_gb', 'peak_ram_used_gb', 'sec_denoise', 'peak_temp_c', 'blocks_logged', 'seconds_per_block', 'steps', 'completed', 'crashed', 'opt_nodes', 'scheduler', 'tespeed', 'shift', 'sampler', 'text_encoder', 'head_chunks'}) and 13 missing columns ({'ram_total_gb', 'mem_clock_mhz', 'temp_c', 'sm_clock_mhz', 'timestamp', 'pagefile_avail_gb', 'vram_used_mib', 'vram_total_mib', 'gpu_util_pct', 'gpu_mem_util_pct', 'power_w', 'ram_load_pct', 'ram_avail_gb'}).
              
              This happened while the csv dataset builder was generating data using
              
              hf://datasets/FlowForgeLabAi/minimax-h3-8gb-bench/results.csv (at revision 210fe4cb0a38cdbb34522b6af1626b6be2fa2bc8), ['hf://datasets/FlowForgeLabAi/minimax-h3-8gb-bench@210fe4cb0a38cdbb34522b6af1626b6be2fa2bc8/h3_monitor_log.csv', 'hf://datasets/FlowForgeLabAi/minimax-h3-8gb-bench@210fe4cb0a38cdbb34522b6af1626b6be2fa2bc8/results.csv']
              
              Please either edit the data files to have matching columns, or separate them into different configurations (see docs at https://hf.co/docs/hub/datasets-manual-configuration#multiple-configurations)

Need help to make the dataset viewer work? Make sure to review how to configure the dataset viewer, and open a discussion for direct support.

timestamp
string
ram_avail_gb
float64
ram_total_gb
float64
ram_load_pct
float64
pagefile_avail_gb
float64
gpu_util_pct
float64
gpu_mem_util_pct
float64
sm_clock_mhz
float64
mem_clock_mhz
float64
temp_c
float64
power_w
float64
vram_used_mib
float64
vram_total_mib
float64
15:03:49
4.67
15.26
69
15.64
5
4
285
11,001
45
9.7
4,435
8,151
15:03:59
4.63
15.26
69
13.46
100
68
2,400
11,001
61
91
6,469
8,151
15:04:09
4.52
15.26
70
13.64
99
42
2,527
11,001
66
93
6,367
8,151
15:04:19
4.38
15.26
71
13.74
100
68
2,437
11,001
70
94.7
6,359
8,151
15:04:29
0.72
15.26
95
7.24
99
8
2,587
11,001
68
62.8
8,037
8,151
15:04:39
1.03
15.26
93
7.37
99
70
2,790
9,001
69
85.6
8,104
8,151
15:04:49
0.87
15.26
94
7.29
100
5
2,797
9,001
69
58.9
8,100
8,151
15:04:59
0.88
15.26
94
7.33
100
12
2,805
9,001
63
42.3
8,097
8,151
15:05:09
0.79
15.26
94
7.22
99
7
2,805
9,001
62
39.9
8,083
8,151
15:05:19
0.75
15.26
95
7.17
100
5
2,812
9,001
62
36.2
8,070
8,151
15:05:29
0.73
15.26
95
7.2
100
9
2,805
9,001
66
47.5
8,061
8,151
15:05:39
0.77
15.26
94
7.22
100
5
2,812
9,001
61
34.9
8,059
8,151
15:05:49
0.8
15.26
94
7.34
100
9
2,812
9,001
60
35.6
8,034
8,151
15:05:59
0.79
15.26
94
7.31
100
6
2,805
9,001
67
55.3
8,055
8,151
15:06:09
0.67
15.26
95
7.14
100
29
2,797
9,001
63
52.8
8,083
8,151
15:06:19
0.7
15.26
95
7.3
100
5
2,812
9,001
60
33.4
8,009
8,151
15:06:29
0.73
15.26
95
7.44
100
3
2,812
9,001
59
32.5
7,934
8,151
15:06:39
0.71
15.26
95
7.41
100
3
2,812
9,001
59
32.3
7,934
8,151
15:06:49
0.7
15.26
95
7.33
100
3
2,812
9,001
59
32.4
8,002
8,151
15:06:59
0.72
15.26
95
7.43
100
4
2,805
9,001
66
50.9
7,933
8,151
15:07:09
0.64
15.26
95
7.38
100
3
2,812
9,001
60
33.1
7,933
8,151
15:07:19
0.65
15.26
95
7.4
100
3
2,812
9,001
59
31.6
7,933
8,151
15:07:29
0.68
15.26
95
7.43
100
3
2,812
9,001
59
32.7
7,933
8,151
15:07:39
1.76
15.26
88
7.43
100
3
2,812
9,001
58
32.5
7,933
8,151
15:07:49
1
15.26
93
6.4
100
40
2,805
9,001
61
56.8
7,933
8,151
15:07:59
1.01
15.26
93
6.41
100
7
2,812
9,001
62
50.3
7,933
8,151
15:08:09
1.02
15.26
93
6.42
100
3
2,812
9,001
59
32.8
7,933
8,151
15:08:19
0.92
15.26
94
6.32
100
3
2,812
9,001
59
33.2
7,977
8,151
15:08:29
0.82
15.26
94
6.32
100
4
2,812
9,001
59
33
7,970
8,151
15:08:39
0.87
15.26
94
6.38
100
4
2,812
9,001
59
33
7,979
8,151
15:08:49
0.88
15.26
94
6.38
100
3
2,812
9,001
59
32.4
8,004
8,151
15:08:59
0.87
15.26
94
6.37
100
4
2,797
9,001
65
51.2
8,003
8,151
15:09:09
0.85
15.26
94
6.4
100
5
2,812
9,001
60
33
8,004
8,151
15:09:19
0.86
15.26
94
6.4
100
3
2,812
9,001
59
32.5
8,009
8,151
15:09:29
0.87
15.26
94
6.44
100
4
2,812
9,001
59
32.4
7,978
8,151
15:09:39
0.89
15.26
94
6.47
100
3
2,812
9,001
58
32.1
7,979
8,151
15:09:49
0.89
15.26
94
6.41
44
4
2,167
9,001
57
30.8
8,004
8,151
15:09:59
0.91
15.26
94
6.48
100
8
2,812
9,001
62
48.4
7,978
8,151
15:10:09
0.91
15.26
94
6.48
100
3
2,812
9,001
59
32.7
7,978
8,151
15:10:19
0.85
15.26
94
6.43
100
3
2,812
9,001
59
33
8,004
8,151
15:10:29
0.85
15.26
94
6.44
100
3
2,812
9,001
59
33
8,009
8,151
15:10:39
0.85
15.26
94
6.47
100
3
2,812
9,001
58
32.5
7,978
8,151
15:10:49
0.83
15.26
94
6.45
100
27
2,805
9,001
60
46
8,004
8,151
15:10:59
1.05
15.26
93
6.48
100
6
2,805
9,001
62
47.3
7,978
8,151
15:11:09
1.04
15.26
93
6.44
100
3
2,812
9,001
59
32.8
8,003
8,151
15:11:19
1.01
15.26
93
6.39
100
3
2,812
9,001
59
32.9
8,044
8,151
15:11:29
1
15.26
93
6.39
100
3
2,812
9,001
59
32.8
8,040
8,151
15:11:39
1.04
15.26
93
6.52
100
3
2,812
9,001
58
32.6
7,965
8,151
15:11:49
1.02
15.26
93
6.43
100
4
2,805
9,001
62
49.9
8,034
8,151
15:11:59
1
15.26
93
6.41
100
8
2,812
9,001
61
49.4
8,040
8,151
15:12:09
0.99
15.26
93
6.39
100
5
2,812
9,001
59
33.7
8,045
8,151
15:12:19
1.02
15.26
93
6.51
100
5
2,812
9,001
59
33.1
7,978
8,151
15:12:29
1.02
15.26
93
6.48
100
4
2,812
9,001
58
32.9
8,003
8,151
15:12:39
1.45
15.26
90
6.46
100
4
2,812
9,001
58
32.2
8,035
8,151
15:12:49
1.36
15.26
91
6.37
100
4
2,805
9,001
62
55
8,050
8,151
15:12:59
1.35
15.26
91
6.36
100
5
2,812
9,001
59
33.9
8,055
8,151
15:13:09
1.45
15.26
90
6.33
100
3
2,812
9,001
59
32.3
8,035
8,151
15:13:19
1.45
15.26
90
6.32
100
4
2,812
9,001
58
32.3
8,035
8,151
15:13:29
1.47
15.26
90
6.36
100
7
2,812
9,001
59
33
8,055
8,151
15:13:39
1.46
15.26
90
6.32
100
20
2,812
9,001
61
47.5
8,060
8,151
15:13:49
1.47
15.26
90
6.36
100
4
2,797
9,001
62
52.6
8,055
8,151
15:13:59
1.44
15.26
90
6.35
100
3
2,812
9,001
59
32.8
8,055
8,151
15:14:09
1.41
15.26
90
6.35
100
3
2,812
9,001
59
32.7
8,035
8,151
15:14:19
1.42
15.26
90
6.34
100
3
2,812
9,001
58
32.5
8,055
8,151
15:14:29
1.42
15.26
90
6.33
100
3
2,812
9,001
58
32.5
8,055
8,151
15:14:39
1.42
15.26
90
6.34
100
30
2,805
9,001
61
52
8,060
8,151
15:14:49
1.43
15.26
90
6.36
100
4
2,805
9,001
63
49.6
8,055
8,151
15:14:59
1.41
15.26
90
6.35
100
3
2,812
9,001
59
32.8
8,055
8,151
15:15:09
1.42
15.26
90
6.36
100
3
2,812
9,001
59
32.8
8,055
8,151
15:15:19
1.4
15.26
90
6.33
100
4
2,812
9,001
58
33
8,060
8,151
15:15:29
1.4
15.26
90
6.36
100
3
2,812
9,001
58
32.6
8,040
8,151
15:15:39
1.39
15.26
90
6.33
100
9
2,805
9,001
60
46.7
8,055
8,151
15:15:49
1.38
15.26
90
6.33
100
4
2,805
9,001
63
48.4
8,055
8,151
15:15:59
1.39
15.26
90
6.34
100
4
2,812
9,001
59
32.7
8,055
8,151
15:16:09
1.4
15.26
90
6.35
100
4
2,812
9,001
58
32.8
8,020
8,151
15:16:19
1.44
15.26
90
6.44
100
5
2,812
9,001
58
33.4
8,041
8,151
15:16:29
1.42
15.26
90
6.41
100
5
2,812
9,001
58
32.8
8,046
8,151
15:16:39
1.42
15.26
90
6.4
100
4
2,805
9,001
63
58.9
8,046
8,151
15:16:49
1.43
15.26
90
6.32
100
5
2,812
9,001
60
33.5
8,054
8,151
15:16:59
1.33
15.26
91
6.31
100
3
2,812
9,001
59
32.3
8,041
8,151
15:17:09
1.21
15.26
92
6.3
100
4
2,812
9,001
58
32.1
8,062
8,151
15:17:19
1.28
15.26
91
6.36
100
4
2,812
9,001
58
34.8
8,037
8,151
15:17:29
1.41
15.26
90
6.39
53
4
2,452
9,001
56
32.9
8,024
8,151
15:17:39
1.39
15.26
90
6.38
100
9
2,805
9,001
59
49.3
8,018
8,151
15:17:49
1.33
15.26
91
6.34
100
3
2,812
9,001
58
32.1
8,020
8,151
15:17:59
1.34
15.26
91
6.35
100
3
2,812
9,001
57
32.7
8,026
8,151
15:18:09
1.36
15.26
91
6.37
100
3
2,812
9,001
56
32.2
8,021
8,151
15:18:19
1.34
15.26
91
6.35
100
4
2,812
9,001
56
32.7
8,029
8,151
15:18:29
1.26
15.26
91
6.15
100
24
2,812
9,001
59
46
8,078
8,151
15:18:39
1.29
15.26
91
6.22
100
4
2,805
9,001
63
48.3
8,048
8,151
15:18:49
1.21
15.26
92
6.1
100
4
2,812
9,001
57
32.9
8,034
8,151
15:18:59
1.07
15.26
92
6.02
100
4
2,812
9,001
57
31.7
8,025
8,151
15:19:09
1.17
15.26
92
6.16
100
4
2,812
9,001
56
31.5
8,050
8,151
15:19:19
1.17
15.26
92
6.14
100
5
2,812
9,001
56
32.5
8,049
8,151
15:19:29
1.17
15.26
92
6.1
100
3
2,812
9,001
56
30.8
8,004
8,151
15:19:39
1.4
15.26
90
6.28
100
4
2,805
9,001
61
51.4
7,988
8,151
15:19:49
1.4
15.26
90
6.3
100
4
2,805
9,001
63
49.8
7,986
8,151
15:19:59
1.39
15.26
90
6.27
100
4
2,812
9,001
56
31.9
7,991
8,151
15:20:09
1.41
15.26
90
6.27
100
6
2,812
9,001
56
31.7
7,996
8,151
15:20:19
1.35
15.26
91
6.17
100
5
2,812
9,001
56
32
8,022
8,151
End of preview.

MiniMax-H3 on an 8 GB laptop GPU — measured benchmark

First-hand measurements of running MiniMax-H3 video generation on a single consumer laptop: RTX 5060 Laptop (8 GB VRAM) + 15.26 GiB system RAM, Windows, ComfyUI 0.35.0.

This is not a leaderboard. It is a record of what actually happens on this class of hardware, including the runs that failed, because the failure mode turned out to be the interesting part.

Draft. Single machine, single seed (12345), single operating system. Read Limitations before quoting any number.

Why this dataset exists

The published envelope for H3 assumes datacentre GPUs. On an 8 GB laptop the two things that actually break are not the ones usually discussed:

  1. The binding constraint is system RAM, not VRAM. The working set is roughly 19.5 GiB (main model) + 15.0 GiB (NVFP4 text encoder) ≈ 34.5 GiB against 15.26 GiB of RAM. The process lives in the pagefile.
  2. That failure is invisible in the usual metrics. utilization.gpu keeps reading 99–100 % the whole time, because the GPU is not idle — it is starved between kernels. The signal that does move is board power.

h3_monitor_log.csv is included specifically to make point 2 checkable by anyone.

Files

results.csv — 18 runs × 33 columns

One row per generation run: a sweep over steps (4/6/8/20), sampler, turbo-LoRA on/off, TE-Speed on/off, resolution/aspect, and the video-encoding path.

column meaning
date, label, prompt_id run timestamp, human label, ComfyUI prompt UUID
model, text_encoder, lora, lora_strength exact checkpoint filenames used
attn_backend, opt_nodes attention backend and the optimisation nodes in the graph
shift, head_chunks, chunkff sampling shift; chunked-attention and chunked-FFN settings
steps, sampler, scheduler, seed sampling configuration
peak_vram_mib, peak_ram_used_gb, peak_pagefile_gb peak memory, not mean
peak_power_w, peak_temp_c peak board power and temperature
wall_sec end-to-end wall clock for the run
sec_load, sec_denoise, sec_post wall clock split across load / denoise / post
sec_total, blocks_logged, seconds_per_block denoise accounting (blocks logged by the model)
tespeed TE-Speed setting, where used
completed, crashed, process_crash outcome flags — failed runs are kept in
quality_note free-text note for that run (Chinese)

h3_monitor_log.csv — 534 samples × 13 columns, 10 s interval

A continuous telemetry trace of one session, sampled every 10 seconds.

timestamp, ram_avail_gb, ram_total_gb, ram_load_pct, pagefile_avail_gb, gpu_util_pct, gpu_mem_util_pct, sm_clock_mhz, mem_clock_mhz, temp_c, power_w, vram_used_mib, vram_total_mib

The power column is the point of this file:

power_w (n = 534)
min 0.0
median 32.8
max 106.6

…while gpu_util_pct reads 99–100 % across the same window. A healthy run on this machine draws 64–98 W. A flat ~33 W at 99 % utilisation is the pagefile-starvation signature.

What the data shows

  • T_wall ≈ 102 + 34.1 × steps seconds (10 s clip, 768×1024). 8 steps ≈ 858 s; 20 steps 2029–2613 s.
  • Fresh session: 0.9–6 s/block. After ~1.7 h of accumulated pagefile pressure: 212 s/block — same graph, same resolution.
  • A 259-module LoRA applied 208 modules (80.3 %); all 51 adaln_proj.linear silently skipped.
  • At H3's real attention geometry (56 heads × 128 head_dim, S = 8192, bf16), a measured backend comparison: SDPA 143.1 ms, comfy-kitchen INT8 22.8 ms, SageAttention 29.0 ms.
  • SaveVideo's H.264 re-encode fails when the frame's width or height is odd (yuv420p needs even dimensions). Even dimensions work. Diagnosed and reported upstream.

Limitations

Stated plainly, because a dataset like this is easy to over-read:

  • One machine, one seed. No run-to-run variance study was done; treat single digits as indicative, not reproducible to the second.
  • Peaks, not means. peak_* columns are maxima. Two runs with identical means can differ in peak, and peak is what pushes a 15 GiB box into the pagefile.
  • h3_monitor_log.csv has no date column — only HH:MM:SS. It covers exactly one session and cannot be aligned to the other machine.
  • Windows only. Pagefile behaviour is not portable to Linux swap.
  • quality_note is Chinese and informal, written during the run, not a formal rating.
  • Quality was judged by eye on a handful of clips; no reference metric is included.
  • No model weights are redistributed here. Filenames are recorded for reproducibility only.

Provenance

Collected by the author on their own hardware while building 8 GB-targeted ComfyUI workflows, using an instrumented custom sampler node that logs per-block timing, peak memory and board power. Scripts are not yet included in this repository.

License

MIT for the data and this card. Model weights are not included and remain under their own licenses.
Downloads last month
40
Total file size:
51.1 kB