The dataset viewer is not available for this split.
Error code: StreamingRowsError
Exception: CastError
Message: Couldn't cast
model: string
s_cap: struct<NF4: double, INT8: double>
child 0, NF4: double
child 1, INT8: double
arch_label: struct<turing: string, ada: string, blackwell: string, ampere: string, hopper: string>
child 0, turing: string
child 1, ada: string
child 2, blackwell: string
child 3, ampere: string
child 4, hopper: string
borrow: struct<hopper: string>
child 0, hopper: string
curves: struct<ada: struct<INT8: struct<A: double, S: double, Nstar: double, resid_std: double, n_min: doubl (... 1324 chars omitted)
child 0, ada: struct<INT8: struct<A: double, S: double, Nstar: double, resid_std: double, n_min: double, n_max: do (... 359 chars omitted)
child 0, INT8: struct<A: double, S: double, Nstar: double, resid_std: double, n_min: double, n_max: double, crossov (... 118 chars omitted)
child 0, A: double
child 1, S: double
child 2, Nstar: double
child 3, resid_std: double
child 4, n_min: double
child 5, n_max: double
child 6, crossover_b: null
child 7, anchors: list<item: struct<N: double, dE: double, model: string, gpu: string, n: int64>>
child 0, item: struct<N: double, dE: double, model: string, gpu: string, n: int64>
child 0, N: double
child 1, dE: double
child 2, model: string
child 3, gpu: string
child 4, n: int64
child 8, loo_mae: double
child 1, NF4: struct<A:
...
string, batch_size: int64, context_length: i (... 5 chars omitted)
child 0, model_name: string
child 1, params_b: double
child 2, precision: string
child 3, batch_size: int64
child 4, context_length: int64
results: struct<total_energy_joules: double, tokens_generated: int64, energy_per_token_mj: double, avg_power_ (... 124 chars omitted)
child 0, total_energy_joules: double
child 1, tokens_generated: int64
child 2, energy_per_token_mj: double
child 3, avg_power_watts: double
child 4, throughput_tokens_per_s: double
child 5, basis: string
child 6, fp16_energy_per_token_mj: double
child 7, vs_fp16_energy_pct: double
notice: string
system_under_test: struct<gpu: string, gpu_arch: string, accelerator_count: int64, host: string>
child 0, gpu: string
child 1, gpu_arch: string
child 2, accelerator_count: int64
child 3, host: string
measurement: struct<method: string, sample_rate_hz: int64, tokens_per_run: int64, iterations: int64, warmup: int6 (... 27 chars omitted)
child 0, method: string
child 1, sample_rate_hz: int64
child 2, tokens_per_run: int64
child 3, iterations: int64
child 4, warmup: int64
child 5, iterations_note: string
measurement_source: string
provenance: struct<tool: string, dataset_doi: string, rtx4090_deep_dive_doi: string, paper_ssrn: string, code: s (... 6 chars omitted)
child 0, tool: string
child 1, dataset_doi: string
child 2, rtx4090_deep_dive_doi: string
child 3, paper_ssrn: string
child 4, code: string
to
{'schema_version': Value('string'), 'benchmark': Value('string'), 'follows_mlcommons_energy_reporting_conventions': Value('bool'), 'certified_benchmark_result': Value('bool'), 'scenario': Value('string'), 'scenario_note': Value('string'), 'timestamp_utc': Value('string'), 'system_under_test': {'gpu': Value('string'), 'gpu_arch': Value('string'), 'accelerator_count': Value('int64'), 'host': Value('string')}, 'workload': {'model_name': Value('string'), 'params_b': Value('float64'), 'precision': Value('string'), 'batch_size': Value('int64'), 'context_length': Value('int64')}, 'measurement': {'method': Value('string'), 'sample_rate_hz': Value('int64'), 'tokens_per_run': Value('int64'), 'iterations': Value('int64'), 'warmup': Value('int64'), 'iterations_note': Value('string')}, 'software': {'python': Value('string'), 'packages': {'torch': Value('string'), 'transformers': Value('string'), 'bitsandbytes': Value('string'), 'accelerate': Value('string'), 'nvidia-ml-py': Value('string'), 'sentencepiece': Value('string')}, 'torch_cuda': Value('string'), 'nvidia_driver': Value('string'), 'reference_pins': {'transformers': Value('string'), 'bitsandbytes': Value('string'), 'accelerate': Value('string'), 'nvidia-ml-py': Value('string'), 'sentencepiece': Value('string')}, 'matches_reference_pins': Value('bool')}, 'provenance': {'tool': Value('string'), 'dataset_doi': Value('string'), 'rtx4090_deep_dive_doi': Value('string'), 'paper_ssrn': Value('string'), 'code': Value('string')}, 'notice': Value('string'), 'measurement_source': Value('string'), 'results': {'total_energy_joules': Value('float64'), 'tokens_generated': Value('int64'), 'energy_per_token_mj': Value('float64'), 'avg_power_watts': Value('float64'), 'throughput_tokens_per_s': Value('float64'), 'basis': Value('string'), 'fp16_energy_per_token_mj': Value('float64'), 'vs_fp16_energy_pct': Value('float64')}, 'quality': {'metric': Value('string'), 'basis': Value('string'), 'value': Value('float64'), 'mean_nll': Value('float64'), 'tokens_evaluated': Value('int64'), 'windows': Value('int64'), 'seq_len': Value('int64'), 'corpus': {'name': Value('string'), 'bytes': Value('int64'), 'sha256': Value('string')}, 'note': Value('string'), 'fp16_value': Value('float64'), 'delta_vs_fp16_pct': Value('float64')}, 'share_url': Value('string')}
because column names don't match
Traceback: Traceback (most recent call last):
File "/src/services/worker/src/worker/utils.py", line 147, in get_rows_or_raise
return get_rows(
dataset=dataset,
...<4 lines>...
column_names=column_names,
)
File "/src/libs/libcommon/src/libcommon/utils.py", line 272, in decorator
return func(*args, **kwargs)
File "/src/services/worker/src/worker/utils.py", line 127, in get_rows
rows_plus_one = list(itertools.islice(safe_iter(ds, dataset=dataset), rows_max_number + 1))
File "/src/services/worker/src/worker/utils.py", line 483, in safe_iter
yield from ds.decode(False) if ds.features else ds
File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 2951, in __iter__
for key, example in ex_iterable:
^^^^^^^^^^^
File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 2461, in __iter__
for key, pa_table in self._iter_arrow():
~~~~~~~~~~~~~~~~^^
File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 2486, in _iter_arrow
for key, pa_table in self.ex_iterable._iter_arrow():
~~~~~~~~~~~~~~~~~~~~~~~~~~~~^^
File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 547, in _iter_arrow
for key, pa_table in iterator:
^^^^^^^^
File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 430, in _iter_arrow
for key, pa_table in self.generate_tables_fn(**gen_kwags):
~~~~~~~~~~~~~~~~~~~~~~~^^^^^^^^^^^^^
File "/usr/local/lib/python3.14/site-packages/datasets/packaged_modules/json/json.py", line 343, in _generate_tables
self._cast_table(pa_table, json_field_paths=json_field_paths),
~~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/usr/local/lib/python3.14/site-packages/datasets/packaged_modules/json/json.py", line 132, in _cast_table
pa_table = table_cast(pa_table, self.info.features.arrow_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
model: string
s_cap: struct<NF4: double, INT8: double>
child 0, NF4: double
child 1, INT8: double
arch_label: struct<turing: string, ada: string, blackwell: string, ampere: string, hopper: string>
child 0, turing: string
child 1, ada: string
child 2, blackwell: string
child 3, ampere: string
child 4, hopper: string
borrow: struct<hopper: string>
child 0, hopper: string
curves: struct<ada: struct<INT8: struct<A: double, S: double, Nstar: double, resid_std: double, n_min: doubl (... 1324 chars omitted)
child 0, ada: struct<INT8: struct<A: double, S: double, Nstar: double, resid_std: double, n_min: double, n_max: do (... 359 chars omitted)
child 0, INT8: struct<A: double, S: double, Nstar: double, resid_std: double, n_min: double, n_max: double, crossov (... 118 chars omitted)
child 0, A: double
child 1, S: double
child 2, Nstar: double
child 3, resid_std: double
child 4, n_min: double
child 5, n_max: double
child 6, crossover_b: null
child 7, anchors: list<item: struct<N: double, dE: double, model: string, gpu: string, n: int64>>
child 0, item: struct<N: double, dE: double, model: string, gpu: string, n: int64>
child 0, N: double
child 1, dE: double
child 2, model: string
child 3, gpu: string
child 4, n: int64
child 8, loo_mae: double
child 1, NF4: struct<A:
...
string, batch_size: int64, context_length: i (... 5 chars omitted)
child 0, model_name: string
child 1, params_b: double
child 2, precision: string
child 3, batch_size: int64
child 4, context_length: int64
results: struct<total_energy_joules: double, tokens_generated: int64, energy_per_token_mj: double, avg_power_ (... 124 chars omitted)
child 0, total_energy_joules: double
child 1, tokens_generated: int64
child 2, energy_per_token_mj: double
child 3, avg_power_watts: double
child 4, throughput_tokens_per_s: double
child 5, basis: string
child 6, fp16_energy_per_token_mj: double
child 7, vs_fp16_energy_pct: double
notice: string
system_under_test: struct<gpu: string, gpu_arch: string, accelerator_count: int64, host: string>
child 0, gpu: string
child 1, gpu_arch: string
child 2, accelerator_count: int64
child 3, host: string
measurement: struct<method: string, sample_rate_hz: int64, tokens_per_run: int64, iterations: int64, warmup: int6 (... 27 chars omitted)
child 0, method: string
child 1, sample_rate_hz: int64
child 2, tokens_per_run: int64
child 3, iterations: int64
child 4, warmup: int64
child 5, iterations_note: string
measurement_source: string
provenance: struct<tool: string, dataset_doi: string, rtx4090_deep_dive_doi: string, paper_ssrn: string, code: s (... 6 chars omitted)
child 0, tool: string
child 1, dataset_doi: string
child 2, rtx4090_deep_dive_doi: string
child 3, paper_ssrn: string
child 4, code: string
to
{'schema_version': Value('string'), 'benchmark': Value('string'), 'follows_mlcommons_energy_reporting_conventions': Value('bool'), 'certified_benchmark_result': Value('bool'), 'scenario': Value('string'), 'scenario_note': Value('string'), 'timestamp_utc': Value('string'), 'system_under_test': {'gpu': Value('string'), 'gpu_arch': Value('string'), 'accelerator_count': Value('int64'), 'host': Value('string')}, 'workload': {'model_name': Value('string'), 'params_b': Value('float64'), 'precision': Value('string'), 'batch_size': Value('int64'), 'context_length': Value('int64')}, 'measurement': {'method': Value('string'), 'sample_rate_hz': Value('int64'), 'tokens_per_run': Value('int64'), 'iterations': Value('int64'), 'warmup': Value('int64'), 'iterations_note': Value('string')}, 'software': {'python': Value('string'), 'packages': {'torch': Value('string'), 'transformers': Value('string'), 'bitsandbytes': Value('string'), 'accelerate': Value('string'), 'nvidia-ml-py': Value('string'), 'sentencepiece': Value('string')}, 'torch_cuda': Value('string'), 'nvidia_driver': Value('string'), 'reference_pins': {'transformers': Value('string'), 'bitsandbytes': Value('string'), 'accelerate': Value('string'), 'nvidia-ml-py': Value('string'), 'sentencepiece': Value('string')}, 'matches_reference_pins': Value('bool')}, 'provenance': {'tool': Value('string'), 'dataset_doi': Value('string'), 'rtx4090_deep_dive_doi': Value('string'), 'paper_ssrn': Value('string'), 'code': Value('string')}, 'notice': Value('string'), 'measurement_source': Value('string'), 'results': {'total_energy_joules': Value('float64'), 'tokens_generated': Value('int64'), 'energy_per_token_mj': Value('float64'), 'avg_power_watts': Value('float64'), 'throughput_tokens_per_s': Value('float64'), 'basis': Value('string'), 'fp16_energy_per_token_mj': Value('float64'), 'vs_fp16_energy_pct': Value('float64')}, 'quality': {'metric': Value('string'), 'basis': Value('string'), 'value': Value('float64'), 'mean_nll': Value('float64'), 'tokens_evaluated': Value('int64'), 'windows': Value('int64'), 'seq_len': Value('int64'), 'corpus': {'name': Value('string'), 'bytes': Value('int64'), 'sha256': Value('string')}, 'note': Value('string'), 'fp16_value': Value('float64'), 'delta_vs_fp16_pct': Value('float64')}, 'share_url': Value('string')}
because column names don't matchNeed help to make the dataset viewer work? Make sure to review how to configure the dataset viewer, and open a discussion for direct support.
Direct NVML measurements of quantized LLM inference energy across 0.5B–14B parameters on four NVIDIA architecture classes: RTX 5090 (Blackwell), RTX 4090D (Ada), A800 80GB (Ampere) and Tesla T4 (Turing).
360+ configurations (v1.1.0) covering FP16, NF4, INT8 and FP8, n=2 repeated trials per configuration, CV < 2%. Companion deposits add an RTX 4090 deep dive (three sessions, paired perplexity) and an RTX 5090 four-precision session.
Key findings
Small-model quantization paradox: below the break-even size, quantization costs energy — NF4 at 0.5B is +46% on RTX 4090 (measured)
NF4 break-even (crossover): ≈2.1B (Tesla T4, Turing) · ≈3.7B (RTX 4090, Ada) · ≈4.8B (RTX 5090, Blackwell, v1.1.0 stack). A September 2026 re-run on a newer stack (bitsandbytes 0.50.2 / torch 2.14 / CUDA 13.0) moved the 5090 crossover to ≈1.8B. The crossover belongs to a card+stack combination, not to an architecture — above it NF4 saves (−28% at 7B on RTX 4090, measured)
INT8 is not automatically cheaper: default load_in_8bit=True costs +50–242% energy on RTX 4090 (deep dive, n=1) and +17–147% across the broader test set. Root cause: mixed-precision decomposition (INT8↔FP16 conversion overhead), not INT8 compute itself. Fix: set llm_int8_threshold=0.0 to recover +79% throughput and −36% energy on average.
FP8 paradox: up to +701% energy overhead on RTX 5090 (software immaturity — confirmed by torchao maintainers in #4094)
Batch size is the largest single lever: BS=1→64 on A800 reduces energy per request by 95.7% (55.5× throughput scaling).
🎯 Call for replications — help fix this dataset's biggest weakness Almost every point in this dataset was measured by one person on rented GPUs. Independent replication is the single most valuable contribution you can make. If you have any CUDA GPU, one hour is enough:
Option A — one command (Docker):
docker run --rm --gpus all --user "$(id -u):$(id -g)"
-v "$PWD/out:/workspace/outputs"
ghcr.io/hongping-zh/ecocompute-mlcube:latest energy_estimate
--model TinyLlama/TinyLlama-1.1B-Chat-v1.0 --params_b 1.1
--precision NF4 --gpu_arch auto --output_dir /workspace/outputs --share
Option B — zero install (Google Colab, free T4): Open the measurement notebook
Then drop the resulting energy.json on quantenergy.tech/replications — it is validated in your browser (13 checks: schema, GPU/arch consistency, sampling rate, thermal state, FP16 baseline) and plotted against the published curves, and you are credited by name on the page.
Most valuable hardware right now: H100 / H200 (Hopper — currently estimated from Ampere curves, no measured data at all) · A100 / RTX 3090 (Ampere) · L40S (Ada) · consumer Blackwell (RTX 5080/5070) · independent RTX 4090 / 5090 replications of the crossover.
Contributed measurements are folded into the dataset archive (CC BY 4.0) with attribution.
Try it interactively Website & estimator — every number computed in the browser, no install, no API. Query these benchmarks conversationally with the EcoLobster AI advisor on ClawHub.
Dataset details Main dataset v1.1.0 (supersedes v1, 113 configurations): 360+ configurations, 0.5B–14B, 5 GPU models / 4 architecture classes, FP16/NF4/INT8/FP8, n=2, CV < 2% Models: Qwen2 (0.5B/1.5B/7B/14B), TinyLlama-1.1B, Mistral-7B, Yi-1.5 (6B/9B) Measurement: direct NVML GPU-package power sampling, trapezoidal energy integration, under protocol EcoCompute v1.1 Links & references Resource URL Website, estimator & replications https://quantenergy.tech Measurement container (MLCube) https://github.com/hongping-zh/ecocompute-mlcube Main dataset v1.1.0 (version DOI) https://doi.org/10.5281/zenodo.19647290 Main dataset (concept DOI, always latest) https://doi.org/10.5281/zenodo.19647289 RTX 4090 deep dive (3 sessions) https://doi.org/10.5281/zenodo.22037483 RTX 5090 four-precision session https://doi.org/10.5281/zenodo.22855133 Protocol v1.1 https://doi.org/10.5281/zenodo.22958675 ClawHub skill (interactive) https://clawhub.ai/hongping-zh/ecocompute HF Optimum docs (findings integrated, dataset v1.1.0) https://huggingface.co/docs/optimum/concept_guides/quantization torchao FP8 issue #4094 https://github.com/pytorch/ao/issues/4094 Community adoption Findings integrated into HuggingFace Optimum official quantization documentation (PR #2410, merged Mar 2026; based on dataset v1.1.0) FP8 energy anomaly confirmed by torchao maintainers (Issue #4094) Discussed with the MLCommons Power Working Group on energy metrics for MLPerf Inference (inference #2558) Related contributions: bitsandbytes PR #1882, Transformers PR #44407 Citation @dataset{zhang2026ecocompute, author = {Hongping Zhang}, title = {EcoCompute: Energy Efficiency Benchmark for Quantized Language Models}, year = {2026}, publisher = {Zenodo}, doi = {10.5281/zenodo.19647290}, url = {https://doi.org/10.5281/zenodo.19647290} }
The DOI above is the frozen v1.1.0 version; the concept DOI 10.5281/zenodo.19647289 always resolves to the newest version.
License CC BY 4.0
- Downloads last month
- 236