Datasets:

Intelion
/

gpuark-gpu-dataset

license: cc-by-4.0
language:
  - en
  - ru
tags:
  - gpu
  - hardware
  - nvidia
  - amd
  - benchmarks
  - semiconductors
pretty_name: GPU Ark — GPU specifications & benchmarks (13.5k GPUs, 1999–2025)
size_categories:
  - 10K<n<100K
source_datasets:
  - original
configs:
  - config_name: gpu_specs
    data_files: gpuark-gpu-specs.csv
  - config_name: benchmarks
    data_files: gpuark-benchmarks.csv

GPU Ark — open GPU specifications & benchmarks dataset

Specifications of 13,566 GPUs released between 1999 and 2025 — from the GeForce 256 to NVIDIA Blackwell and AMD Instinct MI355X — plus 993 third-party benchmark results. Curated and maintained by GPU Ark (a GPU catalog & price comparison project). Canonical source and always-fresh copy: https://gpuark.com/datasets/.

Files

File	Rows	What
`gpuark-gpu-specs.csv`	13,566	One row per GPU — public spec columns
`gpuark-benchmarks.csv`	993	Third-party benchmark results, join on `gpu_id`
`gpuark-gpu-dataset.sqlite`	—	Both tables (`gpu_specs`, `benchmarks`) for SQL

Key columns (`gpu_specs`)

id, name, slug, vendor (nvd/amd/int), manufacturer, arch_name, card_release_date, proc_foundry, proc_process_size, proc_transistors, cores, tensor_cores, base_clock, boost_clock, ram, ram_type, bus_width, ram_bandwidth, fp16/fp32/fp64/bf16/tf32/int8_performance, tdp, multi_gpu, api_cuda, is_retail_board, gpi_value.

Quick start

import pandas as pd
df = pd.read_csv("gpuark-gpu-specs.csv", parse_dates=["card_release_date"])
nv = df[df.vendor == "nvd"]
# peak FP32 flagship per year
print(nv.groupby(nv.card_release_date.dt.year).fp32_performance.max())

Known issues (read before drawing conclusions)

vendor is set for ~2,360 of ~13,566 rows (nvd/amd/int); the rest are mostly partner/OEM board variants without a chip-vendor tag. Filter on vendor for vendor-level work.
ram (VRAM) unit is inconsistent across eras — older cards store MB, newer store GB (a value ≥ 256 on a pre-2018 card is almost certainly MB).
fp16/bf16/int8 are sparse and not consistently tensor-vs-non-tensor across vendors (NVIDIA Ampere+ tensor figures are often listed with structured sparsity = 2× dense). Don't compare low-precision peaks cross-vendor without checking the card.
card_release_date has a handful of implausible years — filter to 1998..2025.
is_retail_board=True = AIB/OEM editions of a reference chip (near-duplicates).

License & attribution

CC BY 4.0 — free to use with attribution to gpuark.com.

Citation

GPU Ark (2026). GPU specifications & benchmarks dataset. https://gpuark.com/datasets/