README.md

metadata

annotations_creators:
  - machine-generated
  - expert-generated
language:
  - en
  - es
  - pt
license: mit
multilinguality:
  - multilingual
size_categories:
  - n<1K
source_datasets:
  - original
task_categories:
  - tabular-classification
  - tabular-regression
  - feature-extraction
task_ids: []
pretty_name: 2026 LLM Hardware & VRAM Memory Profiling Dataset
tags:
  - llm
  - vram
  - inference
  - hardware
  - llama-cpp
  - ollama
  - vllm
  - gqa
  - moe
  - benchmarks
  - iaexplore
configs:
  - config_name: default
    data_files:
      - split: train
        path:
          - hardware_2026.json
          - models_2026.json

🧠 IAExplore 2026 LLM Hardware & VRAM Memory Profiling Dataset

Official dataset providing verified architectural parameters for cutting-edge Large Language Models (LLMs) and 2026 consumer/workstation AI hardware compute profiles.

Curated by the IAExplore Research & Systems Infrastructure Team.

---

📌 Dataset Summary

This dataset serves as the ground-truth technical foundation for:

1. Mathematical VRAM Allocation: Weight sizing with unquantized metadata overhead factor ($1.08$), KV-Cache memory consumption accounting for Grouped-Query Attention (GQA), and CUDA/Metal runtime execution buffers.
2. Mixture-of-Experts (MoE) Bandwidth Profiling: Distinguishing active parameter footprints ($P_{\text{active}}$) from total stored parameters ($P_{\text{total}}$) per token generation step.
3. Hardware Latency & Tokens/Second Estimation: Exact physical memory bandwidths ($\text{BW}_{\text{mem}}$), bus topologies (GDDR7, LPDDR5X, GDDR6X, HBM), and bus saturation efficiency coefficients ($\eta$).

---

🗂️ Data Structure

1. models_2026.json

Contains verified tensor specifications for:

• llama-3.3-70b (Meta AI)
• llama-3.1-8b (Meta AI)
• deepseek-r1-14b (DeepSeek AI)
• deepseek-r1-32b (DeepSeek AI)
• qwen-2.5-coder-32b (Alibaba Cloud)
• mixtral-8x7b-moe (Mistral AI)
• gemma-2-9b (Google DeepMind)
• phi-4-14b (Microsoft Research)

Key fields:

• total_params_b: Total billion parameters.
• active_params_b: Active parameters loaded per generation token (vital for MoE architectures).
• num_hidden_layers ($L$): Number of transformer blocks.
• num_attention_heads ($H_{\text{q}}$): Query heads.
• num_key_value_heads ($H_{\text{kv}}$): Key/Value heads for GQA/MHA.
• head_dim ($D_{\text{head}}$): Dimension per head.
• gqa_factor: Compression ratio ($H_{\text{q}} / H_{\text{kv}}$).

2. hardware_2026.json

Profiles for 2026 silicon across Discrete GPUs, Apple Silicon, and NPU/APU SoC systems:

• NVIDIA GeForce RTX 5090 (32 GB GDDR7, 1792 GB/s, 575W)
• NVIDIA GeForce RTX 4090 (24 GB GDDR6X, 1008 GB/s, 450W)
• NVIDIA GeForce RTX 4070 SUPER (12 GB GDDR6X, 504 GB/s, 220W)
• NVIDIA GeForce RTX 4060 Mobile (8 GB GDDR6, 256 GB/s, 115W)
• Apple M4 Max (128 GB Unified LPDDR5X, 546 GB/s, 90W)
• Apple M4 Pro (48 GB Unified LPDDR5X, 273 GB/s, 60W)
• Apple M4 (16 GB Unified LPDDR5X, 120 GB/s, 30W)
• AMD Ryzen AI Max+ 395 Strix Halo (96 GB assignable, 256 GB/s, 126 TOPS, 120W)
• Intel Core Ultra 9 288V Lunar Lake (24 GB shared, 136 GB/s, 120 TOPS, 30W)
• Qualcomm Snapdragon X Elite (48 GB assignable, 135 GB/s, 45 TOPS, 45W)

---

🔬 Mathematical Formulas

Model Weights Memory:

$${\text{VRAM}}_{\text{weights}}=\frac{P_{\text{total}}\times {10}^9\times ({\text{bits}}_{\text{quant}}\mathrm{/}8)}{{1024}^3}\times 1.08\backslash text\{VRAM\}\_\{\backslash text\{weights\}\}\; =\; \backslash frac\{P\_\{\backslash text\{total\}\}\; \backslash times\; 10^9\; \backslash times\; (\backslash text\{bits\}\_\{\backslash text\{quant\}\}\; /\; 8)\}\{1024^3\}\; \backslash times\; 1.08$$

Exact KV-Cache Memory (with GQA):

$${\text{VRAM}}_{\text{kv}}=\frac{2\times L\times H_{\text{kv}}\times D_{\text{head}}\times C_{\text{ctx}}\times B_{\text{batch}}\times ({\text{bits}}_{\text{kv}}\mathrm{/}8)}{{1024}^3}\backslash text\{VRAM\}\_\{\backslash text\{kv\}\}\; =\; \backslash frac\{2\; \backslash times\; L\; \backslash times\; H\_\{\backslash text\{kv\}\}\; \backslash times\; D\_\{\backslash text\{head\}\}\; \backslash times\; C\_\{\backslash text\{ctx\}\}\; \backslash times\; B\_\{\backslash text\{batch\}\}\; \backslash times\; (\backslash text\{bits\}\_\{\backslash text\{kv\}\}\; /\; 8)\}\{1024^3\}$$

Autoregressive Inference Throughput (Tokens/Second):

$${\text{TPS}}_{\text{gen}}=\frac{{\text{BW}}_{\text{mem}}\times {\eta }_{\text{efficiency}}}{{\text{VRAM}}_{\text{active}}+({\text{VRAM}}_{\text{kv}}\times 0.15)}\backslash text\{TPS\}\_\{\backslash text\{gen\}\}\; =\; \backslash frac\{\backslash text\{BW\}\_\{\backslash text\{mem\}\}\; \backslash times\; \backslash eta\_\{\backslash text\{efficiency\}\}\}\{\backslash text\{VRAM\}\_\{\backslash text\{active\}\}\; +\; (\backslash text\{VRAM\}\_\{\backslash text\{kv\}\}\; \backslash times\; 0.15)\}$$

---

🔗 Authoritative References & In-Depth Guides

• AI Hardware Architecture Guide (EN): NPU, TPU and GPU Complete Guide
• Guía de Hardware de IA (ES): Arquitecturas NPU, TPU y GPU para Inteligencia Artificial
• Guia de Hardware de IA (PT): Hardware para IA: Guia Completo NPU, TPU e GPU
• 2026 PC & Hardware Benchmarks (EN): Best Computers of 2026: Benchmark Guide
• Comparativas de Ordenadores 2026 (ES): Los Mejores Ordenadores de 2026: Guía y Benchmarks
• Melhores Computadores 2026 (PT): Melhores Computadores 2026: Guia Completo de Compra
• ARM & Mobile AI Devices 2026 (PT): Melhores Tablets e Dispositivos ARM
• Landauer Principle & Reversible Computing (ES): Física de la Computación y Límite Termodinámico
• Interactive Neural Network Simulator: IAExplore Real-Time Simulator

---

📜 Citation

If you use this dataset in academic research, benchmark pipelines, or LLM serving infrastructure, please cite:

@misc{iaexplore_llm_profiler_2026,
  author = {IAExplore Research Team},
  title = {IAExplore LLM Profiler: Deterministic VRAM and Hardware Inference Dataset},
  year = {2026},
  publisher = {Zenodo},
  doi = {10.5281/zenodo.14889210},
  url = {https://www.iaexplore.com/}
}