System Requirements
Hardware, operating system, GPU, storage, and network requirements for BCILattice.
Requirements at a Glance
| Requirement | Minimum | Recommended | High-Performance |
|---|---|---|---|
| RAM | 8 GB | 16 GB | 32 GB+ |
| CPU | 4-core x86-64, 2.0 GHz | 8-core, 3.0 GHz+ | 16-core workstation |
| Storage | 5 GB free | 20 GB+ SSD | 500 GB+ NVMe SSD |
| GPU | Not required | NVIDIA RTX 3060+ (8 GB VRAM) | NVIDIA RTX 4090 / A100 |
| OS | Current packaged build from download page | Modern 64-bit workstation OS | Same |
| Internet | Required (sign-in and licence validation) | 10 Mbps+ for cloud sync | 100 Mbps+ for team workspaces |
Hardware Requirements
RAM
| Use Case | Recommended RAM |
|---|---|
| Single-subject sessions, up to 64 channels | 8 GB |
| Multi-subject sessions (10–20 subjects), up to 128 channels | 16 GB |
| Large datasets (>50 subjects, high-density EEG 256+ channels) | 32 GB |
| fMRI-EEG co-registration or concurrent fNIRS+EEG | 32 GB+ |
| Deep learning (large batch training, high-res data) | 16–32 GB (+ GPU VRAM) |
CPU
BCILattice uses multi-threading for preprocessing and parallel subject training. More cores = faster batch training on CPU.
| Scenario | Recommended CPU |
|---|---|
| Single-subject, classical ML (LDA, SVM) | 4-core, any modern CPU |
| Multi-subject batch training (10+ subjects) | 8-core, Intel i7 / AMD Ryzen 7 |
| Large cohort (50+ subjects), fast iteration | 16-core, Intel i9 / AMD Ryzen 9 / Threadripper |
| Deep learning (without GPU) | 8-core+ strongly recommended |
Apple Silicon can run supported macOS builds where provided. PyTorch acceleration depends on the packaged runtime and local MPS support.
Operating System Support
| Platform | Supported Versions | Notes |
|---|---|---|
| Windows | Windows 10 (64-bit, 21H2+), Windows 11 | Use the currently published installer; user-level packaging depends on release |
| macOS | macOS 12 Monterey – macOS 15 Sequoia | Use the current macOS package when listed; PyTorch MPS support depends on runtime availability. |
| Linux | Ubuntu 20.04 LTS, 22.04 LTS, 24.04 LTS; Debian 11+; Fedora 38+; RHEL 8+; Arch (any recent) | Use the current Linux package when listed; otherwise run from the supported source/developer environment. |
GPU & CUDA Support
A GPU is optional but strongly recommended for deep learning models (EEGNet, DeepConvNet, EEGTransformer). CPU-only training is always available.
| GPU Platform | Support | Details |
|---|---|---|
| NVIDIA CUDA | Full support | CUDA 11.8 – 12.4. Requires matching NVIDIA driver (≥520.61). RTX 3060 or better recommended. |
| Apple Metal (MPS) | Runtime-dependent | Available only when the packaged PyTorch runtime and local hardware support MPS. |
| AMD ROCm | Not supported | AMD GPUs fall back to CPU automatically. ROCm support is planned. |
| Intel Arc / Xe | Not supported | Falls back to CPU. Intel XPU support planned. |
Recommended NVIDIA GPUs
| Tier | GPU | VRAM | Use Case |
|---|---|---|---|
| Entry | RTX 3060 / RTX 4060 | 8 GB | Single-subject deep learning, medium datasets |
| Mid-range | RTX 3080 / RTX 4070 | 10–12 GB | Multi-subject batch deep learning |
| High-end | RTX 4090 / RTX 3090 | 24 GB | Large cohort, transformer models, large batch sizes |
| Workstation | NVIDIA A100 / H100 | 40–80 GB | Institutional high-performance compute clusters |
Storage Requirements
| Component | Space Required | Notes |
|---|---|---|
| BCILattice application | Varies by package | Includes the desktop app and packaged runtime dependencies where provided |
| Local database directory | ~100 MB to 2 GB+ | Grows with studies, sessions, and experiment history |
| MLflow experiment store | ~50 MB – 5 GB | Depends on number of runs and stored artifacts |
| Session data (per 10-subject EEG) | ~500 MB – 5 GB | Varies by channel count, duration, sampling rate |
| Preprocessed arrays cache | ~200 MB – 2 GB per session | Cleared by Settings → Storage → Clear Cache |
| Trained model files | ~1 MB – 500 MB per model | Deep learning models are larger (50–500 MB) |
Recommended storage type: NVMe SSD for the BCILattice data directory significantly improves dataset import speed and preprocessing throughput. Spinning HDDs work but are 3–10x slower for large multi-subject datasets.
Network Requirements
BCILattice requires an internet connection. There is no offline mode. An active account is needed to run it, and paid features are unlocked by a licence check against the server, not by anything cached on disk. The heavy work — import, preprocessing, pipeline building, training, analysis, reports — all runs on your machine, but the connection has to be there.
A dropped connection does not close the app: it keeps running, and paid features stay available for the remaining life of the current licence token, about ten minutes. After that the app falls back to Free features and keeps retrying in the background, unlocking again on the first successful check. At a cold start with no network there is no token to fall back on, so the session is Free until connectivity returns.
| Feature | Required Bandwidth | Hosts to Allow |
|---|---|---|
| Initial download & install | Varies by package | bcinexus.xyz |
| Account sign-in & activation | Minimal (JSON API calls) | api.bcinexus.xyz |
| Study/cloud sync | Depends on selected package | api.bcinexus.xyz |
| Optional cloud upload | Limited by active plan and workspace quota | BCINexus storage endpoint |
| Community study download | <10 MB per study | api.bcinexus.xyz |
| Team workspace sync and requests | Depends on selected files and actions | BCINexus API endpoint |
| App updates | ~100–500 MB | updates.bcinexus.xyz |
Software Dependencies
BCILattice uses the following core libraries. Exact versions and packaging vary by release and by source/developer setup. This table is for reference and enterprise IT inventory planning.
| Dependency | Typical Version Family | Licence |
|---|---|---|
| Python | 3.10.x | PSF Licence |
| PySide6 | 6.x | LGPL v3 / Commercial |
| PyTorch | 2.2.x | BSD 3-Clause |
| scikit-learn | 1.4.x | BSD 3-Clause |
| MNE-Python | 1.6.x | BSD 3-Clause |
| NumPy | 1.26.x | BSD 3-Clause |
| SciPy | 1.12.x | BSD 3-Clause |
| Pandas | 2.2.x | BSD 3-Clause |
| MLflow | 2.11.x | Apache 2.0 |
| FastAPI | 0.110.x | MIT |
| PostgreSQL | 15.x | PostgreSQL Licence |
| ONNX Runtime | 1.17.x | MIT |
| SHAP | 0.44.x | MIT |
| Jinja2 | 3.1.x | BSD 3-Clause |
| WeasyPrint | 61.x | BSD 3-Clause |
A full Software Bill of Materials (SBOM) is available on request for enterprise security reviews. Email [email protected].
Performance Benchmarks
Representative training times on a standard 10-subject EEG motor imagery dataset (200 epochs, 128 channels, 250 Hz, 4-second trials):
| Model | CPU (8-core) | GPU (RTX 3080) | GPU (RTX 4090) |
|---|---|---|---|
| LDA (Leave-One-Subject-Out) | ~8 seconds | Same (CPU model) | Same (CPU model) |
| SVM (RBF, LOSO) | ~45 seconds | Same (CPU model) | Same (CPU model) |
| EEGNet (50 epochs, LOSO) | ~18 minutes | ~2.5 minutes | ~1.1 minutes |
| DeepConvNet (50 epochs, LOSO) | ~25 minutes | ~3.5 minutes | ~1.5 minutes |
| EEGTransformer (50 epochs, LOSO) | ~55 minutes | ~7 minutes | ~3 minutes |
Preprocessing a 10-subject EEG dataset (bandpass + notch + CAR + resampling): ~15 seconds on any modern CPU.
Large-Scale Deployments
For institutions deploying BCILattice across many workstations:
- Silent installer: Run
BCILattice-Setup.exe /S /D=C:\BCILattice(Windows) for unattended installation - Group policy / MDM: Deployable via SCCM, Intune, Jamf, or Ansible, MSI/PKG packages available on request
- External database: scope a supported PostgreSQL-compatible deployment model with the enterprise team
- On-premise BCINexus: Docker Compose stack deployable on any Linux server; supports Kubernetes
- Restricted networks: scope activation, updates, and catalog distribution during enterprise deployment planning
Contact [email protected] for enterprise deployment documentation and support.
Pre-Install Checklist
| Check | Details |
|---|---|
| ✓ 64-bit OS | A currently supported 64-bit package or source/developer setup |
| ✓ 8 GB+ RAM | 16 GB recommended for multi-subject work |
| ✓ 5 GB free disk space | 20 GB+ recommended for datasets |
| ✓ NVIDIA driver ≥ 520 | If using GPU acceleration; check with nvidia-smi |
| ✓ Firewall allows outbound 443 | Required — sign-in and licence validation both use it, as do cloud features |
| ✓ Install permissions | Depends on selected package and institutional policy |