Deployment planning
Pulsar System Requirements
Compare preliminary CPU-only, AMD GPU, and NVIDIA GPU sizing profiles for private Pulsar deployments.
Pulsar supports CPU-only, AMD GPU, and NVIDIA GPU deployments. Actual model capacity, context length, and concurrency are calibrated during installation against available memory and the selected workload.
Preliminary planning baseline
These requirements reflect current tested deployment profiles and are subject to change as installers, model packages, runtimes, and supported platforms evolve. Final sizing is confirmed during a deployment review.
Deployment profiles at a glance
Start with the profile closest to the intended deployment. Workload mix, model selection, context requirements, concurrent users, and media generation may increase the final specification.
| Requirement | CPU only | AMD GPU | NVIDIA GPU |
|---|---|---|---|
| Minimum CPU | 8 logical cores | 8 logical cores | 8 logical cores |
| Recommended CPU | 16+ logical cores | 12-16+ logical cores | 12-16+ logical cores |
| Minimum RAM | 16 GB | 32 GB | 32 GB |
| Recommended RAM | 32-64 GB | 64 GB | 64 GB |
| Minimum VRAM | Not applicable | 8 GB | 8 GB |
| Recommended VRAM | Not applicable | 16-24 GB | 24 GB |
| Minimum storage | 100 GB SSD | 150 GB SSD | 150 GB SSD |
| Recommended storage | 200 GB NVMe | 250 GB NVMe | 250 GB NVMe |
| Complete creative suite | External media runner recommended | Approved external media runner currently recommended | 500 GB NVMe and a dedicated 24 GB media GPU |
| Primary runtime | Ollama | Ollama with ROCm | Ollama or vLLM |
| Supported server OS | Ubuntu or Windows/WSL2 | Ubuntu; Windows preview | Ubuntu or Windows/WSL2 |
CPU-only deployment
CPU-only deployments provide the lowest infrastructure threshold and fit light, single-user chat workloads where GPU acceleration is not required.
Minimum
- 64-bit x86 processor with 8 logical cores
- 16 GB system memory
- 100 GB free SSD storage
- Ubuntu 22.04/24.04 or Windows 10/11 with WSL2
- Suitable for Gemma 4 E2B and light, single-user workloads
Recommended
- 16 or more logical cores
- 32-64 GB system memory
- 200 GB free NVMe storage
- 64 GB RAM for larger local models
- External media runner for image and video generation
CPU model guidance
| System RAM | Recommended models |
|---|---|
| 16-23 GB | Gemma 4 E2B |
| 24-31 GB | Gemma 4 E2B or E4B |
| 32-47 GB | Gemma 4 12B |
| 48-63 GB | Gemma 4 26B |
| 64 GB+ | Gemma 4 31B |
AMD GPU deployment
AMD acceleration is supported through a validated ROCm stack. Native Ubuntu is the supported production configuration; Windows through WSL2 remains preview-only.
Minimum
- Supported AMD GPU with 8 GB VRAM
- 8 logical CPU cores
- 32 GB system memory
- 150 GB free SSD storage
- Native Ubuntu 22.04/24.04 with a compatible ROCm stack
Recommended
- AMD GPU with 16-24 GB VRAM
- 12-16 logical CPU cores
- 64 GB system memory
- 250 GB free NVMe storage
- 24 GB VRAM for Gemma 4 26B-class models
AMD support notes
- Native Ubuntu is the supported production configuration.
- Windows AMD support through WSL2 is currently preview-only.
- The installer verifies GPU, driver, ROCm userspace, and container access before enabling acceleration.
- The current managed FLUX and Wan creative profile is CUDA-based. AMD deployments should use an approved external ComfyUI service for image and video generation.
NVIDIA GPU deployment
NVIDIA deployments provide the broadest current path for local chat and the complete managed creative suite.
Minimum
- NVIDIA GPU with 8 GB VRAM
- 8 logical CPU cores
- 32 GB system memory
- 150 GB free SSD storage
- Ubuntu 22.04/24.04 or Windows 10/11 with WSL2
Recommended for local chat
- NVIDIA GPU with 24 GB VRAM
- 12-16 logical CPU cores
- 64 GB system memory
- 250 GB free NVMe storage
- Suitable for Gemma 4 26B and selected 31B configurations
Complete creative suite
- Two NVIDIA GPUs with 24 GB VRAM each
- 16+ logical CPU cores
- 64-128 GB system memory
- 500 GB or more free NVMe storage
- One GPU assigned to chat and one assigned to ComfyUI
- Local chat, image creation, image editing, and video generation concurrently
Single-GPU operation
A single 24 GB NVIDIA GPU can run either the primary chat model or the creative runtime effectively. Sharing one GPU is supported only with reduced concurrency and runtime model loading and unloading.
GPU model guidance
| GPU VRAM | Recommended models |
|---|---|
| 8-11 GB | Gemma 4 E2B |
| 12-15 GB | Gemma 4 E4B or 12B |
| 16-19 GB | Gemma 4 12B |
| 20-23 GB | Gemma 4 26B |
| 24-31 GB | Gemma 4 26B or compatible 31B |
| 32 GB+ | Gemma 4 31B with increased context or concurrency |
Storage profiles
Storage planning must include application data, model packages, container images, generated media, temporary files, backups, and upgrade rollback space.
| Deployment | Minimum free storage | Recommended |
|---|---|---|
| Application with external AI services | 50 GB | 100 GB |
| Self-contained application and local chat | 100 GB | 150-250 GB |
| Chat plus image creation and editing | 200 GB | 350 GB |
| Chat, image, edit, and video | 300 GB | 500 GB+ |
Media package allocation
The approved FLUX and Wan media package contains approximately 44 GiB of model files before chat models, container images, generated media, temporary files, backups, and upgrade rollback space are included.
Additional requirements
- SSD storage is required; NVMe is strongly recommended.
- Reliable broadband is needed for initial container and model downloads.
- Outbound HTTPS access is required during installation.
- Some approved models require Hugging Face authentication and license acknowledgement.
- NVIDIA requires a compatible driver and container GPU runtime.
- AMD requires a compatible ROCm GPU, driver, and container runtime.
- Published model context limits do not guarantee that the maximum context will fit a particular host.
- Production systems should retain at least 15% GPU memory or 3 GiB as operational reserve.
Installation-time calibration
Pulsar automatically calibrates context length and concurrency to the selected workload and available memory. The deployment review confirms the operating reserve, target models, media profile, and expected concurrency before production acceptance.
Private deployment consultation
Review Pulsar against your environment.
Bring the infrastructure, security boundaries, model runners, and use cases. The Pulsar team will map the appropriate deployment path.