128GB
vram
3.4x
ngcSpeedup
$0
perExperimentCost
What it is
The DGX Spark is the foundation of the personal research lab: 128GB of unified VRAM that enables experiments typically reserved for well-funded research teams.
Why DGX Spark
- GB10 Blackwell GPU with 128GB unified memory
- Runs models that don't fit in consumer GPU memory
- NGC containers provide 3.6x speedup over vanilla PyTorch
- Local execution means no cloud costs, no data privacy concerns
What It Enables
- Training OncoVLM's 4B parameter model from scratch
- Running multiple large language models simultaneously for multi-agent experiments
- Vector indexing of entire knowledge bases (32,000+ chunks)
- Continuous operation of ARIA autonomous research agent
Infrastructure Integration
- Syncs with MacBook Pro for development workflow
- Raspberry Pi handles production trading bots
- Together they form a distributed personal compute cluster
Cost Analysis
Equivalent cloud compute would cost $50K+/year. The DGX Spark pays for itself in months of heavy research use.
Timeline
How it went
- 2024-09-15The Lab Gets Real. DGX Spark arrives. 128GB VRAM. A personal GPU research lab rivaling small teams.
- 2024-12-01The Distributed Lab. Mac + DGX + Pi working together. Vector indexing on DGX. Trading bots on Pi. Development on Mac.
Connected work
What it touches
Enables
OncoVLM
Oncology multimodal AI models at three scales
Read the story →
Enables
ARIA
Autonomous research that runs for months. One kernel, three organizations, a published paper.
Read the story →
Enables
Autonomous Researcher
Self-running ML experiments on DGX
Project page →
Enables
Nanochat Experiments
Training LLMs from scratch on domain data
Project page →
Enables
Text Diffusion
Discrete diffusion for text generation
Project page →
Related
Local AI Stack
M4 Max multi-model inference
Project page →
Related
Gemma Fine-tuning
Fine-tuning Gemma 3 4B for meeting intelligence
Project page →
Enables
Dendrite
O(1) fork LLM inference engine for tree-structured reasoning
Project page →
Enables
AI Explorer (AIXplore)
Automated research digital garden with 80+ articles
Project page →
Related
Obsidian Knowledge System
2,353 documents with semantic search across 32K vectors
Project page →
Enables
DeepCode Simplified
3-agent paper-to-code matching 10-agent performance
Project page →
Enables
AI-Wire
AI-news intelligence engine feeding a public wire
Project page →
Enables
MakerVideo
Turns blog posts and stats into branded short videos
Project page →
Related
Headroom
Menu-bar gauge for your remaining AI coding subscriptions
Project page →