JANCTIONDISTRIBUTED COMPUTE
日本語 Request Materials Contact Us
Home  ›  JANCTION (Distributed GPU Infrastructure / DePIN)
JANCTION — Core Infrastructure · DePIN

Distributed, so
Fast.

Jobs are assigned directly to idle GPUs that are already pooled — so you start computing without waiting on procurement. JANCTION is the distributed GPU infrastructure (DePIN) behind every JasmyLab service. This page publishes its structure and mechanisms in a verifiable form.

Connected GPUs
1,000+ (updated periodically)
Virtualization
vGPU / Docker isolation
Host-to-host transfer
GPUDirect RDMA
Consensus
PoR + PoT (two layers)
View the docs Use GPUs (GPU POOL)
System Architecture — vGPU PoolPhysical GPUs are split into vGPUs and pooled while staying isolated per container
JANCTION vGPU pool architecture: Task → Janction Server API → vGPU POOL (Docker/MPI) → CUDA Driver → Physical hosts
Source: JANCTION technical documentation (docs.janction.ai)
01

Compute plane

Tasks reach the Janction Server API over VxLAN, framework by framework. They are assigned to isolated Dockerized vGPU containers, and nodes coordinate over MPI for multi-GPU parallel execution.

02

Data plane

Physical hosts supply GPUs through CUDA drivers. Hosts transfer data directly via GPUDirect RDMA, and data is shared across IPFS / memory.

03

Service mesh

Aggregators handle resource discovery and pool formation. They are elected via on-chain staking, so operation does not depend on any single administrator.

Whitepaper — Architecture

A distributed GPU network,
built in three layers.

01

Blockchain layer

settlement / data availability
  • Handles settlement and data availability (DA) as a Layer 2 blockchain
  • Contribution evaluation algorithm: Proof of Contribution
  • Reward allocation recorded on-chain
02

Distributed resource pooling layer

pooling / scheduling
  • Physical GPUs split into vGPUs, isolated in Docker containers
  • GPU pools formed over a VxLAN overlay and organized as microservices
  • Pricing mechanism: PVCG (Procurement VCG)
03

GPU marketplace layer

front / backend services
  • Front/backend services for user and node management
  • Reputation system based on node operation history
  • Compare specs and prices across providers to procure

We can also walk you through the architecture in our materials.

Request MaterialsContact UsUse GPUs
Specifications

Technical components.

GPU virtualizationPhysical GPUs split into vGPUs and allocated in isolated Docker containers
NetworkPooling via VxLAN overlay; jobs accepted through the Janction Server API
Inter-node communicationInter-vGPU communication via MPI (multi-GPU parallel execution)
Host-to-host transferDirect data transfer via GPUDirect RDMA
Data layerDistributed data sharing via IPFS / memory (data plane)
Execution environmentGPU supply from physical hosts via CUDA drivers
ConsensusTwo-layer consensus of Proof of Resource + Proof of Task; contribution is evaluated by Proof of Contribution
PricingFair pricing and profit allocation via the PVCG (Procurement VCG) mechanism
Aggregator electionElected via on-chain staking (service mesh)

Source: JANCTION Whitepaper / docs.janction.ai

Whitepaper — Consensus

Proof that it ran
can be verified.

Miners are rewarded based on performance, while validator cross-verification and reputation scores safeguard quality.

01

Proof of Resource.

The network verifies that offered GPU resources exist and perform as claimed. Misreported resources cannot join the pool.

02

Proof of Task.

Validators cross-verify that tasks completed correctly, and rewards are allocated based on the quality of results.

03

Reputation scoring.

Node operation history accumulates as a reputation score, deterring fraud and improving allocation quality.

GPU providers
Offer idle GPUs and earn based on utilization
Aggregators
Mesh participants responsible for resource discovery and pool formation
Users
Procure GPUs from the pool and use compute services
Node participants
Participate in verifying and supervising each process via staking
Whitepaper — Pricing & Data

Pricing and data design.

PVCG PRICING

Pricing and profit allocation via PVCG

A pricing mechanism designed so that honest reporting is the winning strategy. A provider's true costs and performance, and how much value a user places on the result, are known only to themselves — PVCG takes this information asymmetry as a premise and derives fair prices mathematically.

  • Misreporting costs never pays — honest reporting is the optimal strategy
  • Participation is guaranteed not to leave you worse off
  • Prices are set to maximize network-wide surplus

For the mathematical details, see "PVCG Pricing" in the whitepaper.

DATA ECOSYSTEM

Data operations on distributed storage

Training data is stored on a distributed storage network, with distinct roles maintaining dataset quality.

  • Data providers — provide data
  • Indexers — monitor data changes
  • Curators / Delegators — ML preprocessing such as classification and labeling
Documents

The primary sources are public.

For due diligence and technical evaluation, please refer to the primary sources in the official documentation.

WhitepaperThe primary source covering design philosophy, network structure, and economic designdocs.janction.ai ArchitectureThree-layer architecture and the vGPU poolDOCS Actor ModelThe roles that make up the networkDOCS GPU MarketplaceThe trading environment and how procurement worksDOCS Proof of Resource & TaskTwo-layer consensus and verification mechanismDOCS PVCG PricingPricing that addresses information asymmetryDOCS Node DeploymentNode setup guides for Windows / macOS / LinuxDOCS RoadmapDevelopment roadmapDOCS

Technical questions welcome.

Our engineering team can walk you through what you need for technical evaluation and due diligence. Material requests are welcome, too.

Contact Us Request Materials
Request Materials Contact Us