Skip to content

SiliconLanguage/tensorplane

Repository files navigation

TensorPlane: The Recursive Self-Improving AI Foundry

https://img.shields.io/badge/Architecture-Compute--Continuum-blue

https://img.shields.io/badge/Foundry-Recursive--Self--Improving-orange

https://img.shields.io/badge/Stack-Go%20%7C%20Rust%20%7C%20Cpp%20%7C%20RISC--V-green

Executive Vision & Overview

TensorPlane (also operating as dataplane-emu) is an end-to-end Recursive Self-Improving AI Foundry. It is engineered as a hardware-accurate data plane emulator and orchestration framework designed to span the entire compute continuum: from Go-based cloud orchestration to C++/Rust kernel-bypass storage, down to bare-metal RISC-V execution.

The system utilizes a specialized Multi-Agent System (MAS) driven by the Model Context Protocol (MCP) to autonomously monitor, evaluate, and rewrite its own compute kernels. This recursive loop ensures the continuous elimination of I/O and latency bottlenecks without human intervention.

The "Team of Agents" Architecture

The /team_of_agents/ directory implements "Scalable Agency" through a series of hyper-specialized pools:

  • Orchestrator: The AgentOps manager responsible for task routing and MAS state.
  • Ops (Telemetry & Triage): Monitors P99 tail latency via eBPF (Cilium/Hubble), manages GPU fractioning (ops_resource_optimizer), and handles thermal/DVFS scaling (ops_power_monitor).
  • Dev (The Builders):
    • /dev_cpp_dataplane/: SPDK and io_uring high-performance implementations.
    • /dev_rust_kernel/: eBPF and XRP (eXpress Resubmission Path) hooks.
    • /dev_go_controlplane/: Kubernetes and FaaS orchestration logic.
    • /dev_bare_metal/: RISC-V Assembly and hardware-software co-design.
  • Eval & Deployment: Agents that sandbox, benchmark, and safely hot-swap compiled C++/CUDA/Triton kernels into the live data plane.

The Compute Continuum & Tech Stack

The Zero-Copy Storage Plane

TensorPlane leverages the Rust (usrbio crate) and C++ for lock-free Submission/Completion Queues (SQ/CQ). By bypassing the kernel via SPDK and utilizing eBPF XRP hooks, the system achieves near-theoretical hardware performance.

The Bare-Metal Silicon Layer

Targeting the RISC-V "Snitch" pseudo dual-issue processor, the foundry utilizes Stream Semantic Registers (SSR) to elide load/store instructions. Performance is further tuned via the Zawrs (Wait-on-Reservation-Set) and Zihintntl (Non-Temporal Locality) extensions for energy-efficient polling.

The Control Plane

A distributed FaaS (Serverless Lambda) deployment of MCP servers managed by Go, providing a scalable gateway for agentic orchestration.

Infrastructure as Code (IaC) & Provisioning

The /provision/ directory ensures the Foundry is fully reproducible across hyperscale environments:

/provision/
├── ai_common/      # Foundational GPU dependencies (CUDA 12.x / ROCm 7).
├── ai_inference/   # vLLM/SGLang engines with PagedAttention.
├── ai_training/    # Distributed Ray clusters for multi-node LLM training.
├── dev_control/    # Go toolchain and eBPF kernel headers for observability.
└── dev_data/       # AWS Graviton (ARM64) setup with 2MB hugepages and LSE atomics.

Prior Art & Inspiration

TensorPlane incorporates paradigms from: * DeepSeek 3FS: For RDMA and zero-copy NVMe separation logic. * Intel DAOS: For transparent POSIX interception models. * MemPool/Snitch: For RISC-V tightly-coupled memory and SSR architectures.

About

An enterprise-grade AI data plane accelerator achieving true zero-copy I/O. Inspired by DeepSeek 3FS USRBIO, it separates command and data queues for GPUDirect RDMA streaming and SmartNIC/DPU offloading, bypassing POSIX bottlenecks for hyperscale AI workloads.

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

 
 
 

Contributors