Skip to content
FrankX.AI
Research Hub/Open Silicon & RISC-V AI Accelerators (Tenstorrent)

Open Silicon & RISC-V AI Accelerators (Tenstorrent)

Wormhole, Blackhole, open-source ISA architectures, chiplet scaling, and decoupling AI from closed hardware ecosystems

TL;DR

To prevent global AI compute from becoming monopolized by closed proprietary hardware ecosystems (NVIDIA CUDA), open silicon initiatives like Tenstorrent leverage open RISC-V instruction set architectures (ISA). By pairing programmable RISC-V cores with dedicated tensor math engines and open-source compilers (TT-Metalium), open silicon delivers cost-effective, royalty-free hardware scalability.

Updated 2026-08-186 source references4 claims indexed

Research briefs like this, when the evidence is ready. Source links, limitations, and open questions.

Subscribe

RISC-V

Open-source, royalty-free instruction set architecture foundation

RISC-V International

Tenstorrent

Wormhole & Blackhole chiplet-based AI processors

Jim Keller & Tenstorrent Architecture

TT-Metalium

Open-source low-level kernel programming framework bypassing proprietary CUDA

Tenstorrent Open Source

Chiplets

Modular silicon packaging reducing manufacturing fab costs

Open Compute Project (OCP)
01

The Open RISC-V Architecture for AI

Proprietary ISAs (like x86 and ARM) require expensive licensing and restrict microarchitectural modification. RISC-V provides an open, modular instruction set that allows designers to add custom matrix and vector math extensions directly to the CPU cores.

Royalty-Free Open Standard

Open

Eliminates licensing gatekeeping, enabling global universities, startups, and sovereign nations to manufacture custom AI silicon.

Custom Matrix Extensions

Extensions

Extends standard RISC-V instructions with specialized vector and matrix multiply registers.

Heterogeneous Compute Arrays

Cores

Grids of hundreds of small, power-efficient RISC-V cores managing local data movement and tensor math.

02

Tenstorrent Wormhole & Blackhole Silicon Architecture

Led by legendary chip architect Jim Keller, Tenstorrent builds modular AI processors that combine high-speed 2D Torus network-on-chip (NoC) topologies with low-cost standard GDDR6 and LPDDR memory.

2D Torus Network-on-Chip (NoC)

NoC

Every compute tile acts as an autonomous router, passing data packets across the silicon die with zero centralized bus contention.

Cost-Effective GDDR6 Memory

Memory

Uses affordable consumer-grade GDDR6 graphics memory instead of expensive, supply-constrained HBM.

Direct Ethernet Chip-to-Chip Links

Ethernet

Integrates 100GbE physical layers directly on the processor die, connecting thousands of chips without external PCIe switches.

03

Breaking the CUDA Software Moat (TT-Metalium & PyTorch Native)

NVIDIA's primary moat is not just silicon, but the 15-year CUDA software ecosystem. Open silicon initiatives bypass CUDA by providing open-source C++ and Triton-compatible compiler runtimes that compile PyTorch models directly to bare-metal hardware.

TT-Metalium (Bare-Metal Programming)

Metalium

An open-source C++ framework that gives developers direct control over RISC-V cores, SRAM buffers, and NoC data movement.

Direct PyTorch & vLLM Integration

PyTorch

Runs standard Hugging Face and vLLM models without rewriting custom hardware-specific code.

Democratized Compute Sovereignty

Sovereignty

Allows nations and enterprises to manufacture and deploy sovereign AI hardware free from single-vendor lock-in.

Key Findings

1

Open-source RISC-V architectures provide a viable path to break single-vendor proprietary hardware monopolies in AI acceleration.

2

Using standard GDDR6 memory with direct-on-die Ethernet links delivers high-performance AI inference at 3x lower silicon hardware cost than HBM clusters.

3

Tenstorrent's 2D Torus Network-on-Chip (NoC) allows linear performance scaling from single PCI cards up to thousands of interconnected server nodes.

4

Open-source compiler frameworks (like TT-Metalium) enable developers to write custom tensor kernels with full hardware transparency.

5

Sovereign nations are increasingly adopting RISC-V AI designs to guarantee national technological independence and supply-chain resilience.

Research Transparency

Limitations

  • Software compiler maturity is still evolving compared to NVIDIA's highly polished 15-year CUDA ecosystem.
  • Requires active community and enterprise developer adoption to expand optimized kernel libraries.

What We Don't Know

  • ?The exact market share inflection point where open RISC-V AI compilers achieve full parity with CUDA performance out of the box.
  • ?Optimal packaging interconnect standards for multi-vendor modular chiplet interoperability.
Evidence Grade:Grade A(Backed by Tenstorrent technical architecture publications, RISC-V International standards, and open-source compiler repository benchmarks.)

Frequently Asked Questions

Tenstorrent is an AI hardware and software company building open-architecture AI processors using RISC-V. It is led by Jim Keller, the legendary microprocessor architect behind Apple A4/A5, AMD Zen, and Tesla Autopilot silicon.

From research to practice

Learn these tools hands-on

The research maps the landscape. These portals curate the videos, docs, and experts to actually build with the platforms it covers.