Tenstorrent
Tenstorrent delivers open, sovereign AI compute: Blackhole accelerator cards, whisper-quiet liquid-cooled workstations, scale-out TT-QuietBox servers, licensable Galaxy IP, and TT-Forge, an MLIR-based compiler for PyTorch, JAX, and ONNX—built to run large models efficiently from desks to data centers.
Overview
Build in familiar frameworks, compile with TT-Forge, and deploy to Blackhole accelerators inside workstations or TT-QuietBox servers. Select passive, active, or liquid cooling to fit constraints, and keep workloads on-prem while preserving a portable, open toolchain across development and production.
Hardware, IP, and Compiler Stack
Tenstorrent fits ML engineers, researchers, and MLOps teams who need deterministic performance, on-prem sovereignty, and a clear path from prototype to rack-scale. It also suits OEMs and silicon teams licensing Galaxy IP to specialize for targeted AI workloads. Labs, startups, and enterprises that value open tooling, inspectable kernels, and a responsive developer community will appreciate the velocity gains from TT-Forge and the practical deployment envelope of Blackhole cards, deskside workstations, and TT-QuietBox servers.
- Compile PyTorch, JAX, and ONNX models with TT-Forge into optimized hardware kernels.
- Run up to 120B-parameter models on whisper-quiet, liquid-cooled deskside workstations.
- Scale the same architecture into TT-QuietBox servers for sovereign, production AI deployments.
- License Galaxy IP to tailor accelerators for specific workloads without vendor lock-in.
- Choose passive, active, or liquid cooling to balance thermals, acoustics, and density.
Why engineers choose Tenstorrent
Editorial perspective
Start by installing TT-Forge (public beta) and reviewing the docs and examples in the developer hub. Bring PyTorch, JAX, or ONNX models, compile with TT-Forge, and validate on a Blackhole-powered workstation before promoting to TT-QuietBox servers. Join the community Discord for support, file issues or PRs on GitHub, and track bounty tasks to accelerate kernel improvements. When ready, select cooling and chassis options that match your environment, then operationalize builds using your existing MLOps pipelines. TT-Deploy demos showcase approaches for packaging, scheduling, and scaling Tenstorrent workloads.
AI still obeys the old laws of compute; Tenstorrent aligns silicon, software, and systems to deliver practical, sovereign performance.
Getting started
Tenstorrent’s advantage is coherence: a consistent software toolchain, flexible thermals, and options from licensable IP to turnkey systems. Teams can prototype quietly on a desk, keep sovereignty by staying on-prem, and scale without retooling. With open-source TT-Forge and a responsive developer community, performance tuning remains transparent and iterative.
Open the tool and review its core product experience.
Create your account or access your existing workspace.
Use your own task to judge speed, quality, and fit.
Check similar AI tools before making a final decision.
Tenstorrent
Tenstorrent designs AI accelerators, systems, and licensable processor IP with an open, developer-first philosophy. Its team builds transparent architectures and open-source software, enabling developers to edit, select, fork, and own their silicon future while deploying performant, sovereign AI from desks to data centers.
Tenstorrent is an end-to-end AI compute platform spanning silicon IP, accelerator cards, quiet workstations, and scale-out servers. Its Blackhole hardware powers local model execution, while TT-QuietBox systems enable sovereign, on-prem deployments. Developers target the stack using TT-Forge, an MLIR-based, open-source compiler that ingests PyTorch, JAX, and ONNX graphs to generate performant kernels. For OEMs and hyperscalers, Galaxy IP provides licensable, flexible processor designs to tailor specific workloads without vendor lock-in. Practical engineering touches—passive, active, and liquid cooling options—let teams right-size thermals and acoustics. A developer hub, public docs, GitHub repos, Discord, and a paid bounty program tighten the feedback loop from kernel to silicon. The result is a transparent, developer-first path to run models up to 120B parameters on deskside systems and scale the same stack into production racks.
tenstorrent.com/en
Comments (0)
No Comments Found