Tenstorrent 推出 TT-QuietBox 2：全球首款采用全开源堆栈并支持万亿次 (Teraflop) 级推理的 RISC-V AI 工作站

搜索最新SoC解决方案
关于汽车电子...物联网...安全...音频...视频

www.design-reuse.com – May. 18, 2026 –

Liquid-Cooled Desktop System Runs Models up to 120B Parameters Locally With a Fully Open-Source Stack, Starting at $9,999

SANTA CLARA, CA -- Tenstorrent, the AI computing company led by CEO Jim Keller, today announced TT-QuietBox™ 2 (Blackhole™). This whisper-quiet, liquid-cooled AI workstation runs models up to 120 billion parameters directly at your desk, ships with an entirely open-source software stack from compiler to kernel, and starts at $9,999. It marks the industry's first desktop AI workstation built on RISC-V architecture to deliver teraflop-class inference.

The Inference Imperative

The timing matters. Inference has quietly overtaken training as the dominant AI workload, now accounting for more than 55% of cloud AI infrastructure spending at $37.5 billion - and it is still accelerating. Yet, developers running these workloads face a stark choice: pay per-token cloud fees that compound as usage scales, or buy hardware locked to proprietary stacks they cannot inspect, modify, or truly own.

QuietBox 2 is built around a different proposition: Developers doing the actual work of AI should be able to see, control, and own every layer of their compute - from silicon architecture to the compiler.

"Tenstorrent is working hard on open source AI software and we wanted to build a teraflop development system that was easy to use in a lab or office, fast and quiet. It's open top to bottom including the mechanical engineering. Build your own software or hardware. You can own your AI future," said Jim Keller, CEO of Tenstorrent.

Real Workloads Out of the Box

QuietBox 2 ships ready for quick deployment. It excels across diverse AI domains. LLMs & Coding: GPT-OSS 120B runs entirely on-device - a full 120-billion-parameter model operating privately at your desk. Llama 3.1 70B runs at 476.5 tokens per second. Qwen3-32B deploys as a private coding agent, reasoning through entire codebases without cloud token limits. Creative & Multimodal: Flux handles image generation and Wan 2.2 handles video synthesis entirely locally, ensuring creative IP remains off third-party servers. Scientific Research: Boltz-2, a biomolecular ML model, predicts the structure of a 686-amino-acid protein in just 49 seconds on a single Blackhole processor.

Silicon Innovation Without Memory + Networking Bottlenecks

Four Blackhole ASICs work as a unified mesh inside a single desk-friendly enclosure. The system features 480 Tensix cores delivering 2,654 TFLOPS at BlockFP8 precision, backed by 128 GB of GDDR6 high-speed memory and 256 GB of DDR5 system memory. This architecture integrates compute and high-density SRAM on a single die. By utilizing GDDR6 and on-chip SRAM, QuietBox 2 entirely avoids the High-Bandwidth Memory (HBM) supply shortages currently driving price hikes across the rest of the AI hardware market. The system runs on Ubuntu 24.04, plugs into a standard 120V wall outlet, and requires no rack, specialized electrical work, or server room.

Open Source at Every Layer

Every layer of QuietBox 2's software is open source. TT-Forge gives developers total visibility into graph lowering, transformation, optimization, and execution. TT-Metalium, the low-level AI SDK, provides kernel-level control with deterministic execution. TT-LLK handles low-level kernel software.

About Tenstorrent

Tenstorrent is an AI compute company. Led by CEO Jim Keller - architect of AMD Zen, Apple A4/A5, and Tesla's Full Self-Driving chip - the company builds RISC-V-based AI processors and systems for developers, enterprises, and sovereign infrastructure worldwide. Backed by Bezos Expeditions, Samsung, LG Electronics, Hyundai Motor Group, Fidelity, and others, Tenstorrent has raised over $1 billion. Learn more at tenstorrent.com.

 Back