📊Bloomberg Technology•較早收集於 41m
Intel 與 Qualcomm 考慮收購 AI 新創公司 Tenstorrent
💡大型晶片製造商正尋求 Nvidia 以外的 AI 硬體替代方案，Tenstorrent 正是目前的目標。
⚡ 30-Second TL;DR
有什麼變化
Tenstorrent 目前吸引了來自 Intel 與 Qualcomm 的早期收購興趣。
為什麼重要
潛在的收購案將顯著改變 AI 硬體格局，為 Intel 或 Qualcomm 提供能與 Nvidia 抗衡的專業架構。這也標誌著 AI 基礎設施堆疊正朝向垂直整合發展。
下一步行動
監控 Tenstorrent 的硬體發展藍圖，並將其效能與現有的 Nvidia H100/B200 進行基準測試比較，以評估未來的基礎設施採購。
誰應關注：Founders & Product Leaders
關鍵要點
- •Tenstorrent 目前吸引了來自 Intel 與 Qualcomm 的早期收購興趣。
- •該新創公司被視為挑戰 Nvidia 與 AMD 等市場主導者的潛在對手。
- •此趨勢反映了業界確保自有 AI 矽晶片能力的戰略需求。
🧠 深度解析
Web-grounded analysis with 15 cited sources.
🔑 增強重點摘要
- •Tenstorrent achieved unicorn status in 2021 with a $1 billion valuation after raising over $200 million in a Series C round led by Fidelity.
- •The company has raised a total of $1.18 billion across 10 funding rounds, with a significant Series D round of $693 million in December 2024, valuing the company at $2.6 billion, backed by investors including Jeff Bezos, Samsung Securities, and Hyundai Motor Group.
- •Tenstorrent differentiates itself by embracing an open-source approach, utilizing RISC-V for its instruction set architecture and providing an open-source compiler stack (TT-Metal/Buda), contrasting with Nvidia's proprietary CUDA ecosystem.
- •Tenstorrent's technology focuses on an asynchronous, networked "Grayskull" architecture and subsequent "Wormhole" and "Blackhole" designs, prioritizing efficiency and scalability through explicit data movement and dynamic sparsity.
- •In July 2025, Tenstorrent acquired Blue Cheetah, an analog IP design company, and opened its Open Chiplet Architecture (OCA) specification, signaling a strategic move towards a composable chiplet ecosystem.
🛠️ 技術深入
- Architecture Philosophy: Tenstorrent's design philosophy emphasizes an open ISA (Instruction Set Architecture) built on RISC-V, an open-source compiler stack, and a chip design centered on explicit data movement rather than a hardware cache hierarchy, allowing for an auditable compute stack.
- Core Components:
- Tensix Cores: Proprietary AI processing units that exploit dynamic sparsity to improve performance-per-watt.
- RISC-V CPU Cores: High-performance RISC-V CPU cores, including the "Ascalon" series, which are licensable IP and positioned against Arm's Neoverse cores for data centers.
- Chip Architectures:
- Grayskull: An asynchronous, networked architecture designed for efficiency and scalability.
- Wormhole: Features a RISC-V data movement processor, an 8x8 BF16 tile multiply-accumulate unit (Matrix engine), and a vector unit for element-wise operations.
- Blackhole: Successor to Wormhole, designed with up to 120 Tensix compute cores and 16 big RISC-V cores (in the p100 card), integrating CPU and AI acceleration on a single die. It includes 28 GB GDDR6 memory and a PCIe 5.0 x16 interface.
- Interconnect: Utilizes a mesh-network chip fabric for on-chip communication to reduce memory bottlenecks.
- Software Stack:
- Buda/Metalium: Open-source AI compiler stacks (MIT-licensed) that enable native execution of PyTorch and JAX models, aiming to improve developer portability and reduce total cost of ownership.
- Manufacturing: Employs a modular "Chipletoid" architecture and plans to use Samsung Foundry's 4nm and SF2 (2nm) nodes for production, with initial chips manufactured by GlobalFoundries.
- Chiplet Strategy: Acquired Blue Cheetah in July 2025 to bring die-to-die interconnect IP (BlueLynx, supporting BoW and UCIe standards) in-house and opened its Open Chiplet Architecture (OCA) specification to foster a composable chiplet ecosystem.
🔮 前景展望AI analysis grounded in cited sources
The potential acquisition could accelerate the adoption of RISC-V and open-source AI hardware in the data center.
If Intel or Qualcomm acquires Tenstorrent, their significant resources could push Tenstorrent's open-source RISC-V and software stack into broader commercial deployment, offering a strong alternative to proprietary ecosystems.
Increased competition in the AI chip market could lead to more specialized and energy-efficient solutions.
Tenstorrent's focus on efficiency, dataflow architecture, and dynamic sparsity, if scaled by a major player, could drive other incumbents to innovate further in power-efficient and specialized AI accelerators beyond general-purpose GPUs.
The move towards chiplet architectures will intensify among AI hardware developers.
Tenstorrent's acquisition of Blue Cheetah and its Open Chiplet Architecture initiative highlight a strategic industry shift towards modular chip designs, which major players like Intel and Qualcomm would likely leverage and further develop.
⏳ 時間線
2016
Tenstorrent founded in Toronto.
2020
Jim Keller joined as CTO.
2021 H2
Grayskull AI processor launched; company achieved unicorn status with $200M Series C funding.
2023
Jim Keller became CEO; closed $100M strategic funding round with Hyundai and Samsung.
2024-12
Raised $693M in Series D funding, valuing the company at $2.6 billion, with Jeff Bezos among investors.
2025-07
Acquired Blue Cheetah and opened its Open Chiplet Architecture (OCA) specification.
📎 來源 (15)
Factual claims are grounded in the sources below. Forward-looking analysis is AI-generated interpretation.
📰
AI 週報
閱讀本週精選 AI 大事摘要 →
👉相關動態
AI 策展新聞聚合。所有內容版權歸原始發布者所有。
原始來源： Bloomberg Technology ↗