【survey】Tenstorrent

# 重要資訊 ## Tenstorrent 代理資訊 * 基本資料 Tenstorrent是一家由Apple及英特爾（Intel）資深工程師創立的美國人工智慧（AI）晶片新創公司，近日宣布與日本政府達成合作協議，先進半導體技術中心LSTC（日本政府推動成立的研究機構）將派遣200名工程師前往美國，未來5年內在Tenstorrent培訓為晶片設計師。 * 評估 ||| |---|---| |產品定位|PCIe Card, PC, Server| |價格定位|明確,高性價比的產品| |技術文件|完整| |軟體服務|豐富且開源| |核心服務|不採用額貴的 HBM | * 主要販售硬體,提供靈活的API 讓客戶可以高度客製化, 彌補高單價的 HBM 產品 * CTO Jim Keller * [低價挑戰 NVIDIA！晶片大神 Jim Keller 曝設計關鍵是「捨棄 HBM」](https://money.udn.com/money/story/11162/8105990) * [工作經歷](https://www.linkedin.com/in/jimbkeller/details/experience/) * Tenstorrent > Intel > Tesla > AMD > Apple ## Tenstorrent 產品 ||類型|算力(TeraFLOPs)|DDR|SRAM|記憶體速度/頻寬|功耗|價格| |---|---|---|---|---|---|---|---| |Wormhole™ n150d|PCIe Card|262|12 GB | 108 MB|12 GT/S, 288 GB/S|160W|1,099| |Wormhole™ n300d|PCIe Card|466|24 GB | 192 MB|12 GT/S, 576 GB/S|300W|1,449| |TT-LoudBox|PC(Intel)|4x n300s |512GB|||1200W|12,000| |TT-QuietBox|PC (AMD)|4x n300s |512GB|||1650W|15,000| |Galaxy Wormhole|Server|9.3 Peta|384 GB | 3.8 GB||7.5KW|TBC| |Grayskull™ e75|PCIe Card(推理)|221 | 8GB | 96 MB|3.2GT/S, 102.4 GB/S|75W|600| |Grayskull™ e150|PCIe Card(推理)|332 | 8GB | 120 MB|3.7GT/S, 118.4 GB/S|200W|800| * 對比 * NV A100 => 624 (峰值) / 1248(稀疏) TFLOPs * NV H100 => 1670 (峰值) / 3341(稀疏) TFLOPs => $ 30,000 * Galaxy Server (抽取式) ![Image](https://tenstorrent.com/_next/image?url=https%3A%2F%2Fcdn.sanity.io%2Fimages%2Fjpb4ed5r%2Fproduction%2Fe4ee6af96040b66319024b157d1842ac4d73d53c-5712x2884.png&w=3840&q=75) * Cloud 架構 ![Image](https://tenstorrent.com/_next/image?url=https%3A%2F%2Fcdn.sanity.io%2Fimages%2Fjpb4ed5r%2Fproduction%2Fe7039421244037e2d149281b78cc2d8e595dfbb0-5720x3100.png&w=3840&q=75) ## Tenstorrent 技術支援/特色 * Tenstorrent 專注於AI晶片設計，並與全球半導體市場領導者合作，擁有強大的技術研發能力。 * 第二代產品不採用~**昂貴的 HBM 記憶體** * 致力解決 AI 服務器價格昂貴的問題，希望可以服務到更小型的用戶 (CTO Jim Keller) * (1) 軟體資源 ![Image](https://tenstorrent.com/_next/image?url=https%3A%2F%2Fcdn.sanity.io%2Fimages%2Fjpb4ed5r%2Fproduction%2F9a062b48bf0541a1e76b17e0d2bc4dbff63d7d08-2682x1310.png&w=3840&q=75) * TT-Metalium™ (更底層的 API 支援) * A fully open source, low level SDK * Simple, tractable * Customize * TT-Buda (上層的 Framwork ) * compile AI/ML models from several different frameworks * execute them on Tenstorrent hardware * [github](https://github.com/tenstorrent/tt-buda) ![Image](https://tenstorrent.com/_next/image?url=https%3A%2F%2Fcdn.sanity.io%2Fimages%2Fjpb4ed5r%2Fproduction%2F51ccc2e0f5c6a7cecf25cc859fb1b9624b456835-2682x1492.png&w=3840&q=75) * (2) 技術文件 * [DOC](https://docs.tenstorrent.com/) * [Starting Guide](https://docs.tenstorrent.com/getting-started/README.html) * [TT-Buda / PyBuda](https://docs.tenstorrent.com/pybuda/latest/index.html) * 是高階 Python API，將 PyTorch 模型轉成適合 Tenstorrent 硬體 * [Example](https://docs.tenstorrent.com/pybuda/latest/user_guide.html#examples-of-pybuda-use-cases) * [TT-Metalium](https://docs.tenstorrent.com/tt-metal/latest/tt-metalium/index.html) * Metalium provides developers to do more than running models, facilitating a transition from running models effortlessly out of the box, engaging in lightweight optimizations, and progressing into more sophisticated, heavyweight optimizations. * [Model Demos](https://github.com/tenstorrent/tt-metal) ![image](https://hackmd.io/_uploads/SJ-B8zyTJl.png) ![image](https://hackmd.io/_uploads/rJbU8Gkpye.png) * [TT-NN](https://docs.tenstorrent.com/tt-metal/latest/ttnn/ttnn/about.html) * user-friendly library * [TT-MLIR](https://docs.tenstorrent.com/tt-mlir/) * MLIR is a compiler infrastructure that is designed to be modular and extensible. * (3) 架構 ![Image](https://tenstorrent.com/_next/image?url=https%3A%2F%2Fcdn.sanity.io%2Fimages%2Fjpb4ed5r%2Fproduction%2F4a01dc7da5c67df97786ab922aec8f3974b49e8d-5364x2728.png&w=3840&q=75) --- # 調查資料 --- ## **(1) 時間軸** --- * 2025/03/25 評估公司 --- ## **(2) 募資** --- ||金額|日期| |---|---|---| |C |2億美元 |2021.05| |D |6.93億美元 |2024.12| * [人工智能芯片初创企业Tenstorrent再获超2亿美元C轮融资](https://www.eetrend.com/content/2021/100113089.html) * [Tenstorrent closes $693M+ of Series D funding led by Samsung Securities and AFW Partners](https://www.prnewswire.com/news-releases/tenstorrent-closes-693m-of-series-d-funding-led-by-samsung-securities-and-afw-partners-302319584.html) --- ## **(3) 獲獎** --- --- ## **(4) 新聞** --- * [半導體｜搶攻AI晶片！日本和Tenstorrent攜手培養先進技術人才](https://uanalyze.com.tw/articles/966757810) * [低價挑戰 NVIDIA！晶片大神 Jim Keller 曝設計關鍵是「捨棄 HBM](https://technews.tw/2024/07/15/us-chip-designer-tenstorrent-jim-keller-vs-nvidia/) --- ## **(5) 視頻** --- * [【人工智能】Tenstorrent能否挑战HBM市场垄断地位 | 硅仙人Jim Keller | DRAM | Blackwell GPU | Groq | 3D DRAM | Neo半导体](https://www.youtube.com/watch?v=uhHdz_aI-mo&t=47s)