At a glance
| Booth | Q6158 |
| Country | US |
| Website | tenstorrent.com/vision |
Company profile
Tenstorrent Inc. builds computers for AI and describes itself as a next-generation computing company headquartered in the United States, with offices in Austin, Texas, and Silicon Valley as well as global offices in Toronto, Belgrade, Seoul, Tokyo, and Bangalore. Its teams bring together expertise in computer architecture, ASIC design, RISC-V technology, advanced systems, and neural network compilers, giving the company a combined hardware, systems, and software foundation for AI computing. Chief Operating Officer Keith Witek leads operations, legal, HR, finance, fundraising, IT, strategic deals, compliance, licensing, IP, automotive, and robotics, while the Experience organization led by Christine Blizzard spans brand, product, communications, go-to-market activities, and customer success. A co-founder and Senior Fellow leads systems engineering and low-level foundational software teams: systems engineering designs, builds, and deploys advanced systems at large scale, while foundational software provides the lower layers of the AI stack, abstracts hardware details, and delivers APIs and tools for high-performance kernel development. He has also worked across Tensix processor architecture, RTL, verification, and post-silicon validation, reflecting experience that extends from processor design through implementation and validation. Senior Fellow Jasmina Vasiljevic leads the Pathfinding team, which explores new programming models and prototypes compilation methods, and her AI and HPC work is applied to making Tenstorrent's RISC-V and AI architecture more accessible to AI customers, researchers, and hobbyists. Physical Engineering leader Dan Bailey is associated with high-performance flip-flop design and global clock-distribution methodologies adopted across the industry, and the company profile notes that he has nearly 40 publications and more than 20 patents. Chief CPU Architect and Senior Fellow Wei-Han Lien leads work defining a high-performance RISC-V CPU, fabric, system caching, and a high-performance memory subsystem for Tenstorrent's heterogeneous AI and HPC platform, and also leads chiplet-architecture definitions for scalable, configurable, composable SiP with power, security, and system-management compatibility. Tenstorrent states that its IP is transparent, its architectures are open, and its software is open source, allowing users to edit, select, fork, and own elements of their silicon future; its listed locations span North America, Asia, and Europe.
Exhibits
Blackhole® PCIe boards are designed for high-performance AI processing and scalable deployment, with 16 big RISC-V cores and up to 32 GB of GDDR6 memory per chip. Blackhole® p100a uses the Blackhole® Tensix Processor with 120 Tensix Cores and 28 GB of GDDR6, operates at up to 300 W in an active-cooled desktop-workstation form factor, and is supported by Tenstorrent's fully open source software stack. Blackhole® p150a also has 120 Tensix Cores, carries 32 GB of GDDR6, operates at up to 300 W with active cooling, and adds four passive QSFP-DD 800G ports so multiple cards can be linked to pool memory and improve performance. The Blackhole comparison lists p100a, p150a, and p150b with a 1.35 GHz AI clock, 16 big RISC-V cores, 180 MB SRAM, 16 GT/sec memory speed, PCI Express 5.0 x16, and 664 TeraFLOPS in BLOCKFP8; p100a has 448 GB/sec memory bandwidth while p150a and p150b list 512 GB/sec. The three Blackhole boards are listed at 300 W total board power and dimensions of 42 mm x 270 mm x 111 mm; p100a is active-cooled with no listed card-to-card connectivity, p150a is active-cooled with four passive QSFP-DD 800G links, and p150b is passive-cooled with the same four-link connectivity. Wormhole™ n150d uses a scalable Wormhole™ Tensix Processor at up to 160 W and can join a multichip mesh for workstations and servers such as Galaxy, with TT-Forge™ for high-level development and TT-Metalium™ for low-level development. Wormhole™ n300d integrates two Wormhole™ Tensix Processors at up to 300 W, supports the same multichip-mesh approach and broad data-precision formats, and includes a 2.5-slot active cooling solution with support from Tenstorrent's fully open source software stack. Interconnect accessories include a passive 0.5 m QSFP-DD 800G cable for Blackhole®-based boards and systems, a validated 0.6 m QSFP-DD 400G cable for Wormhole™ systems, and the Warp 100 Bridge in dual-slot TX-01002 and triple-slot TX-01004 versions for internal Wormhole add-in-board interconnection. Tenstorrent Galaxy™ is a scalable, ultra-dense AI-compute server for training and inference workloads including video, speech, language, and code, and its TT-Forge open-source MLIR-based compiler is positioned for broad model support. A Tenstorrent Galaxy™ Blackhole system contains 32 Blackhole® ASICs and is specified at 23 PFLOPS Block FP8, 6.2 GB of accelerator SRAM at 2.9 PB/s, 1 TB GDDR6 at 16 TB/s, and ten 400 GbE links per ASIC for a 32 TB/s accelerator fabric. For scale-out, Galaxy Blackhole lists up to 56 QSFP-DD 800 GbE ports for 11.2 TB/s, and four-system superclusters can extend into configurations from 4 to 36 or more Galaxy systems for AI video generation, large-scale LLM inference, and private AI infrastructure. The server host configuration lists one AMD EPYC 9004 Zen 4 CPU with up to 32 cores, up to 576 GB DDR5-4800 ECC RDIMM, an OCP NIC 3.0 PCIe Gen5 x16 interface with a default 2 x 200 GbE configuration, dedicated 1 GbE management, and NVMe storage for the operating system and internal expansion. Galaxy is a 6U air-cooled rackmount system measuring 263.4 mm high, 446.8 mm wide, and 884.5 mm long, with a listed weight of 119 kg, average power usage of 8 to 10 kW, a 12 kW maximum and configurable maximum up to 14.5 kW, and an operating range of 10 to 35 °C. Tenstorrent states that 90% of models from HuggingFace work on the platform across LLMs, image generation, speech, vision, embeddings, encoders, and other workloads, while Tensix Neo™ is designed to scale from one chip to thousands under one programming model. The Galaxy software path is open source from compiler to kernel, allowing a model to be compiled and run directly or kernels to be tuned at a lower level, and the system is positioned to use the same general-purpose AI platform for decode and prefill in latency-sensitive and long-context inference workloads.
Capabilities and products
- AI computing computers
- Blackhole® PCIe compute boards
- Wormhole™ PCIe compute boards
- Galaxy™ scalable AI-compute servers
- open-source AI software stack