Wormhole
RISC-V AI accelerator with mesh networking
Wormhole is a series of PCIe accelerator cards from Tenstorrent, designed for artificial intelligence workloads. It uses RISC-V based Tensix cores and GDDR6 memory, offering an open-source software stack for developers to build scalable AI systems.
- Uses RISC-V derived Tensix Cores for AI processing
- Supports open-source software stacks TT-Forge and TT-Metalium
- n300d card contains two processors with up to 300W power
- Can network multiple chips into a scalable multichip mesh
- Features GDDR6 memory and local SRAM caches
What it is
Wormhole is a family of processor cards built by Tenstorrent for high performance artificial intelligence tasks. The chips rely on Tensix Cores, which are derived from the RISC-V instruction set architecture. These processors include compute units, network-on-chip structures, local cache memory, and smaller control cores known as baby RISC-V cores. This design allows for efficient data movement across the chip. The company states that Wormhole offers better cost performance than traditional graphics processing units and supports various data precision formats.
How it came about
Tenstorrent, led by CEO Jim Keller, developed Wormhole to challenge established players in the AI server market using RISC-V technology. The company launched developer kits featuring these cards to allow engineers to test the architecture. By providing hardware that is currently available, Tenstorrent aimed to distinguish itself from startups with only theoretical designs. The launch included pre-built workstation options to facilitate immediate development and testing of the new silicon.
Inside the chip
The architecture centres on Tensix Cores, which handle the primary computational load. These are supported by baby RISC-V cores for control tasks. Each chip integrates a network-on-chip to manage internal communication and local SRAM caches for fast data access. The system uses GDDR6 memory for larger storage needs. A key feature is the ability to link multiple chips into a multichip mesh, enabling scalable performance for servers and workstations. High-speed Ethernet links, such as 400GbE on the n300s, facilitate this external connectivity.
How it was made
Specific manufacturing details such as the foundry or process node size are not provided in the available information. The cards are assembled into PCIe form factors. The n150d model operates at up to 160 watts and includes an active cooling kit. The n300d model contains two processors on a single card, operating at up to 300 watts, and uses a 2.5-slot active cooling solution. Interconnects like the Warp 100 Bridge allow for internal linking between boards.
Where it went
These processors are intended for use in workstations and servers, such as the Galaxy system. They target developers who need scalable AI processing power. The cards can be used individually or linked together to pool memory and improve performance. Tenstorrent offers them as standalone PCIe boards or integrated into pre-built systems. These systems may use Intel Xeon or AMD EPYC processors as hosts, allowing multiple developers to share the resources or one user to harness all available ASICs.
Versions and family
The main variants are the n150 and n300 series. The n150d features a single Wormhole Tensix Processor. The n300d includes two of these processors on one card. The n150s model is noted for having more compute per ASIC, while the n300s provides two ASICs on the card. Connectivity options vary, with the n300s including a local 400GbE link. Cables such as QSFP-DD 400G and 800G are used to connect these boards in a mesh configuration.
Why it matters
Wormhole provides an alternative to dominant GPU providers, offering a path for companies to reduce costs in AI development. By making the hardware and its software stack fully open source, Tenstorrent grants developers direct access to the underlying metal. This approach aims to lower the barrier for adopting non-NVIDIA hardware. While it may not immediately overturn existing market leaders, it establishes Tenstorrent as a viable provider of physical AI silicon rather than just conceptual designs.
Questions people ask about Wormhole
What instruction set does Wormhole use?
Wormhole uses cores derived from the RISC-V instruction set. The main compute engines are called Tensix Cores, and they are assisted by smaller baby RISC-V cores for control and data movement tasks within the chip.
How much power do the Wormhole cards consume?
The power consumption depends on the specific model. The n150d variant operates at up to 160 watts. The n300d, which contains two processors on a single card, operates at up to 300 watts and requires a larger active cooling solution.
Can multiple Wormhole cards be connected together?
Yes, the processors can be networked into a multichip mesh. This allows users to link multiple cards to pool memory and improve performance. Connections use high-speed interfaces like QSFP-DD cables and internal bridges such as the Warp 100.
Good reading and sources
There is no English Wikipedia article on Wormhole yet, so the facts on this page come from the sources listed above: tenstorrent.com/hardware/wormhole, servethehome.com/tenstorrent-wormhole-developer-kits-launched-risc-v/. Some are the subject's own pages, so its claims about itself are reported as its claims. EveryChip wrote the text in its own words, with AI assistance, and compares it with the sources to make sure no passage is copied.
It may contain mistakes, and figures change over time. Nothing here is investment or buying advice. If something is wrong, tell us. Last refreshed 2026-11-14.