# Wormhole (Tenstorrent)

Wormhole is a series of AI accelerator cards developed by Tenstorrent, designed for high-performance deep learning and large language model inference and training workloads.

Wormhole is a family of AI accelerator cards produced by Tenstorrent, a fabless semiconductor company. These cards are designed to accelerate [artificial-intelligence](https://www.wikiprompt.org/wiki/artificial-intelligence) and [machine-learning](https://www.wikiprompt.org/wiki/machine-learning) workloads, particularly [deep-learning](https://www.wikiprompt.org/wiki/deep-learning) models such as [large-language-model](https://www.wikiprompt.org/wiki/large-language-model)s and [transformer](https://www.wikiprompt.org/wiki/transformer)-based architectures. The Wormhole series targets both inference and training tasks, offering a scalable architecture that can be linked together for larger deployments. Tenstorrent positions Wormhole as a flexible alternative to GPUs, emphasizing programmability and a focus on dataflow-based computation.

The first generation, Wormhole, was announced in 2020 and began shipping in 2021. It was followed by Wormhole n150 in 2023, a single-card version, and Wormhole n300, a dual-card configuration. These later cards are built on a 6nm process and feature a RISC-V-based control processor. The architecture is designed to handle sparse and dense tensor operations efficiently, with a focus on reducing memory bottlenecks common in traditional GPU designs.

## Architecture and Design

The Wormhole accelerator is built around a grid of Tensix cores, each containing a RISC-V CPU, a matrix multiply unit, and a vector engine. This heterogeneous design allows for flexible execution of various neural network operations. The cores are interconnected via a high-bandwidth, low-latency mesh network, enabling efficient data sharing and scaling across multiple chips. Unlike GPUs, which rely on a SIMT (Single Instruction, Multiple Thread) model, Wormhole uses a dataflow architecture that can be more efficient for certain workloads, particularly those with irregular memory access patterns.

Each Tensix core includes dedicated memory, reducing the need for frequent off-chip memory access. The n150 card features 72 Tensix cores, while the n300 has 144 cores across two dies. The cards support a range of data types, including FP32, FP16, BF16, and INT8, allowing for precision flexibility in training and inference. The architecture also includes a dedicated Ethernet interface for scaling out to multi-card systems, using a protocol that supports direct memory access between cards.

## Performance and Specifications

The Wormhole n150 delivers up to 100 teraflops of FP16 compute and 200 teraops of INT8 compute, with 16 GB of GDDR6 memory providing 256 GB/s of bandwidth. The n300 doubles these figures, offering 200 teraflops of FP16 and 400 teraops of INT8, with 32 GB of memory. Power consumption is rated at 150W for the n150 and 300W for the n300, making them relatively power-efficient compared to some competing accelerators. The cards use a PCIe Gen 4 x16 interface for host connectivity, and can be interconnected via Ethernet for multi-card configurations.

In benchmark tests, Wormhole has shown competitive performance on [neural-network](https://www.wikiprompt.org/wiki/neural-network) inference tasks, particularly for transformer-based models. Tenstorrent has published results showing linear scaling in performance when adding multiple n150 cards, a key feature for large-scale deployments. The cards also support model parallelism and data parallelism, enabling efficient training of large models across multiple accelerators.

## Software Stack

Tenstorrent provides a comprehensive software stack for Wormhole, including a low-level driver, a runtime library, and higher-level frameworks. The primary programming model is based on a C++ API, but there is also support for popular machine learning frameworks. The company has developed a custom compiler that translates models from frameworks like PyTorch and ONNX into optimized code for the Tensix cores. This compiler performs graph-level optimizations, including operator fusion and memory planning, to maximize performance.

For users familiar with GPU programming, Tenstorrent offers a programming model that abstracts away some of the low-level details, while still allowing advanced users to write custom kernels. The software stack includes a debugger and profiling tools to aid in development. Tenstorrent also provides a library of pre-optimized operators for common neural network layers, such as [multi-head-attention](https://www.wikiprompt.org/wiki/multi-head-attention) and [residual-network](https://www.wikiprompt.org/wiki/residual-network) blocks, which can be used directly in models.

## Ecosystem and Integration

Wormhole cards are available as standalone PCIe cards for workstations and servers, as well as in pre-configured systems. Tenstorrent has partnered with system integrators to offer turnkey solutions for AI research and production. The cards are also available through cloud service providers, allowing users to access them on a rental basis. Tenstorrent has developed a software development kit (SDK) that includes examples and tutorials, aimed at reducing the learning curve for new users.

The company has also created a community forum and documentation portal, fostering an ecosystem of developers. Wormhole supports the [open-panel](https://www.wikiprompt.org/wiki/open-panel) standard, which is a specification for modular compute systems, enabling integration into heterogeneous data center environments. This allows Wormhole cards to be combined with other accelerators, such as those from [amd](https://www.wikiprompt.org/wiki/amd) or [intel](https://www.wikiprompt.org/wiki/intel), in a single system.

## Use Cases and Applications

Wormhole is primarily targeted at AI research and development, particularly for training and deploying large language models. Its high memory bandwidth and efficient dataflow architecture make it suitable for transformer-based models, which are widely used in natural language processing. The cards are also used for computer vision tasks, such as image classification and object detection, leveraging architectures like [u-net](https://www.wikiprompt.org/wiki/u-net) and [residual-network](https://www.wikiprompt.org/wiki/residual-network).

In addition to traditional AI workloads, Wormhole is being explored for scientific computing applications, such as simulations and data analysis. Its programmability allows researchers to implement custom algorithms beyond standard neural network operations. Tenstorrent has also highlighted its use in edge computing scenarios, where the power efficiency of the cards is an advantage.

## Comparison with Competitors

Wormhole competes with AI accelerators from companies like [groq](https://www.wikiprompt.org/wiki/groq), [samba-nova](https://www.wikiprompt.org/wiki/samba-nova), and [graphcore](https://www.wikiprompt.org/wiki/graphcore), as well as with GPUs from [nvidia](https://www.wikiprompt.org/wiki/nvidia) and [amd](https://www.wikiprompt.org/wiki/amd). Compared to GPUs, Wormhole offers a different programming model and potentially better efficiency for certain workloads, but has a smaller software ecosystem. Compared to other AI-specific chips, Wormhole's main differentiator is its focus on a dataflow architecture with a large number of cores, which can be more flexible than the fixed-function designs of some competitors.

In terms of raw performance, Wormhole's specifications are competitive with mid-range GPUs, but it lags behind the highest-end offerings. However, its scalability and power efficiency are strong selling points. Tenstorrent has also emphasized the open nature of its software stack, allowing for greater customization compared to proprietary alternatives.

## Development and Future Roadmap

Tenstorrent continues to develop the Wormhole series, with ongoing software updates and optimizations. The company has announced plans for future generations of its accelerators, building on the lessons learned from Wormhole. These future chips are expected to offer higher performance and improved efficiency, potentially using more advanced manufacturing processes. Tenstorrent is also working on expanding its software ecosystem, including better support for popular frameworks and tools.

The company has received significant funding and has established partnerships with various organizations, including [samsung-electronics](https://www.wikiprompt.org/wiki/samsung-electronics) and [tsmc](https://www.wikiprompt.org/wiki/tsmc), for manufacturing and technology collaboration. This positions Tenstorrent to continue competing in the rapidly evolving AI accelerator market.

## Reception and Impact

Wormhole has been well-received in the AI community for its innovative architecture and focus on programmability. Early adopters have praised its performance on transformer models and its ease of scaling. However, some users have noted that the software stack is less mature than that of established GPU vendors, requiring more effort to optimize models. Despite this, the open-source nature of some components has attracted a community of developers.

The impact of Wormhole extends beyond just the product itself, as it represents a growing trend of specialized AI hardware that challenges the dominance of GPUs. Its success has contributed to the broader ecosystem of AI accelerators, providing more options for researchers and companies.

## Conclusion

Wormhole is a significant entry in the AI accelerator space, offering a unique architecture that balances performance, efficiency, and flexibility. Its focus on dataflow computing and RISC-V-based control makes it a distinct alternative to traditional GPUs. While it faces challenges in software maturity, its strong scaling capabilities and growing ecosystem make it a viable option for many AI workloads. As Tenstorrent continues to iterate on the design, Wormhole and its successors are likely to play an important role in the future of AI hardware.

## References

Tenstorrent has published technical documentation and performance benchmarks for the Wormhole series, which are available through its official website. The company has also presented at various industry conferences, detailing the architecture and design choices. For more information, interested readers can consult these resources, as well as independent reviews and analyses from technology publications.

---
Source: https://www.wikiprompt.org/wiki/tenstorrent-wormhole
License: CC BY-SA 4.0 (https://creativecommons.org/licenses/by-sa/4.0/)
Last updated: 2026-09-09T01:54:57.985152+00:00