Tenstorrent
Tenstorrent provides RISC-V based AI hardware and an open-source software stack for developers who need low-level control over compute kernels and want to bypass the CUDA ecosystem.
Excellent for hardware-level AI researchers and RISC-V enthusiasts, weaker for teams requiring plug-and-play parity with Nvidia's software maturity.
Analysis based on product data, pricing structure, traffic signals, and public user sentiment.
Who Should Use Tenstorrent?
Typical users
AI infrastructure engineers, hardware researchers, and systems programmers at mid-to-large labs or specialized startups.
Maturity fit
advanced
Choose this if…
- You require direct access to hardware kernels via an open-source stack
- Your priority is avoiding proprietary vendor lock-in like CUDA
- You are building custom AI models that benefit from decentralized, non-host-dependent compute
- You are specifically developing for the RISC-V architecture
Skip this if…
- You need immediate, high-performance support for every niche PyTorch operator
- Your team lacks the engineering bandwidth to optimize for non-standard hardware
- You rely on third-party software that is strictly optimized for Nvidia GPUs
About Tenstorrent
Tenstorrent is a computing company led by Jim Keller that designs AI processors based on the RISC-V instruction set. It offers an alternative to traditional GPU architectures by focusing on modular chiplets and an open-source software philosophy that allows developers to program hardware directly.
Official profiles
What it actually does
The company sells AI accelerator cards and workstations that run machine learning workloads without requiring a heavy host CPU. Their software stack, including TT-Buda and TT-Metalium, allows users to compile high-level models or write custom low-level kernels for their Tensix cores.
What makes it different
Unlike Nvidia, Tenstorrent uses a decentralized architecture where each core can handle its own compute and communication. Their commitment to open-sourcing their software stack (TT-Metalium) allows developers to see and modify how data moves through the chip, which is impossible with proprietary alternatives.
Key Features
Tensix Cores
Programmable cores designed specifically for tensor operations and data movement.
TT-Metalium
A low-level programming framework that gives developers direct control over hardware resources.
Wormhole Architecture
Features integrated 100Gb Ethernet ports on the chip for native scaling without external switches.
Grayskull Accelerators
Entry-level PCIe cards designed for AI inference and development at a lower price point.
Host-independent Execution
The ability for the hardware to manage its own graph execution, reducing CPU bottlenecks.
RISC-V General Purpose Cores
Integrated cores that handle complex branching and non-tensor logic on-chip.
Blackhole (Upcoming)
Next-generation architecture designed for higher throughput and standalone compute nodes.
Pricing
Grayskull e75
- Small form factor PCIe card
- 75W TDP
- Ideal for entry-level AI inference
- TT-Buda and TT-Metalium support
Grayskull e150
- Standard PCIe card
- 200W TDP
- Higher throughput for model development
- Full software stack access
Wormhole n150
- Wormhole architecture
- 12GB GDDR6 memory
- Optimized for networking and scaling
- Direct Ethernet connectivity
Wormhole n300
- Dual-chip Wormhole card
- 24GB GDDR6 memory
- Best for multi-chip development
- Integrated networking
Galaxy / Nebula Systems
- Multi-node server racks
- Liquid or air-cooled options
- Enterprise-grade support
- Turnkey AI compute clusters
Pricing checked 5 months ago
Pricing guidance
- When moving from single-model testing to multi-chip scaling (Wormhole)
- When inference latency on Grayskull becomes a bottleneck for production
- When transitioning from individual development to rack-scale deployment
- Requires specific Linux kernel versions for driver stability
- PCIe power requirements may necessitate PSU upgrades in standard workstations
- Physical cooling requirements for high-TDP cards in non-server chassis
Aggressively priced for developers to encourage ecosystem adoption, with premium pricing reserved for large-scale cluster deployments.
Pros & Cons
Strengths
-
Open-source software transparency
By open-sourcing their compiler and kernel code, they allow developers to debug and optimize at a level that proprietary stacks block.
-
Cost-effective development entry
Dev kits like the Grayskull e75 are priced under $600, making hardware-level AI experimentation accessible to individual researchers.
-
Native scaling via Ethernet
The Wormhole architecture allows cards to be networked directly, simplifying the creation of large-scale compute clusters.
-
Reduced host CPU dependency
The architecture handles data movement and instruction sequencing on-chip, which can lower the total system cost by requiring less powerful host processors.
Weaknesses
-
Software ecosystem immaturity
While improving, the stack does not yet support the vast library of pre-optimized kernels available in the CUDA ecosystem.
Affects: Teams needing to deploy production models with zero code modification
-
Documentation gaps
As a developing platform, technical documentation can be sparse or change rapidly, requiring more trial and error from users.
Affects: Junior developers or teams without deep systems engineering experience
-
Limited third-party integrations
Most MLOps and orchestration tools are built for GPUs first; integrating Tenstorrent into existing cloud workflows requires manual effort.
Affects: Enterprise teams with highly automated deployment pipelines
Real User Sentiment
Users are generally excited about the hardware's potential and the open-source philosophy, though they acknowledge the software is still a work in progress.
Users tend to like
- The transparency of the open-source stack
- Jim Keller's architectural vision
- Affordable entry price for dev kits
- The move away from CUDA dependency
Users commonly complain about
- Initial setup can be difficult
- Software bugs in early releases
- Limited support for older or very obscure ML frameworks
Recurring tradeoffs
- Lower cost and more control vs. higher engineering effort to get models running
Happiest users
Systems engineers and researchers who enjoy 'tinkering' with hardware and want to optimize their own kernels.
Often frustrated
Data scientists who expect a 'one-click' experience to run a standard Hugging Face model at peak performance.
Use Cases
Custom Kernel Development
Writing specialized math operations that are inefficient on standard GPUs.
RISC-V Research
Testing how AI workloads perform on an open instruction set architecture.
Edge AI Prototyping
Using low-power Grayskull cards to develop inference solutions for local deployment.
Distributed Training Research
Utilizing Wormhole's direct Ethernet to study new ways of scaling models without traditional switches.
Cost-Effective Inference
Deploying trained models on hardware that is significantly cheaper than Nvidia's enterprise offerings.
Frequently Asked Questions
How much does Tenstorrent hardware cost?
Developer kits start at $599 for the Grayskull e75 and go up to $1,399 for the dual-chip Wormhole n300. Large-scale systems like the Galaxy require a custom quote from sales.
Does Tenstorrent support PyTorch?
Yes, Tenstorrent supports PyTorch through their TT-Buda compiler, which maps PyTorch graphs to their hardware. They also offer TT-Metalium for those who want to bypass the compiler and write kernels manually.
How does Tenstorrent compare to Nvidia?
Nvidia offers a more mature software ecosystem (CUDA) and higher raw performance for standard models, but it is closed-source and expensive. Tenstorrent offers an open-source stack, RISC-V flexibility, and lower hardware costs, but requires more manual optimization.
Can I run LLMs on Tenstorrent cards?
Yes, models like BERT, ResNet, and various LLMs are supported. However, performance and ease of setup vary depending on whether the specific model architecture has been optimized for the Tensix cores yet.
What is the difference between Grayskull and Wormhole?
Grayskull is an older architecture focused on pure AI acceleration via PCIe. Wormhole adds integrated 100Gb Ethernet ports and RISC-V cores, making it better suited for networking multiple chips together into a larger compute fabric.
Is the software really open source?
Yes, Tenstorrent has open-sourced significant portions of their stack, including TT-Metalium and their kernel code on GitHub, allowing developers to inspect and contribute to the codebase.
Why trust this page?
This evaluation combines product positioning, pricing analysis, traffic and market signals, and public user sentiment into a single decision-support page. Content is generated editorially — not copied from the vendor's website.
Funding & Company
Founded
2016
Stage
Late stage
Total Raised
$1.02B
Latest Round
Series D (Dec 2024)
Notable Investors
Tenstorrent has raised over $1 billion, culminating in a massive $693 million Series D in late 2024. This significant, recent funding from strategic investors like Samsung and Hyundai provides a very long runway and validates its technology for use in major industries like automotive and consumer electronics.
Market Signals & Traffic
Estimated visits, global rank, geography, traffic sources, monthly visit trends, and organic search keywords (Similarweb)—on a dedicated page built for depth and search.
- Estimated visits
- 118,502
- Global rank
- #313,024
- Snapshot
- Apr 2026
- Traffic trend
- Surging
Estimated monthly visits
Alternatives to Tenstorrent
View all alternatives
Similar Tools
Etched
Developer Tools, Research
Specialized ASIC hardware for accelerating Transformer model inference and throughput.
CVAT
Developer Tools, Research
Collaborative image and video annotation platform for computer vision
Nous Research
Developer Tools, Research
Open-source research collective developing high-performance large language models.
Akash Network
Developer Tools, Research
Decentralized cloud marketplace for high-performance GPU and compute resources.
Label Studio
Developer Tools, Research
Open-source data labeling platform for training machine learning models.
FluidStack
Developer Tools, Research
Global GPU cloud platform for AI training and inference.