Edge AI Hardware Hack with Tenstorrent @ Studio 45

**To register, please join on the luma page: https://lu.ma/s4ievd0i**

​Come join for a full day of building on the latest AI hardware with the Tenstorrent team.
​The event is free of charge! Seats are limited.

# ​Agenda

• ​9:30am-10:00am - Registration / Coffee and Networking
• ​10am-10:30am - Kickoff
• ​10:30a-5:30p - Build
• ​Lunch provided
• ​5:30p-6:30p - Demos

# ​5 Different Challenge Tracks

​All Tenstorrent software is open source, and we encourage participants to browse the code ahead of time to pick a challenge that aligns with their interests. Whether you’re into compiler internals, ML training, or graphics experiments, there’s something here to get your hands dirty with.
​Below are a few suggested challenges—these are optional, but they’re all meaningful contributions to the growing Tenstorrent ecosystem.

1. ​Implement an Operator in tt-metal
   Help extend operator support in Tenstorrent’s low-level programming stack, tt-metal, by implementing a commonly used math or neural net primitive. We recommend: GitHub Issue #15939: Implement rsqrt. This is a great way to learn how ops are lowered and scheduled in Tenstorrent’s architecture.
2. ​Ray Marching with the SFPU
   Use the Special Function Processing Unit (SFPU) to build a basic Signed Distance Field (SDF) renderer using ray marching. This challenge explores how Tenstorrent’s hardware handles nonlinear math and graphics-style workloads, showing that it’s not just for transformers and CNNs.
3. ​Implement SRCNN for Image Upscaling
   Deploy a Super-Resolution Convolutional Neural Network (SRCNN)—a simple 3-layer model that enhances low-res images—on Wormhole using tt-metal. This task is approachable for those looking to understand model mapping, tensor layout, and operator sequencing on Tenstorrent hardware.
4. ​Train word2vec with tt-train
   Train a classic word2vec embedding model using Tenstorrent’s training stack, tt-train. This challenge is great for exploring custom training loops, loss functions, and tensor movement in a real-world NLP training scenario.
5. ​Port the CLIP Image Encoder to tt-nn
   Take the vision backbone of CLIP (Contrastive Language-Image Pretraining) and port it to run inference on tt-nn.
   This is ideal for participants excited about foundation models and wanting to explore the boundaries of Tenstorrent’s inference APIs.

# ​Winning Prize

• ​🏆 Best use of Tenstorrent will win a Wormhole n150d ($1099 value).

# ​Hosting partners

• ​Tenstorrent builds high-performance AI processors and RISC-V CPUs designed for scalable, efficient machine learning and deep learning workloads in data centers and edge devices.
• ​Koyeb is the fastest way to deploy full stack apps and AI workloads globally. No ops, servers, or infrastructure management.

# ​Thank you to our venue Studio 45!

• ​​Studio45 is a professional coworking space for professionals building across diferent areas of hardware and robotics.