Big things have been happening at Modular, and we’re bringing it all to GTC. Find us at Booth #3004 in San Jose, CA, March 16–19.

Most AI infrastructure today is assembled from parts that were never designed to work together: one tool for serving, another for optimization, another for custom kernels, something else for scaling. Every layer you add is another place where things can break. We built Modular to fix that, and at GTC, we're showing what that looks like in practice.

‍

Stop by our booth to see:

‍

AI-native GPU development with open source Mojo skills We open sourced the AI tools our engineers use internally to write Mojo and MAX. Last week, an agent one-shotted a port of NVIDIA's CUTLASS Blackwell conv2d to Mojo, hitting 130.7 TFLOPS and 4-7x cuDNN in a single session.

‍

CUTLASS kernel ports in Mojo Our matmul is hitting 1,772 TFLOPS on B200, exceeding cuBLAS, with code that any engineer can read, modify, and own without reverse-engineering CUDA internals.

‍

A first look at Modular Cloud with DeepSeek V3.1 We're giving attendees an early look at Modular Cloud. Stop by the booth to see what inference looks like when the stack is built from scratch (hint: fast, portable, and performant).

‍

FLUX.2-dev image generation on MAX with SqueezeBits In partnership with SqueezeBits, our MAX platform is running FLUX.2-dev end-to-end to show what generative image workloads look like when the entire stack is optimized together, rather than assembled after the fact.

‍

The minds behind the stack will be at the booth all week, ready to get into the details.

‍

In October’s Community Meeting, we explored the latest innovations from the community and the Modular team. Martin Vuyk shared a generic FFT implementation in Mojo, Gabriel de Marmiesse presented a MAX backend for PyTorch, and Brad Larson walked through the Modular 25.6 release, now offering unified GPU support for NVIDIA, AMD, and Apple.Watch to see demos, insights, and discussion from the community and the Modular team!

For our 20th community meeting, Modular hosted a Q&A with the Mojo team to discuss the recently released Mojo vision and roadmap documents.

‍

We also heard from:

Bernardo Taveira on “Porting GSplat Kernels to Mojo”
Seif Lotfy on “Datastructures for DB Development"

Join us at Modular’s Los Altos office for an evening of big ideas and engineering insights on bringing AI from concept to production.

​Talks include:

• ​Chris Lattner, Modular – The future of democratizing AI compute and the role of open collaboration in accelerating progress.
• ​Feifan Fan, Inworld AI – How to integrate state-of-the-art voice AI into consumer applications and make it production ready, featuring insights from Inworld’s collaboration with Modular.
• ​Chris Hoge, Modular – Why matrix multiplication remains one of the most challenging problems in computer science and how the Modular stack helps developers optimize it.

​Enjoy complimentary refreshments, engage in Q&A with the speakers, and network with other builders.

​Please note that in-person space is limited and is first come, first served. Doors open at 6 PM PT and a valid Luma event registration is required for admission. Arriving late may result in being turned away if the venue has reached capacity. We recommend arriving by 6 PM PT, if possible.

​There will also be a virtual livestream option for attendance.

​Agenda

​6 PM PT | Doors open

​6:30 PM PT | Talks start

​7:45-9 PM PT | Networking

‍

During our August Modular Community Meeting, Manuel walks us through optimizations to his Mojo regular expression library, mojo-regex, covering architecture decisions, a hybrid DFA/NFA engine, SIMD optimizations, benchmarking challenges, and the future roadmap. Modular intern Amir then shares his work enabling Apple GPU support in Modular, explaining the technical hurdles, compiler changes, and design choices required. Finally, we wrap up with Q&A with the team.