PostHeaderSection

Symbol

Global NavBar

PostContentSection

Global Footer

Blog Detail Page

Alex Medick

Discover how Lumerian Labs delivers CUDA-level performance - no kernels needed. Reinventing AI compute for PyTorch and heterogeneous GPUs.

Beyond CUDA with Jay Dawani: No Kernels Needed for Speed

Welcome to another episode of Beyond CUDA, where we talk with the people pushing the boundaries of AI compute. In this week’s episode, Jeff Tatarchuk sits down with Jay Dawani, co-founder and CEO of Lumerian Labs, to explore why performance in AI workloads doesn’t have to mean writing kernels, and why the future of compute might depend on a complete software rethink. The short version? Lumerian is building a new AI stack that lets developers stay in PyTorch and still hit CUDA-level (or better) performance on heterogeneous hardware.









Episode Highlights


CUDA Performance Without Writing CUDA

“We can basically give you CUDA-level performance or even better, PTX-level performance without you ever having to write CUDA. And while staying in PyTorch.”

Jay’s bold pitch: Lumerian’s runtime dynamically analyzes workloads, extracts parallelism, and optimizes scheduling... all without the dev needing to know anything about warps, threads, or hardware-specific kernels.


Why Lumerian Exists

“We have everything it takes. We know how to train models. We know how to scale them. But why don’t we have 300 million robots in the world?”

Jay’s journey began with robotics, but it quickly became clear: the bottleneck isn’t ideas, it’s compute. Lumerian originally set out to build a robotics chip, but pivoted toward the data center when they realized the core problem was systemic and software-rooted.


The End of Kernel-Centric Thinking

“Kernels are local optimizations. But real performance, especially in a post-Moore’s Law world, comes from global system-level optimization.”

Instead of relying on a growing list of hand-tuned kernels, Lumerian’s approach is to generate optimized compute plans in real time, tailored to your model and hardware. Think compilers with semantic awareness, not static libraries.


Rethinking the Software Stack

“If you’re living in the current paradigm, you have to choose between productivity, portability, and performance. But if you change the constraints—you can have all three.”

By moving compute to where data resides (rather than the reverse), optimizing for memory hierarchies, and building a new programming abstraction layer, Jay believes we can unlock 10x performance over traditional stacks without burning developers out.


The Market Need: Now

“There’s not enough land or energy to keep scaling the way we are. We’ll need a billion GPUs to support future model workloads. We might get 50 million. That gap has to be filled by better software and better algorithms.”

Jay argues that AI isn’t overhyped, but under-hyped. But current compute stacks aren’t ready for what’s coming. Lumerian aims to fill that gap by making it easier to exploit hardware performance without deep hardware expertise.


Building for Developers

“Our pitch to developers? Are you really comfortable with CUDA? Do you like your sanity?”

Lumerian is releasing an open-source DSL and backend support for PyTorch. Developers will be able to drop in and get performance gains on diverse hardware without rewriting their stack.


Key Takeaway

You don’t need to write low-level CUDA kernels to get high performance. You just need the right runtime.

As compute needs explode and hardware becomes more heterogeneous, developer productivity can’t be sacrificed at the altar of performance. Lumerian Labs is building the stack that makes AI development faster, more accessible, and more portable, all without the burn.

🧠 Want more like this?

Subscribe to Beyond CUDA on YouTube and stay tuned for future episodes diving deep into the future of open AI infrastructure.


About TensorWave

TensorWave is the AMD GPU cloud purpose-built for performance. Powered exclusively by Instinct™ Series GPUs, we deliver high-bandwidth, memory-optimized infrastructure that scales with your most demanding models—training or inference.

Ready to get started? Connect with a Sales Engineer.

Beyond CUDA with Jay Dawani: No Kernels Needed for Speed

Product

Solutions

Resources

Company

Product

Resources

Solutions

Company