GPU Kernel Engineer
San Francisco
ML /
Full-time /
Remote
Join us to build and safely deploy aligned, superhuman AI. We are building an AI pair programmer that feels like a full colleague inside your computer - capable, conversational, and reliable across domains.
As a GPU Kernel Engineer, you will design efficient implementations of novel model architectures and optimize kernels to ensure high throughput and low latency during training and inference.
Responsibilities
- Write efficient custom kernels for training and inference in CUDA/CuTe/Cutlass
- Optimize inference for our novel architectures, both by writing more efficient code and thinking about how we can sacrifice accuracy for speed
- Understand and optimize for H100 GPUs
- Think beyond the kernel level to the broader scheme of how we train these models and suggest improvements
We want someone who delights in optimization and loves seeing numbers go down. Here, a 1% optimization can save us hundreds of thousands of dollars in our most tightly used kernels. We have many ideas for new architectures, but often lack the engineering time necessary to write kernels for them, and they are often impracticable to implement in PyTorch. You will help with these kernels as they are optimized for large scale training and inference.
Requirements
- Understands and has worked on GPU programming, ideally matmul-heavy workloads
Magic's culture
- Integrity. Words and actions should be aligned.
- Hands-on. Most of us have previously led engineering teams. At Magic, there are no managers. We all spend the vast majority of our time on engineering. If you want to solve hard problems, Magic is the right place for you.
- Teamwork. We move as one team, not N individuals.
- Focus. Ethically deploy AGI. Everything else is noise.
- Quality. We have high standards for ourselves and our products. Magic should feel like magic.
Benefits and perks
- Benchmark-based compensation in the 75th or 90th percentile, including base salary, generous equity, and benefits
- 401K with 6% match
- Flexible working hours
- In-person (SF or Vienna) or remote
- A small, fast-paced, highly focused team
FAQ:
What's your motivation?
Automation has led humanity from subsistence farming to becoming a globally connected society. AGI is the ultimate chapter of the story of human tool-building, presenting the potential to decouple productivity and ingenuity from human labor. What if the last 50 years of technological progress happened in 2 days? We want to make this a possibility.
Funding?
We've recently raised $28M.
How do we balance deploying the technology today with ambitions for AGI?
We think deploying AI within the right interfaces is just as important as the technology itself. Building an AI pair programmer helps us do both at the same time. We aim to launch gradually improving AI assistants while pursuing work on what will ultimately become AGI.
Do you train your own models?
Yes
Do you care about the product?
It's funny that this is a question, but many AI companies neglect UX and focus only on their model. Yes, we care.
Can I work from anywhere?
We welcome applications from anyone around the world. We'll look at visa requirements case by case.
I don't meet all the criteria, should I still apply?
If you feel you have something to contribute to the mission and you're a high-energy person, absolutely. We make exceptions for exceptional people. In all hires, we are looking for either 1) difference makers on world class teams or 2) individuals who would become this very quickly if placed on such a team tomorrow.
