About the company: A venture backed AI hardware company building an inference accelerator for the next generation of large model workloads. The company is well funded, has a differentiated architecture, and is moving from design into first silicon, with real customer interest ahead of launch.
Engineering is organized around small teams that own whole problems end to end. Software and hardware engineers work side by side, benchmark against the best GPU results in the industry, and treat every percentage point of efficiency as a competitive edge. The culture is direct, curious and focused on measurable results rather than process.
The role: You will own the low level kernels that determine how quickly customer models run on the chip. There is no mature kernel library to lean on, so you will build it: matrix multiplies, attention, normalization and the operators that follow, tuned for a new memory and compute architecture.
You will profile real workloads, find where time and bandwidth are lost, and work with the compiler and architecture teams on fixes that change the design as well as the code. The results you produce will show up in customer benchmarks and in the company’s own story about performance.
What you’ll do
- Develop high performance kernels for matrix and attention workloads
- Profile and remove bottlenecks across compute and memory
- Build the kernel library and benchmarking tools
- Feed results back to the architecture team
What you’ll bring
- Strong C++ and low level optimization experience (CUDA or similar)
- Knowledge of memory hierarchies, parallelism and profiling
- A record of measurable speed ups
Nice to have
- Non Nvidia target experience (ROCm or custom ISA)
- GEMM and attention kernel work
- Assembly or intrinsics
Why join: Your kernels set the performance customers see, and you will work next to the people designing the chip.
Darwin Recruitment is acting as an Employment Agency in relation to this vacancy.
Oliver Stewart