GPU Engineer

Société : Kog

Kog builds the fastest LLM inference engine on standard datacenter GPUs. Our Kog Inference Engine generates 3,000 output tokens per second per request on a single 8× AMD MI300X node and 2,100 on an 8× NVIDIA H200 node (FP16, batch size 1, no speculative decoding).

The hot path is a monokernel implemented with handwritten CUDA (with PTX inline assembly) on NVIDIA, and HIP (with CDNA ISA inline assembly) on AMD.

We optimize at the low level with engine/kernel/model co-design, using reverse engineering to understand and exploit the details of how the GPU hardware works at the micro level.

We are a team of 11 people, including 10 engineers and 4 PhDs.

Test it at playground.kog.ai. Read the technical details on the Kog Labs blog.

Description du poste

You will perform experiments to understand GPU internals, find creative solutions to accelerate critical computational sections used in LLM inference, and write optimized GPU kernels accordingly. Then test, profile, and optimize again.

Contribute to our monokernel pipeline, the single persistent GPU program that covers the full decode pass from QKV projection to LM head sampling, across AMD and NVIDIA architectures.

Work on low-level GPU optimization, including impossibly-fast grid synchronizations and inter-GPU collectives, and optimized GEMM and attention kernels for specific batch sizes and context lengths.

Build profiling infrastructure inside a monokernel, including custom instrumentation, device-timestamp frameworks, and per-stage analysis to translate machine behavior into concrete engineering decisions.

Scale the stack to third-party MoE models such as DeepSeek v4 and Qwen 3 to push generation speed on the models that matter in production today.

Contribute to building AI agents that will perform GPU Engineering research and kernel optimization autonomously, calibrated to hardware target and workload, starting from the inference foundations we are building now.

Profil recherché

You have written GPU kernels where performance was the central constraint. Showing the code is a requirement to move forward in the process.

PyTorch custom ops are an acceptable starting point if the kernels show a genuine understanding of the hardware below the framework level.

Stronger signals include inline PTX or CDNA ISA in public repositories, experience with latency-sensitive execution paths, understanding of why MBU matters more than MFU at batch size 1, and a background in inference engine components.

A top engineering school or a PhD with concrete GPU work counts, even without industry experience.