- Experience
- Any
- Salary
- —
- Openings
- 1
- Posted
- 1 week ago
- Work mode
- Remote
- Education
- BSc/MSc/PhD in Computer Science, Engineering, Machine Learning or equivalent
- Resume
- Required to apply
Sign in to tell us what does and doesn't work for you here — it sharpens every match we show you.
Job description
Company Overview
TensorX is an AI infrastructure platform based in Dublin, operating NVIDIA Blackwell GPUs across European data centres under strict EU jurisdiction. The company delivers open-weight large language model inference through an OpenAI-compatible API with strict data privacy and compliance adherence. Their clientele includes regulated enterprises in finance, healthcare, government, AI developers, and platforms.
Role Summary
TensorX seeks multiple GPU Performance Engineers to elevate inference engine throughput within latency constraints. Reporting to the CTO, you will take ownership from issue detection through resolution, patch open-source engines, and optimize GPU kernels when needed. The position emphasizes increasing real-world 'goodput'—the count of requests processed within latency targets—rather than theoretical peak performance.
Key Responsibilities
- Own engine-related tickets from initial report to full resolution, including reproducing issues, diagnosing causes, applying patches, and documenting findings.
- Conduct cross-layer diagnosis across GPU memory, engine scheduler, containerization, network router, and request traffic patterns.
- Identify bottlenecks affecting GPU goodput, whether in kernels, schedulers, routers, or configurations; optimize to regain throughput based on production traffic measurements.
- Patch defects in open-source inference engines such as SGLang and vLLM; maintain fleet-wide patch consistency and contribute fixes upstream where applicable.
- Modify or author GPU kernels when profiling highlights engine kernels as bottlenecks, including attention and indexing paths and quantized decoding.
- Bring up new models at day zero; perform A/B testing across parallelism and cache configurations to determine optimal setup for deployment teams.
- Tune prefix caching, KV cache replication, and behaviors under tensor parallelism; collaborate on cache-aware routing and routing code enhancements.
- Collaborate in pre-production testing to ensure all changes meet performance criteria prior to customer release.
- Produce clear, reproducible technical documentation and benchmark methods usable by others and customers.
Required Skills & Experience
- Proven ability to analyze inference engine source code to diagnose and fix root causes, supported by code contributions or reports.
- Strong understanding of GPU architectures and CUDA fundamentals including memory hierarchies, occupancy, and kernel characteristics.
- Familiarity with transformer-based model inference including attention mechanisms, KV caches, batching strategies, and quantization effects.
- Proficiency in Python and competence with reading C++ and CUDA codebases.
- Methodical approach to performance testing with controlled variables and rigorous correctness verification.
- Integrity in handling results, transparently reporting both successful and unsuccessful optimizations.
- Motivation to continuously learn and keep pace with evolving stacks and open-source developments.
- Experience incorporating AI-assisted development tools (e.g., Claude Code, Codex) in daily workflows.
- Effective communication, able to explain complex technical issues to diverse audiences including non-technical stakeholders.
Desirable Qualifications
- Experience writing and benchmarking CUDA kernels on Hopper or Blackwell GPU architectures.
- Contributions to inference engine projects such as SGLang, vLLM, or TensorRT-LLM.
- Familiarity with kernel tools like Triton, CUTLASS, CuTe, or ThunderKittens.
- Achievements in GPU mode leaderboards, MLSys competitions, or similarly relevant contests.
- Expertise profiling memory utilization and resolving out-of-memory errors in large models.
- Hands-on experience with Blackwell GPU features including tcgen05, TMEM, TMA.
- Public technical writing experience.
Why Join TensorX?
- Work at the engine level, directly modifying GPU kernels and inference engines rather than peripheral systems.
- Operate on an owned fleet of NVIDIA B300 GPUs in Dublin and Helsinki rather than rented cloud instances.
- Engage with authentic production traffic featuring long contexts and high concurrency that challenge typical benchmarks.
- Leverage results that roughly double requests served within latency targets compared to standard setups on identical hardware.
- Contribute to and benefit from an open engine where patches are reflected upstream, enhancing broader community knowledge.
- Engage with a broad research agenda including KV cache extensions, context parallelism, and fine-grained model parallelism.
Education & Benefits
- A bachelor’s, master’s, or doctoral degree in computer science, engineering, machine learning, or a related technical discipline, or equivalent demonstrated skills.
- Competitive compensation tailored to experience.
- 25 days of paid annual vacation.
- Hybrid working model combining remote flexibility and a centrally located Dublin office.
- Complimentary access to inference tokens.
Additional Information
No agency assistance is requested for this recruitment.
Minimum education
Doctorate