Write and optimize custom CUDA kernels for transformer inference operations.
Profile and tune kernels to improve GPU occupancy, memory throughput, latency and throughput.
Apply kernel fusion, quantization, mixed precision and caching to improve inference performance.
Benchmark kernels, write tests, and maintain internal CUDA libraries and documentation.
Summarised from the posting by AI and can be wrong. The full description is the source.
Requirements
Experience
4+ years writing production CUDA code, with a track record of shipping performance-critical kernels
Hands-on experience profiling to diagnose compute-bound versus memory-bound bottlenecks
Experience with kernel fusion, memory coalescing and avoiding warp divergence
Experience writing quantised and mixed-precision kernels
Experience with transformer or attention-style kernels, or autoregressive decoding (preferred)
Experience building high-performance GPU libraries from scratch (preferred)
HPC or latency-critical performance engineering experience (preferred)
Multi-GPU or multi-node kernel-level optimisation experience (preferred)
Skills
Deep understanding of GPU microarchitecture, including warps, occupancy, register pressure and memory hierarchy
Strong CUDA C++ skills, including streams and asynchronous execution
Solid grasp of parallel algorithm design and numerical precision tradeoffs
Comfort reading PTX/SASS to validate kernel efficiency (preferred)
Benefits
Health insurance
Private health insurance
Equity
Eligibility for equity
Bonus
Biannual bonus scheme
Meals
Breakfast and dinner allowance for office-based employees
Other
Fully expensed tech to match your needs, Competitive salary
Full job description
From Fuse Energy’s careers site
Fuse Energy is an energy startup on a mission to make energy abundant and affordable, fast. We combine first-principles thinking with cutting-edge technology to build a radically better energy system.
We've raised over $200M from top-tier investors including Balderton, Lakestar, Accel, Creandum, Lowercarbon, Ribbit, 20VC, Hummingbird and Collaborative Fund, alongside strategic angels including Nico Rosberg and GPs behind Meta, Revolut, Spotify and Uber.
Show full descriptionShow less
We're building a fully integrated energy company: developing our own solar, batteries and other generation projects, building our own hardware, improving and developing grid infrastructure, trading power in real time, using AI across the business, and installing distributed energy in homes. By selling directly to consumers we cut out the middleman, lower costs and pass the savings on to our customers.
As data centres become one of the largest and fastest-growing sources of electricity demand, Fuse is expanding into high-performance compute infrastructure at the intersection of energy and AI. We're looking for a CUDA Engineer to write and optimise the low-level GPU code that powers our inference workloads: designing custom CUDA kernels, tuning performance across memory bandwidth and compute bottlenecks, and squeezing maximum throughput out of every GPU in our fleet, working at the level of SMs, warps and memory hierarchies.
Responsibilities
Write and optimise custom CUDA kernels for core transformer inference operations
Profile kernels to identify and eliminate bottlenecks in occupancy, memory throughput and warp divergence
Apply kernel fusion to reduce memory round-trips and launch overhead across inference pipelines
Optimise memory access patterns and manage the memory hierarchy for maximum bandwidth utilisation
Implement quantisation-aware kernels and mixed-precision arithmetic to reduce latency and memory footprint
Build and tune caching mechanisms for efficient autoregressive decoding
Tune kernel launch configurations for target GPU architectures
Benchmark kernels against existing baselines and drive measurable throughput and latency improvements
Write tests for CUDA code to catch performance and correctness regressions
Maintain internal CUDA libraries and contribute to team coding standards and documentation
Requirements
4+ years writing production CUDA code, with a track record of shipping performance-critical kernels
Deep understanding of GPU microarchitecture: warps, occupancy, register pressure and memory hierarchy
Strong CUDA C++ skills, including streams and asynchronous execution
Hands-on experience profiling to diagnose compute-bound vs memory-bound bottlenecks
Experience with kernel fusion, memory coalescing and avoiding warp divergence
Experience writing quantised and mixed-precision kernels
Solid grasp of parallel algorithm design and numerical precision tradeoffs
Bonus: transformer/attention-style kernels or autoregressive decoding; building high-performance GPU libraries from scratch; HPC or latency-critical performance engineering; multi-GPU or multi-node kernel-level optimisation; comfortable reading PTX/SASS to validate kernel efficiency
Benefits
Competitive salary and eligibility for equity
Biannual bonus scheme
Fully expensed tech to match your needs
Private health insurance
Breakfast and dinner allowance for office-based employees