Freelancer profile translated to English.

Description

GPU Performance Engineer CUDA C++ | PTX | Inference Optimization

GPU kernel optimization, Nsight profiling, CPU → GPU migration. 5 years of experience (Airbus, DPD Group, Melexis)

GPU Engineer specializing in CUDA C++ kernel optimization and performance profiling. 5 years of experience in aerospace (Airbus), logistics (DPD Group), and semiconductors (Melexis).

My approach: Below PyTorch, not above.

I intervene at the kernel level, where real performance gains are made.

What I can offer you:

→ Development and optimization of CUDA C++ kernels with Nsight Compute profiling, roofline analysis, and production-validated gains (up to 33x speed-up)

→ CPU → GPU migration and production deployment on NVIDIA T4/V100/A10G with numerical validation and cloud deployment

→ GPU performance audit: compute/memory/latency-bound classification, bottleneck identification, actionable optimization plan with quantified metrics

→ GPU mentoring and expertise transfer: team training on profiling, memory optimization patterns, and GPU architecture

Production results:

→ 33x CUDA kernel speed-up (Airbus, V100)

→ GPU utilization 9% → 89% (Airbus)

→ 30% throughput gain, CPU → GPU migration in 6 weeks (DPD Group)

→ 40% reduction in processing time (Datashift, A10G)

In parallel to my assignments, I contribute to the open-source GPU ecosystem (ThunderKittens, model-kernels) and I am developing an FP4 fused attention kernel for Blackwell GPUs in inline PTX.

Technical blog and profiling guide (20,000+ words) at florianmattana.com.

Shall we discuss your project?

Contact me, I respond very quickly.

Skills

CUDA, C++, GPU Computing, NVIDIA, Profiling, Performance Optimization, Nsight Compute, Nsight Systems, Tensor Cores, HPC, CUTLASS, PTX, Inference Optimization, CPU-GPU Migration

Industry field of expertise

Languages

English
Native or bilingual
French
Native or bilingual

Workplace preferences

Can work on-site

Paris (up to 50km), Lyon (up to 50km), Lille (up to 50km), Marseille (up to 50km), Bordeaux (up to 50km)

Open Source | LLM | Inference
GPU Performance Engineer
TECH
January 2026 - Today (7 months)
Paris, France
Active contributor to the open-source GPU ecosystem, focused on CUDA inference kernel optimization and performance profiling.

As a GPU Performance Engineer, I work on:

→ Development of an FP4 fused attention kernel for consumer Blackwell (SM120) in inline PTX — GEMM-softmax-GEMM fused in registers with mma.sync and block scaling UE8M0
→ Bug fixing for compilation and precision issues on existing inference kernels
→ Profiling and performance auditing of real-world GPU kernels with Nsight Compute
→ Writing technical documentation on GPU profiling
Main contributions:
→ model-kernels: 4 PRs merged — fix for 5 compilation bugs and 2 precision bugs on an INT8 fused attention kernel, max error reduced from 1.69 to 1.37
→ ThunderKittens (Stanford HazyResearch): PR #179 — fix for a narrowing-conversion bug in base-type packing
→ fp4-fused-attention-sm120: FP4 fused attention kernel from scratch for consumer Blackwell GPUs in inline PTX (mma.sync.aligned.mxf8f6f4)
→ CUDA-Kernels: collection of optimized kernels from scratch (GEMM, reduction, prefix scan, softmax, Flash Attention) with full NCU profiling — best GEMM at 58.8% of cuBLAS on RTX 5070 Ti
→ GPU Profiling Guide (20,000+ words) covering Nsight Systems and Nsight Compute end-to-end

Technical environment: CUDA C++, PTX inline, Tensor Cores, Nsight Compute, Nsight Systems, RTX 5070 Ti (SM120), Git, Linux
ThunderKittens CUDA Generative AI AI Engineer HPC
Melexis
GPU Performance Engineer
AUTOMOBILE
March 2024 - December 2025 (1 year and 9 months)
Brussels, Belgium
Melexis is a company specializing in the testing of semiconductor sensors on AWS cloud infrastructure.

I joined the GPU Compute team to manage the GPU computation pipeline for sensor testing on AWS EC2 g5 (NVIDIA A10G).

As a GPU Performance Engineer, my responsibilities included:

→ Development and maintenance of the GPU compute pipeline in CUDA C++
→ Multi-precision numerical validation (FP64, FP32, FP16, FP8) with automated CI (cosine similarity ≥ 0.9995)
→ Diagnosis and correction of numerical corruptions (NaN in FP16) via adversarial fuzzing and dynamic range scaling
→ Optimization of host-device transfers with CUDA streams and pinned memory

I contributed to the following developments:

→ 40% reduction in end-to-end processing time
→ Implementation of an automated multi-precision numerical validation CI gate
→ Improvement of daily throughput through host-device transfer optimization

Technical environment: CUDA C++, Python, Nsight Compute, Nsight Systems, AWS EC2 (g5, A10G), Docker, GitLab CI, Linux
CUDA Linux HPC Performance Improvement C++
Airbus via Accenture
GPU Performance Engineer
AVIATION AND AEROSPACE
April 2022 - March 2024 (1 year and 11 months)
Toulouse, France
Airbus is the world leader in aeronautics and space.

I joined the satellite inspection team to optimize a CUDA kernel for leak detection on Tesla V100.

As a GPU Performance Engineer, my responsibilities included:

→ Optimization of the CUDA kernel with Nsight Compute profiling (coalesced memory access, warp divergence elimination)
→ Reduction of shared memory bank conflicts through tile padding and double buffering
→ Production validation of the obtained speed-up
→ Training of 5+ engineers on Nsight Compute profiling workflows

I contributed to the following developments:

→ 33x kernel speed-up (30 min → < 1 min), validated in production
GPU utilization from 9% → 89%
→ 40% reduction in shared memory bank conflicts
→ Transition from overnight batch processing to same-day turnaround for satellite inspection jobs

Technical environment: CUDA C++, Nsight Compute, Nsight Systems, Tesla V100, Python, Linux
CUDA Linux GPU C/C++ Programming C++ Development