About the service
I'm a GPU developer and make heavy numerical tasks run orders of magnitude faster by moving them to the graphics card. When a simulation, signal processing or ML pipeline keeps the CPU loaded at full throttle for hours, I profile where time actually goes, restructure the hot section for thousands of parallel threads and rewrite it as CUDA kernels that load all stream multiprocessors without idle time. I care not only about correctness but about throughput: how data flows through global, shared and register memory, how coherent warps are and what fraction of theoretical bandwidth the kernel actually pulls on live hardware.
Most projects start with existing CPU code that's simply outlived its runtime. I port that logic to GPU step by step, verifying results against the original at each stage so numbers don't drift, then raise occupancy, coalesced memory access and thread overlap until acceleration becomes real and measurable. I work fluently with CUDA Toolkit and companion libraries — cuBLAS, cuFFT, Thrust — and rely on Nsight and profilers to prove exactly where the bottleneck is, not guess. Whether it's scientific HPC, signal processing or ML training and inference, I deliver kernels documented, reproducible and ready to integrate into your build.
If code is slow and you're hitting the processor wall, describe the task and where it hurts — I'll honestly tell you how much GPU can gain and how I'll get there.
— CUDA kernel development and optimization
— Algorithm porting from CPU to GPU
— Memory and bandwidth tuning (global, shared, registers)
— cuBLAS, cuFFT, Thrust integration
— Nsight profiling and bottleneck analysis
— Multi-GPU and thread concurrency
— HPC acceleration and ML training/inference speedup
Most projects start with existing CPU code that's simply outlived its runtime. I port that logic to GPU step by step, verifying results against the original at each stage so numbers don't drift, then raise occupancy, coalesced memory access and thread overlap until acceleration becomes real and measurable. I work fluently with CUDA Toolkit and companion libraries — cuBLAS, cuFFT, Thrust — and rely on Nsight and profilers to prove exactly where the bottleneck is, not guess. Whether it's scientific HPC, signal processing or ML training and inference, I deliver kernels documented, reproducible and ready to integrate into your build.
If code is slow and you're hitting the processor wall, describe the task and where it hurts — I'll honestly tell you how much GPU can gain and how I'll get there.
— CUDA kernel development and optimization
— Algorithm porting from CPU to GPU
— Memory and bandwidth tuning (global, shared, registers)
— cuBLAS, cuFFT, Thrust integration
— Nsight profiling and bottleneck analysis
— Multi-GPU and thread concurrency
— HPC acceleration and ML training/inference speedup
