nsight-compute

This is my 🔥 100 Days of GPU — a wild, hands-on journey through CUDA/CUTLASS kernels, Triton spells, and PTX sorcery.

Profiling with NVIDIA Nsight Tools Bootcamp

Evidence-driven CUDA, CUTLASS, Triton and GPU workload optimization for ChatGPT · 使用 ChatGPT 驱动 GPU workload 性能优化

Query every CUDA device attribute without profiling, and benchmark kernels against your hardware's theoretical peak.

A comprehensive, hardware-agnostic GPU benchmarking suite that compares CUDA, OpenCL, and DirectCompute performance using identical workloads. Built from scratch with professional architecture, extensive documentation, and production-ready GUI.

Hand-written LLM inference CUDA kernels for consumer Blackwell (RTX 5060 Ti / sm_120) — 1.05x vLLM's Marlin at batch-1 decode, with reproducible Nsight Compute evidence for every level.

基于证据的 vLLM v0.26 Scheduler Trace、KV Cache 压力、Waiting HOL 策略与 GPU Profiling 实验

CACAO is an open-source system that automatically generates, validates, profiles, and optimizes CUDA kernels using LLMs, real GPU performance feedback and KTT-based autotuning.

Reproducible, profiler-driven optimization studies covering CUDA kernels, deep-learning inference, multi-GPU communication and accelerated computing.

References content from the OLCF CUDA Training Series. (https://github.com/olcf/cuda-training-series)

TornadoVM vs hand-written CUDA on an RTX 4090: runnable Java and CUDA C++ versions of every demo, kernel-time comparisons attributed to root cause with Nsight Compute counters, and three upstream bugs found along the way.

A CUDA kernel playground for GEMM, Tensor Cores, Softmax, LayerNorm, validation, profiling, and performance benchmarking.

GPU-accelerated Number-Theoretic Transform for ZK-Proof generation. Targets the NTT bottleneck (91% of Groth16 prover time) via two CUDA optimizations: async double-buffered pipeline eliminating CPU-GPU transfer overhead, and IADD3-path Montgomery multiplication reducing finite-field instruction latency. BLS12-381, Ampere sm_86, Nsight-profiled.

Kernel-only profiling workflow for CUDA and Triton kernels with Nsight Compute, standardized reports, visual analysis, and vendor-portable adapters.

Profiling-driven micro-architectural case study & custom Triton GPU operators for batch-1 edge vision inference on NVIDIA Ada Lovelace.

Hand-written CUDA linear algebra library. A GEMM driven from a naive baseline to a compute-bound tensor-core kernel at ~95% of cuBLAS on an RTX 5070, profiled with Nsight Compute, plus GEMV, TRSM, CSR SpMV, cuSOLVER, a measured roofline, and NVML telemetry.

Feeding the Tensor Cores: a dense FP16 GEMM for NVIDIA Ada (sm_89) at 96.5% of cuBLAS, and what it teaches about how each GPU generation handles async copy and Tensor Core issue.

A hand-optimized CUDA attention kernel for RAG cross-encoder reranking, achieving ~25% lower attention latency than PyTorch SDPA on an NVIDIA T4.

CUDA FP32 GEMM optimization with loop unrolling, shared memory tiling, register tiling, benchmarking, and Nsight profiling.

An autonomous agent framework for evidence-driven GPU kernel optimization—explore, profile, tune, verify, and record every branch.