llm theoretical performance analysis tools and support params, flops, memory and latency analysis.

code for benchmarking GPU performance based on cublasSgemm and cublasHgemm

Hands-on Machine Learning Infrastructure on Kubernetes. Using Microk8s/Ubuntu on Paperspace Cloud.

A systematic CPU/GPU performance study of lightgbm and xgboost classifiers for different data shapes and hardware setups.

Disable GPU Thermal and Change GPU Governor to performance. (Only for snapdragon devices).

gpu thrashingNVIDIA GPU Unified Memory diagnostic tool — architecture-aware, measurement-based, PCIe/coherent transport detection

This repository provides the latest benchmarks for the CHARMM/pyCHARMM program on GPUs

📊 Mobile VR Performance Optimization Project

Comprehensive performance analysis of DeepSeek V3 quantization levels (FP16, Q8_0, Q4_0) on 16GB GPU environments.

Evidence-driven PresentMon diagnostics and policy modeling paired with bounded owned-lab D3D11 runtime actuation.

Reproducible long-context inference benchmark comparing vLLM, SGLang, and TensorRT-LLM on NVIDIA GB10.

Owned-lab D3D11 runtime research with bounded exact-content copy, readback, staging-upload, and UpdateSubresource elision.

Profile-driven FP32 CUDA GEMM optimization: naive --> tiled --> coalesced --> register-blocked --> bank-padded, benchmarked against cuBLAS.

Professional GPU Performance Testing Suite for dual GPU setup (AMD RX 6600 + NVIDIA RTX 3050) with comprehensive monitoring tools, crash-safe scripts, and thermal management optimized for ASRock X570 Taichi + Ryzen 7 5700X

Sparse binary 2D FFT on CUDA/cuFFT with memory-footprint optimization, streaming tiles, Hermitian symmetry, and Nsight analysis.

Reproducible Qwen3-1.7B prefix cache benchmark on RTX 3070 Laptop (8GB). Hand-written reference inference loop + vLLM v1 comparison.

Reproducible Instruction Roofline analysis of cuSPARSE and Ginkgo SpMM on RTX 4090 using Nsight Compute metrics.

WMMA FP16-->FP32 Tensor Core GEMM with shared-memory tiling and cp.async-style pipelining, benchmarked against cuBLAS.

Python lab for exploring memory bandwidth, cache effects, and locality in accelerator workloads

Single-head CUDA attention kernel: naive SDPA --> fused softmax --> occupancy-tuned variants, benchmarked against cuDNN SDPA with Nsight Compute profiling.