gpu-benchmarking

CUDA matrix multiplication benchmarking on Jetson Orin Nano. Four implementations, three power modes, five matrix sizes. 99.5% mathematical validation. C++/CUDA and Python.

Dashboard for AI Studio, Open Source Continuous Inference | Deepseek-R1, Qwen2.5, Llama3.1 | 4xRTX-5090 inside PRU2500, 2xH100 inside PRU2500, 8xMI210 in SuperMicro

LLM benchmarking, GPU workload orchestration backend server | Deepseek-R1, Qwen2.5, Llama3.1 | 4xRTX-5090 inside PRU2500, 2xH100 inside PRU2500, 8xMI210 in SuperMicro

Automated game testing and GPU Cooling Tester Evaluation with reproducible measurements and profile authoring

Benchmark your local LLM server (Ollama, vLLM, any OpenAI-compatible API) in one command: how many users your GPU can serve, time to first token, tokens/s, p99.

Analytical performance modeling, VRAM capacity planning, speculative decoding sizing, and hardware roofline prediction for LLM serving.

Standalone LLM inference benchmarking pipelines on AMD GPUs using ROCm, vLLM, MAD, and data visualization scripts.

Plan your LLM deployment before you rent the GPUs. Open-source capacity planner: 39 models, 31 GPUs, 16 quants, transparent formulas, build-vs-buy calculator.

3-stage CIFAR-10 classification pipeline: EDA, 3 model architectures (CNN from scratch, EfficientNet-V2-S, WRN-28-10), GPU benchmarking. Best: 95.85% test accuracy. PyTorch + mixed precision (AMP/FP16).

LLM inference: compare activation and KV-cache sparsity through memory-traffic models and GPU decoding benchmarks.

Reproducible research for measuring agentic AI inference outcomes on NVIDIA Blackwell across cloud environments.

Public, one-way mirror of jgdynamite10/blackwell-agentic-inference-lab. Development, issues, pull requests, releases, and security reporting live in the canonical repository.

benchHUB is a Python-based project to parse, aggregate, and visualize system and performance benchmarks. It includes a Streamlit dashboard to display and compare results.

Standardized bare-metal hardware benchmark suite measuring LLM prefill TFLOPS, decode memory bandwidth saturation, and attention kernel efficiency

Professional benchmarking guide for evaluating NVIDIA H200 GPU memory bandwidth using the STREAM Triad (FP32) benchmark with CUDA 12.x on Linux. Includes setup, execution, performance analysis, and multi-GPU benchmarking results.

Artifact-backed LLM serving performance lab for vLLM baselines, official metrics, GuideLLM checks, and SGLang/PD scaffolding

RTX 4090 measurement harness - paired A/B levers, drift and control-spread gates, retraction chain

Reproducible GPT-2 distributed-training benchmarks on 1-8 V100 GPUs using Slurm, PyTorch, DeepSpeed, NCCL, NVTX, and Nsight Systems.

Measured CUDA graphs trade-off in vLLM 0.23 on H100: per-token decode always faster, but throughput inverts at 32B under saturation (-5.15%) alongside -1.8% tok/J; cold-start cost decomposed into capture and torch.compile.