STREAM, for lots of devices written in many programming models

Provides a set of benchmarks that can be used to measure the memory bandwidth performance of CPU's

A Fast DNN Accelerator Design Space Exploration Framework.

Modern Memory Bandwidth and Latency Benchmarks

Demo code accompanying the talk "Implementing memory locality optimizations in OpenFOAM based code"

A macOS memory benchmark tool that measures read/write performance in GB/s.

Main Memory Bandwidth Monitoring

General-purpose compile-time Expression Templates library for C++

Measure and visualize why LLM inference is slow: bottleneck analysis, model dissection, KV-cache, GEMM/GEMV, quantization, and memory-bound decoding.

C++23 benchmarking framework with 6 profiler backends, CUDA GPU support, statistical regression detection, cross-compilation for 5 architectures, and CLI tools for analysis and visualization.

Reproducible Pascal GPU Unified Memory benchmark with Nsight and nvprof profiling

Console UI for watching Memory stats

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

Benchmark for memory bandwidth

OpenCL benchmarking tool to measure host-device bandwidth and kernel global memory throughput across GPUs and CPUs.

Production-grade Triton kernel fusing residual add + RMSNorm + packed QKV projection into a single GPU launch for decoder-only transformer inference (Llama-3, Mistral, Qwen2). +2.4% tok/s, -1.5 GB VRAM on A10G.

A practical model (with math + Python) to tell if you’re compute-, memory-, or network-bound—and what to buy next

Transform-domain representation enabling 3–4× storage reduction with direct ANN search and novel multi-resolution signals. UK patent application under accelerated examination (Green Channel).

Bandwidth-focused LLaMA token-generation latency benchmarking on Apple Silicon M4 vs M2 with TTFT, PTL, KV-cache, quantization projections, and reproducible architecture analysis.

CUDA SpMV kernels (scalar, warp-per-row, ELL) on NVIDIA A100 benchmarked against cuSPARSE on SuiteSparse matrices, plus AVX2 + cache-tiled CPU baselines on Intel Xeon Gold. Vector kernel reaches 98-110% of HBM2 peak, beating cuSPARSE by 24-56% on regular matrices.