Benchmarks

This section presents performance benchmarks for TensorCraft-HPC kernels compared against NVIDIA's optimized libraries (cuBLAS, cuDNN, cuSPARSE).

Overview

All benchmarks are measured on:

Parameter	Value
GPU	NVIDIA A100 80GB
CUDA Version	12.4
Data Type	FP16 (Tensor Core)
Measurements	Average of 100 runs

Performance Summary

Kernel	Reference	Relative Performance
GEMM (FP16)	cuBLAS	92%
FlashAttention	cuDNN	85%
LayerNorm	cuDNN	95%
Conv2D	cuDNN	78%
SpMV (CSR)	cuSPARSE	88%

Detailed Benchmarks

• GEMM Performance — Detailed GEMM benchmark across matrix sizes
• Attention Performance — FlashAttention benchmark analysis

Benchmarking Philosophy

TensorCraft-HPC prioritizes readability and educational value over raw performance. Our benchmarks serve to:

1. Validate correctness — Ensure optimized versions produce accurate results
2. Demonstrate progress — Show improvement from naive to optimized
3. Guide optimization — Identify bottlenecks and optimization opportunities

Performance vs Readability

While we strive for competitive performance, we sometimes choose clearer code over marginal speed improvements. The goal is learning, not beating cuBLAS.

Running Your Own Benchmarks

# Build benchmarks
cmake --preset dev
cmake --build --preset dev

# Run GEMM benchmark
./build/dev/benchmarks/gemm_benchmark

# Run all benchmarks
ctest --preset dev -L benchmark

Benchmark Configuration

Benchmarks can be configured via environment variables:

# Set matrix size for GEMM benchmark
export TENSORCRAFT_BENCH_SIZE=4096

# Set number of warmup runs
export TENSORCRAFT_BENCH_WARMUP=10

# Set number of measured runs
export TENSORCRAFT_BENCH_RUNS=100