Academy

The academy section teaches readers how to read the repository as a system, not just how to compile it.

What you learn here

• how the repository layers kernels, memory primitives, and hardware capability checks
• how the optimization path progresses from naive kernels to Tensor Core aware variants
• how to evaluate benchmark claims with the right methodology in mind

Suggested reading order

1. Quick start
2. Architecture lessons
3. Modern C++ and CUDA
4. Benchmarking
5. Examples

Optimization path

Optimization Path

Progressive optimization from naive to Tensor Core implementation

1

Naive

2

Tiled

3

Double Buffer

4

Tensor Core

1

Naive

Direct triple loop implementation

Global memory accessNo parallelism optimization

5%

2

Tiled

Shared memory blocking

Block-level tilingShared memory reuseCoalesced access

45%

3

Double Buffer

Pipeline memory access

PrefetchingLatency hidingWarp synchronization

75%

4

Tensor Core

WMMA hardware acceleration

WMMA instructionsMixed precisionMaximum throughput

92%

📊 Performance relative to cuBLAS on A100 80GB