TensorCraft-HPC

English

简体中文

English

简体中文

Appearance

TensorCraft-HPCTechnical Whitepaper / Kernel Architecture Academy

A header-only C++/CUDA project that explains how modern AI kernels are designed, optimized, benchmarked, and evaluated, with the reader journey shaped like a systems paper instead of a generic docs portal.

Read the whitepaper
Enter the academy
Inspect evidence
⛭

Architecture as narrative

The site leads from system model to operator detail, so evaluators can understand why the repository is organized the way it is.

▣

Evidence over slogans

Benchmarks, methodology, and references sit next to each claim instead of being buried behind marketing language.

∿

Dual-theme diagrams

Logos, charts, and explanatory graphics are rebuilt to stay legible in both light and dark modes.

Read this project like a systems paper ​

Architecture narrative

A repository that explains the optimization path

Move from why the kernels exist, to how the system is layered, to what evidence supports the performance posture.

  • Whitepaper first. Understand the project thesis, kernel philosophy, and constraints before reading implementation detail.
  • Evidence second. Trace performance claims back to methodology and cited work.
  • Atlas third. Drop into operator-level reference only after the system model is clear.
92% FP16 GEMM relative to cuBLAS on A100
SM70–SM100 Compile-time coverage from Volta to Blackwell
Whitepaper + Academy Reader flow designed for interview and peer review
Where to start

Choose the path that matches what you need to prove

Each section has a distinct job in the argument, from conceptual framing to implementation inspection.

Evaluation lens

What makes TensorCraft-HPC worth evaluating

The project is designed to surface engineering judgment. The interesting question is not only whether a kernel is fast, but whether the optimization path, system boundaries, and supporting evidence remain legible.

  • Architecture clarity. Memory abstractions, feature detection, kernel boundaries, and hardware support are explicit design objects.
  • Progressive optimization. Readers can trace kernels from naive baselines to Tensor Core aware implementations.
  • Research literacy. The site connects implementation choices to papers, libraries, and competitor surfaces worth studying.

GPU Architecture Support

TensorCraft-HPC supports NVIDIA GPUs from Volta (2017) to Blackwell (2024), covering CUDA Compute Capability 7.0-10.0.

2017

Volta

GV100
SMSM70
CC7.0
Tensor Cores (1st Gen)NVLink 2.0HBM2
Supported
2018

Turing

TU102
SMSM75
CC7.5
Tensor Cores (2nd Gen)RT CoresGDDR6
Supported
2020

Ampere

GA100
SMSM80/SM86
CC8.0/8.6
Tensor Cores (3rd Gen)BF16TF32Sparsity
Supported
2022

Hopper

GH100
SMSM90
CC9.0
Tensor Cores (4th Gen)FP8Transformer EngineDPX
Supported
2024

Blackwell

GB100
SMSM100
CC10.0
Tensor Cores (5th Gen)NVLink 5.0FP4/FP8RAS
Supported
⚡CUDA 11.0 - 13.1

Released under the MIT License.

Copyright © 2024-present TensorCraft-HPC Contributors