---
license: apache-2.0
tags:
- cuda
- kernel
- gpu-optimization
- hpc
---

# ada_tensor_core_fp16

Ada Tensor Core FP16 GEMM prototype with WMMA-backed 32x64x16 correctness and timing harness.

This repository contains the standalone CUDA source for the `ada_tensor_core_fp16` lane from
 the PyC kernel lab. It is a source artifact for inspection and benchmarking;
it is not a precompiled binary and the result below is not a universal ranking.

## Performance

| Kernel | GPU / architecture | Shape | Best recorded result | Evidence |
|---|---|---|---|---|
| `ada_tensor_core_fp16` | not recorded | 1024x1024x1024 | 0.134 ms | Measured on not recorded, shape 1024x1024x1024; evidence `hopper-tensorcore-bringup-20260421T191817Z.json`. |

![Performance plot](performance.svg)

The result is reported with the original campaign's timing and correctness
context. Compare kernels only when GPU, CUDA version, matrix shape, warmup,
repeats, and reference/correctness mode match.

## Source

- `kernel.cu` — copied from `kernels/prototypes/ada/tensor_core/kernel.cu`.
- Original lane tags: `cuda, matmul, ada, sm89, prototype, tensor-core, fp16`.

## Build/run contract

```text
{nvcc} -O3 -std=c++17 -lineinfo -DPYC_ADA_TENSOR_CORE_USE_BF16=0 -gencode arch=compute_89,code=sm_89 -gencode arch=compute_89,code=compute_89 {source} -o {build_dir}/{name}
{build_dir}/{name} 1024 1024 1024 10 50
```