NKIBench

NKIBench is a benchmark of AWS Neuron Kernel Interface (NKI) kernels paired with NumPy reference implementations. Each task provides a specification, a ground-truth NumPy forward pass, and an optimized NKI kernel targeting AWS Trainium devices, together with tooling to compile, check numerical correctness, and measure on-device latency.

Dataset structure

NKIBench/
├── seeds/               # YAML task specifications (shape-agnostic templates)
├── reference/           # NumPy reference implementations with concrete shapes
├── kernels/             # Initial NKI kernels (one per case)
├── summary.json         # Index mapping task → case → {seed, reference, kernel}
├── save_fields.json     # Useful fields of the neuron profiler
└── kernel_wrapper.py    # Profiler: compile, correctness check, latency benchmark

summary.json

The canonical index. Each entry maps a task name to one or more parameter cases and the files that implement them:

{
  "matmul": {
    "seed": "./seeds/matmul.yaml",
    "cases": {
      "3": {
        "values": {"K": 5120, "M": 4096, "N": 12288},
        "impls": [{
          "task":   "./reference/matmul_M4096_N12288_K5120_numpy_2.py",
          "kernel": "./kernels/matmul_M4096_N12288_K5120_0.py"
        }]
      }
    }
  }
}

seeds/*.yaml

A shape-agnostic specification: the task name, its symbolic parameters, an input generator, and a NumPy forward implementation.

test_name: matmul
parameters: [M, N, K]
input: |
  lhs = np.random.normal(loc=0, scale=1.0, size=(M, K)).astype(np.float32)
  rhs = np.random.normal(loc=0, scale=1.0, size=(K, N)).astype(np.float32)
  return [lhs, rhs]
impl: |
  def forward(lhs, rhs):
    return np.matmul(lhs, rhs)

reference/*.py

A shape-concrete NumPy reference. Exposes:

• get_inputs() — produces randomized numpy input tensors.
• forward(*inputs) — ground-truth computation.
• transform_to_nki_inputs(inputs) — reshapes numpy inputs into the tile layout the NKI kernel expects.
• transform_nki_outputs(k_res, ref) — reshapes kernel outputs back to reference layout.

kernels/*.py

Initial NKI kernels using neuronxcc.nki. Each file defines a kernel function decorated with @nki.jit.

Usage

# Clone the dataset
hf download Genghan/NKIBench --repo-type dataset --local-dir NKIBench
cd NKIBench

# Profile one kernel on an AWS Neuron-enabled instance (e.g. trn1 / inf2).
# Requires: neuronx-cc, neuronx-runtime, and the `neuron-profile` CLI.
import json
from kernel_wrapper import NKIKernel

summary = json.load(open("summary.json"))
save_fields = json.load(open("save_fields.json"))
case = summary["matmul"]["cases"]["3"]["impls"][0]

k = NKIKernel(program_path=case["kernel"], base_numpy_path=case["task"])
result = k.profile(save_fields=save_fields)

print("compiled:", result.compiled)
print("correct :", result.correct)
print("latency :", result.metadata.get("latency"), "ms")

NKIKernel.profile() compiles the kernel, validates numerical correctness against the NumPy reference over multiple random seeds (L2-norm relative tolerance 2e-5), and benchmarks latency via neuron-profile. float16 inside a kernel is rejected to avoid silent precision loss.

Citation

If you use NKIBench in your work, please cite the repository.

@article{zhang2026accelopt, 
  title={AccelOpt: A Self-Improving LLM Agentic System for AI Accelerator Kernel Optimization}, 
  author={Zhang, Genghan and Zhu, Shaowei and Wei, Anjiang and Song, Zhenyu and Nie, Allen and Jia, Zhen and Vijaykumar, Nandita and Wang, Yida and Olukotun, Kunle}, 
  journal={Proceedings of Machine Learning and Systems}, 
  volume={9}, 
  year={2026} 
 }