README.md

---
license: mit
language:
- en
tags:
- gpu
- cuda
- kernels
- benchmarks
- code-generation
- agents
size_categories:
- n<1K
pretty_name: KernelBench-Hard Problems
---

# KernelBench-Hard — Problem Definitions

The 7 problem definitions for **KernelBench-Hard**, a benchmark for autonomous LLM coding agents writing GPU kernels on a single Blackwell GPU (RTX PRO 6000, sm_120, CUDA 13.2).

Companion datasets:
- [`Infatoshi/kernelbench-hard-runs`](https://huggingface.co/datasets/Infatoshi/kernelbench-hard-runs) — 84 agent transcripts, winning solutions, leaderboard, reward-hack annotations
- Live site: https://kernelbench.com/hard
- Methodology blog: https://kernelbench.com/blog/hard
- Source repo: https://github.com/Infatoshi/kernelbench.com (monorepo)

## Problems

| id | task | shapes | regime |
| --- | --- | --- | --- |
| `01_fp8_gemm` | FP8 (E4M3) GEMM with bf16 accumulation | 8 | compute-bound |
| `02_kda_cutlass` | Kimi Delta Attention forward (CUTLASS) | 4 | compute-bound |
| `03_paged_attention` | Paged-attention decode (vLLM-style) | 6 | memory-bound |
| `04_kahan_softmax` | Numerically stable softmax with Kahan compensation | 5 | memory-bound |
| `05_topk_bitonic` | Top-k via bitonic select | 6 | memory-bound |
| `06_sonic_moe_swiglu` | Sonic-MoE forward with SwiGLU | 4 | compute-bound |
| `07_w4a16_gemm` | W4A16 weight-only quantized GEMM | 8 | compute-bound |

## File layout per problem

Each `0X_<name>/` directory contains:

| file | purpose |
| --- | --- |
| `reference.py` | The PyTorch reference implementation. The agent must match its output. |
| `check.py` | Correctness harness — reference vs. submission with `torch.allclose` tolerances |
| `benchmark.py` | Timing harness — L2-flush + 30-trial median, prints `shape= variant= tflops= gbps= ms=` lines |
| `problem.yaml` | Metadata: regime (compute/memory bound), tolerance, dtype, shapes |
| `shapes.py` | Iterable of input shapes the benchmark runs |
| `sota.py` | Hand-written SOTA reference (when available) for the upper-bound row in the leaderboard |
| `PROMPT.txt` | The exact prompt fed to the agent harness |

## Scoring (peak_fraction)

For each (model, problem) cell, we compute `peak_fraction` ∈ [0, 1] as:

```
peak_fraction = geomean over shapes of (achieved_throughput / hardware_peak)
```

where `achieved_throughput` is TFLOPS for compute-bound or GB/s for memory-bound, and `hardware_peak` is the sm_120 spec (peak fp8 TFLOPS or peak HBM bandwidth). This rewards approaching the hardware ceiling rather than the easier-to-game "speedup over PyTorch."

A solution must first pass `check.py` (correctness) before it gets a `peak_fraction`.

## Hardware

- **GPU**: NVIDIA RTX PRO 6000 Blackwell Workstation
- **SM**: sm_120a (Blackwell)
- **VRAM**: 96 GB GDDR7
- **Peak HBM bandwidth**: ~1800 GB/s
- **CUDA**: 13.2 / NVCC 12.8 / Driver 595.58.03
- **Host**: Ryzen 9950X3D, 92 GB DDR5

## Rubric leaks (known issues)

Two of the seven problems leak the rubric — meaning the easiest path to a high score involves something other than writing a fast correct kernel. We publish anyway with these documented inline in `benchmarks/hard/SPEC.md` and per-cell in the runs dataset's annotation files:

- **`01_fp8_gemm`** — agents that downcast to bf16 + tensor cores get ~80–90% peak without doing the actual fp8 quantization. The judge model catches some of these but not all. See annotations with `verdict: rubric_leak`.
- **`04_kahan_softmax`** — agents that skip Kahan compensation pass `check.py`'s atol within float32 limits and get a free ~2× speedup. The numerical instability only shows up at extreme magnitudes that the test inputs don't probe.

These are documented honestly because (a) we want the community to fix them, and (b) the rubric leaks themselves are interesting reward-hacking examples.

## How to use

```python
from datasets import load_dataset
# Or just clone the repo
# git clone https://huggingface.co/datasets/Infatoshi/kernelbench-hard-problems
```

To run a problem locally with your own kernel:

```bash
cd 01_fp8_gemm
# Drop your solution at solution.py
uv run python check.py        # verifies correctness
uv run python benchmark.py    # measures throughput
```

## License

MIT. Cite as:

```
@misc{kernelbench-hard-2026,
  author = {Arledge, Elliot},
  title = {KernelBench-Hard: A GPU Kernel Engineering Benchmark for Autonomous Coding Agents},
  year = {2026},
  url = {https://kernelbench.com/hard},
  note = {Built on top of KernelBench (Ouyang et al., 2025).}
}
```

Original KernelBench: Ouyang et al., https://github.com/ScalingIntelligence/KernelBench