Nodus ComputeRL ENVIRONMENTS

GPU KERNELS · REINFORCEMENT LEARNING

Real GPU work.
A higher bar.

RL environments for agents that write GPU kernels. Evaluate against expert production kernels, with sealed graders, expert-reviewed solutions, and scored trajectories.

Explore environments Book a meeting
Backed byY CombinatorOpenAI
Built by engineers previously at
  • OpenAI
  • SpaceXAI
  • Citadel Securities
  • Google
  • Snowflake
  • Scale AI
  • Tesla
  • AWS
  • Cadence
  • Uber
  • Coinbase
  • Capital One
0/15two-hour attempts by Claude Opus 5.5 at max effort passed attention backward
1.10×expert-reviewed solution vs FlashAttention-4 on B300, 4 sealed grades with hidden sizes
12/12planted cheats scored the reward floor
88GPU-kernel RL environments
234 environment versions run against frontier agents974 recorded agent attemptsB300 and 8× B300 environments8× A100 distributed training environmentsAs of Oct 23, 2026

01 / The environments

Hard for the model.
Solved by the expert.

Each environment ships with a sealed grader, an expert-reviewed solution and scored agent trajectories.

01 / ATTENTIONB300

Attention backward.

256-wide heads, sliding windows, and packed documents.

0/15Opus 5.5 passes
1.10×expert-reviewed solution
vs production baseline
02 / INFERENCEB300

Chunked prefill.

Long-context attention over a growing KV cache.

1/5Opus 5.5 passes
1.21×expert-reviewed solution
vs production baseline
EXPLORE THE SAMPLE

The real work, included.

  • 2 complete B300 environments
  • Sealed graders + verified solutions
  • 16 recorded agent attempts
  • Task cards + verification results
Download the sample ↓

ZIP · 82.4 MB. B300 required to run.

Opus 5.5 at max effort, up to 120 minutes per attempt. Attention backward counts 15 graded attempts across three task versions. Chunked prefill was screened on earlier task versions. Expert-reviewed solutions passed 3/3 reruns on B300.

02 / Distributed training

Eight GPUs. One training run.
Frontier models stop short.

Each environment asks an agent to make a real mixture-of-experts model train faster on 8× A100 without changing the training math. Every step is checked against a reference run, and the bar is the best published training recipe.

01 / MOE TRAINING8× A100

Qwen3-30B-A3B.

128 experts with top-8 routing, expert parallel layout and grouped kernels.

1/5Opus 5.5 passes
2.7×verified solution
vs torchtitan
02 / MOE TRAINING8× A100

DeepSeek-V4-Flash.

256 experts, compressed sparse attention and hyper-connections.

0/5Opus 5.5 passes
4.4×verified solution
vs NVIDIA NeMo
EXPLORE THE SAMPLE

Every attempt, recorded.

  • 2 complete 8× A100 tasks
  • 12 recorded Opus 5.5 attempts
  • Scores + task cards
  • Graders available under NDA
Download the sample ↓

ZIP · 9.8 MB. 8× A100 80 GB required to run.

Opus 5.5 at max effort, up to 120 minutes per attempt. Passes count the first 5 graded attempts on each task version, and every attempt was correct. Verified solutions passed 3/3 sealed grades on 8× A100. Both models run at reduced depth so full training fits 8× 80 GB GPUs.

03 / The evidence

Measured, not claimed.
Every number is a sealed grade.

Real attempts, real gradings and real cheats, all on NVIDIA B300. Rewards follow speed against the expert kernel, so agents get a learning signal on every attempt.

FIG 1 / ATTENTION BACKWARD

Every Opus attempt fell short of the expert.

Speed of every graded two-hour Opus 5.5 attempt on attention backward versus FlashAttention-415 attempts from 0.50 to 1.04 times FlashAttention-4 speed, median 0.85. The expert-reviewed solution scored 1.11 to 1.12 times in 4 sealed grades. 0.4×0.6×0.8×1.0×1.2× FlashAttention-4 = 1.0× Opus median 0.85× Expert 1.11× Speed vs FlashAttention-4 on B300, timed sizes
  • Opus 5.5 attempt, 15 graded
  • Expert-reviewed solution, 4 sealed grades
Claude Opus 5.5 at max effort, two hours per attempt, three task versions, Oct 18 to 23, 2026. Attempts lost to infrastructure are excluded.

FIG 2 / REWARD SIGNAL

Reward tracks speed, not just pass or fail.

Reward against measured speed versus the expert kernelCorrect kernels earn log2 of their speed versus the expert, so reward rises smoothly from minus 3.7 for slow starting kernels to plus 0.27 for expert-reviewed solutions. Twelve planted cheats all scored the floor of minus 24. 1/16×1/8×1/4×1/2×1×2×+10-1-2-3-4 −24 12 planted cheats Starting kernels Opus attempts Expert-reviewed Measured speed vs expert kernel (log scale)
  • Starting kernel
  • Opus attempt
  • Expert-reviewed
  • Cheat
Correct kernels earn log2 of their speed against the expert kernel. Wrong, flagged or cheating kernels earn −24. Sealed grades from attention backward, chunked prefill and a sparse attention indexer, October 2026.

FIG 3 / GRADER INTEGRITY

Cheats that look fast still score the floor.

Reward paid to expert-reviewed solutions, slow correct kernels and planted cheatsExpert-reviewed solutions earned plus 0.09 to plus 0.27. Correct but slow starting kernels earned minus 1.6 to minus 3.7. All 12 planted cheats scored minus 24, including one the timer measured at 23.9 times the expert's speed. 0−8−16−24 ExpertCorrect, slowPlanted cheats One cache cheat timed at 23.9× the expert. Reward −24. Reward per sealed grade
  • Expert-reviewed
  • Correct, slow
  • Planted cheat
12 planted cheats across three environments: cached results, wrong precision, partial outputs and size-specific shortcuts. All 12 scored −24 in sealed grades on B300.

04 / How it's graded

One sealed grader.
No way to see the answers.

The agent writes a kernel. Everything that decides its reward runs where the agent cannot reach it.

  1. 01

    Sealed grader

    The grader, hidden sizes and expert code never enter the agent's workspace.

  2. 02

    Hidden sizes

    Each grade adds input sizes the agent never saw, next to the public ones.

  3. 03

    Fresh inputs

    New random values on every timed call, so cached or stale outputs fail.

  4. 04

    Same-run expert timing

    The expert kernel is timed in the same run on the same B300.

  5. 05

    Graded reward

    Correct: log2 of speed against the expert. Wrong or cheating: −24.

05 / The benchmark

A production kernel.
A different standard.

KernelBench asks models to beat PyTorch. Our environments ask them to beat expert production kernels that serve real models.

KERNELBENCH · LEVEL 312% → 87%

Public benchmarks are saturating

Best fast_1 vs PyTorch: o1 in the 2025 paper, STARK agents by March 2026.

Paper ↗ · Leaderboard ↗
NODUS · ATTENTION BACKWARD0/15

Opus 5.5 passes vs production

Max effort, up to 120 min per attempt, B300.

See the environment →

Different models, metrics and baselines. Shown for context, not as a ranking.

EvaluationKernelBenchNodus
Baseline to beatPyTorch eager / torch.compileProduction inference kernels
HardwareH100, L40SNVIDIA B300
PrecisionFP32BF16 · FP8 · FP4
GradingCorrectness + speedupCorrectness, speed and hidden shapes, sealed

KernelBench details from the official repository and paper.

For AI labs

Your model.
The next challenge.

Book a meeting or write to [email protected]