# Kernalin

> Kernalin is a GPU kernel optimization platform: it profiles PyTorch, Triton, CUDA and ONNX workloads, uses AI to rewrite slow kernels, keeps a rewrite only when a verifier shows it matches the reference (within per-dtype tolerance) and is measurably faster, and ships it as a GitHub pull request guarded by a CI regression gate.

Kernalin is available at https://kerneln.com/app (open, self-serve sign-up; pricing not finalized). Numbers on the site come only from recorded runs, each with its date and context: a dry run on an NVIDIA H100 on 2026-09-27 and an end-to-end demo pull request on 2026-10-09 (CPU sandbox). Features marked "Coming soon" are built and tested but not yet released; features marked "GPU verification pending" have no published results. A full plain-text version of the site's facts is at https://kerneln.com/llms-full.txt.

## Product
- [Kernalin](https://kerneln.com/): Kernalin profiles GPU kernels, has AI rewrite the slow ones, proves each rewrite correct and faster, and ships it as a PR behind a CI regression gate.
- [How Kernalin works: profile, rewrite, verify, ship](https://kerneln.com/how-it-works): The Kernalin loop: connect GitHub, profile and benchmark a kernel, let AI propose rewrites, keep only verified ones, open a PR and gate every merge in CI.
- [GPU and CUDA profiling with an Amdahl plan](https://kerneln.com/profiling): Kernalin ranks every PyTorch op by time, imports Nsight Compute and torch.profiler output, analyzes memory and warp divergence, and says what to fix first.
- [GPU benchmarking with Bayesian verdicts and alerts](https://kerneln.com/benchmarking): Benchmark PyTorch eager, torch.compile, ONNX Runtime and TensorRT with Bayesian verdicts, honest timing, schedules and regression alerts.
- [AI GPU kernel optimization, kept only when verified](https://kerneln.com/optimization): Claude rewrites your PyTorch, Triton or CUDA kernels; Kernalin keeps a rewrite only if it passes a correctness gate, anti-cheat audit and speed verdict.
- [Verifying AI-generated GPU kernels: correctness and anti-cheat](https://kerneln.com/verification): How Kernalin verifies AI-written GPU kernels: a five-stage correctness gate, an anti-cheat audit against reward hacking and a Bayesian speed verdict.
- [CI performance regression gate for GPU kernels](https://kerneln.com/ci-gate): A GitHub Actions gate that benchmarks every PR against main in the same environment, blocks real GPU regressions and never fails on an outage.
- [GPU cost per result: compare hardware and find savings](https://kerneln.com/cost-intelligence): Compare a workload across CPUs and GPUs on latency, throughput and cost per million items, attach your own GPU runners and watch idle spend.
- [LLM inference benchmarking on vLLM: TTFT, TPOT, goodput](https://kerneln.com/llm-serving): Kernalin’s vLLM serving benchmarks measure TTFT, TPOT, inter-token latency, goodput under an SLA, VRAM and the batching knee. Coming soon; GPU verification pending.
- [PyTorch to Triton conversion and TensorRT precision](https://kerneln.com/conversions): Convert 11 PyTorch ops to verified Triton kernels and benchmark TensorRT next to eager and torch.compile. TensorRT precision matrix coming soon.
- [Kernel benchmark suites, speed-of-light score and proof pages](https://kerneln.com/benchmark-suites): KernelBench-Verified and SOL-ExecBench suites, a speed-of-light score and public proof pages that refuse to publish without locked clocks. Coming soon.
- [Teams, roles, SSO and audit log for GPU optimization](https://kerneln.com/enterprise): Workspaces with viewer, engineer and admin roles, OIDC and SAML single sign-on, an audit log with CSV export, and sandboxed jobs.
- [Integrations: GitHub, Actions, Slack, Teams, MCP](https://kerneln.com/integrations): Kernalin works with GitHub, GitHub Actions, Slack, Microsoft Teams, email and webhooks. An MCP server for Claude Code, Cursor and Codex is coming soon.

## Solutions
- [Use cases: LLM inference, GPU cost and CI for kernels](https://kerneln.com/use-cases): How teams use Kernalin: faster LLM inference kernels, lower GPU cost per result, CI performance gates, and runtime choices backed by honest numbers.
- [Kernalin for LLM inference teams](https://kerneln.com/for/llm-inference): For teams serving LLMs: find the attention and MLP kernels that own your latency, rewrite them with AI under a correctness gate, and gate every change in CI.
- [Kernalin for reducing GPU cost](https://kerneln.com/for/gpu-cost): Lower GPU cost per result: faster verified kernels, hardware comparison on cost per million items, and idle-GPU detection.
- [Kernalin for CI performance regression testing](https://kerneln.com/for/ci-regression): Catch GPU kernel performance regressions in pull requests with statistically sound benchmarks that never fail your build because a service was down.
- [Kernalin compared with Nsight, torch.profiler and AI coding tools](https://kerneln.com/compare): A factual comparison of Kernalin with NVIDIA Nsight, torch.profiler, benchmark scripts and AI coding assistants, and where each one fits.

## Resources
- [FAQ: GPU kernel optimization with Kernalin](https://kerneln.com/faq): Direct answers about Kernalin: what it is, supported frameworks and GPUs, how verification works, security, pricing and what is coming soon.
- [Docs: get started with Kernalin](https://kerneln.com/docs): Get started with Kernalin: the workload contract, connecting GitHub, profiling and benchmarking, optimize-to-PR, and the rnx-benchmark CI gate.
- [Changelog](https://kerneln.com/changelog): Kernalin changelog from the real sprint history: what went live on kerneln.com/app and what is coming soon, dated.
- [Blog: GPU optimization notes with real data](https://kerneln.com/blog): Engineering notes from Kernalin on verified GPU kernel optimization, benchmarking and profiling, written from recorded runs.
- [What we measured on an NVIDIA H100 dry run](https://kerneln.com/blog/h100-dry-run): Kernalin on an H100 (27 Sep 2026): a verified 3.48× Claude rewrite, TensorRT fp16 1.94–2.56× faster, one TensorRT result rejected, 33% of VRAM stranded.
- [Why "faster" needs a noise floor: Bayesian verdicts for GPU benchmarks](https://kerneln.com/blog/faster-needs-a-noise-floor): A 17.6% speedup Kernalin called "equivalent": how bootstrap resampling and a ±2% equivalence band stop benchmark noise becoming false claims.
- [One kernel, end to end: a slow attention kernel to a merged, verified PR](https://kerneln.com/blog/one-kernel-end-to-end): Claude rewrote a slow PyTorch attention kernel; Kernalin kept the one rewrite verified 7.64× faster, shipped it as PR #3 and re-measured 3.2–4.2× faster.

## Company
- [Pricing](https://kerneln.com/pricing): Kernalin pricing: sign-up is open and self-serve while pricing is being finalized. Talk to us for teams and enterprise. What is live today and what is coming soon.
- [Security: how Kernalin runs untrusted kernels safely](https://kerneln.com/security): How Kernalin runs untrusted GPU code: static intake scanning, network-isolated sandboxes, results validated as untrusted, audit logs and disclosure.
- [About Kernalin](https://kerneln.com/about): Why Kernalin exists: AI can write fast GPU code, but only verification makes it safe to ship. Who builds it and how we decide what to claim.
- [Contact](https://kerneln.com/contact): Contact Kernalin: hello@kerneln.com for questions, teams and pricing; security@kerneln.com for vulnerabilities. Or sign up and try it yourself.

## Optional
- [Full facts for LLMs](https://kerneln.com/llms-full.txt): every feature with its release stage, recorded results with sources, FAQ and changelog in one file.
- [Privacy Policy](https://kerneln.com/privacy): What Kernalin collects, why, who processes it (including Anthropic via OpenRouter, Cloudflare and RunPod), how long we keep it, and your rights.
- [Terms of Service](https://kerneln.com/terms): The terms for using kerneln.com, the Kernalin app, the rnx CLI and the Kernalin VS Code extension.
- Contact: hello@kerneln.com (general, teams, pricing); security@kerneln.com (security)