FAQ
Direct answers.
What Kernalin is, what it supports, how it verifies, what is live and what is coming soon. If something is missing, ask [email protected].
General
What is Kernalin?
Kernalin is a GPU kernel optimization platform. It profiles PyTorch, Triton, CUDA and ONNX workloads, uses AI (Claude) 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 performance regression gate.
How is Kernalin different from asking an AI assistant to optimize a kernel?
An assistant gives you code and a claim. Kernalin runs every rewrite through a five-stage correctness gate, an anti-cheat audit and a Bayesian speed verdict in a sandbox, keeps only rewrites that pass, shows you every rejected one, and then re-checks the change in CI. You merge evidence, not a claim.
Which frameworks and file types does Kernalin support?
Uploads and linked files can be PyTorch or Triton (.py), CUDA C++ (.cu, .cuh), ONNX (.onnx) and PyTorch checkpoints (.pt, .pth), up to 100 MB. Benchmarks cover PyTorch eager, torch.compile, ONNX Runtime and TensorRT in fp32, fp16, bf16 and fp64.
Do I need a GPU to try it?
No. Profiling, benchmarking and AI rewrites of PyTorch kernels run in Kernalin’s CPU sandbox. Triton and CUDA C++ rewrites, TensorRT, GPU timings and hardware counters need a GPU runner, which you can attach yourself (a RunPod setup script is included).
Why are some of your published numbers from a CPU sandbox?
Because those are the runs we recorded end to end, and we only publish recorded numbers with their date and context. The H100 numbers come from a dry run on 27 September 2026. Features still waiting for GPU verification never show results on this site.
How much does Kernalin cost?
Pricing is not finalized. Sign-up is open and self-serve; teams and enterprises talk to us about terms. There are no published prices yet.
What is live today and what is coming soon?
Live today: profiling, memory and warp-divergence analysis, benchmarks with Bayesian verdicts, schedules and alerts, AI rewrites with verification, optimize-to-PR, the CI gate, GitHub integration, hardware comparison, PyTorch → Triton conversion, workspaces, roles, SSO and the audit log. Coming soon: Deep profile, the stricter verifier, multi-agent search, vLLM serving benchmarks, the TensorRT precision matrix, benchmark suites, Ways to save, the MCP server and the new app navigation. The changelog has dates.
Results
Has Kernalin actually shipped a faster kernel?
Yes, end to end, on 9 October 2026: in Kernalin’s CPU sandbox, Claude rewrote a PyTorch attention kernel, Kernalin kept 1 of 3 rewrites as verified 7.64× faster, opened pull request #3 in JashomTech/kernalin-demo, the CI gate passed, it was merged, and a fresh benchmark measured 3.2–4.2× faster in eager mode and 1.8–2.4× under torch.compile. On an NVIDIA H100 on 27 September 2026, a Claude rewrite of a deliberately wasteful GELU was verified 3.48× faster.
Verification
How does Kernalin know an AI-written kernel is correct?
Each candidate must match the reference across a smoke test, a shape sweep, adversarial input scaling, a determinism check and edge cases such as non-power-of-two shapes, with per-dtype tolerances. An anti-cheat audit then checks provenance, fresh inputs, poisoned buffers and whether the output responds to every input. Only then is speed judged.
What is reward hacking in kernel optimization, and does Kernalin catch it?
Reward hacking is when an optimizer "wins" the benchmark without doing the work: skipping operations, ignoring inputs, returning memorized outputs or special-casing the measured shape. Kernalin’s anti-cheat audit is built to catch these: fresh and poisoned inputs, ignored-input checks and red-team tests, and a rewrite that trips any of them is rejected.
What does "faster" mean in Kernalin?
A rewrite or a run is called faster only when a Bayesian analysis gives at least 95% probability that it lies outside a ±2% band of practical equivalence, from at least 8 samples per side and a 2,000-resample bootstrap. Otherwise Kernalin says "equivalent", or abstains.
Profiling
Can I import Nsight Compute or torch.profiler output?
Yes. Kernalin imports Nsight Compute source CSV, torch.profiler JSON traces, torch.cuda.memory_stats evidence and ourprof bundles today. Nsight Systems, .ncu-rep and ONNX Runtime profile imports are coming soon. Imported files are parsed, never executed.
What is Deep profile?
Coming soon (GPU verification pending). One click will run an op profile, then an Nsight Systems timeline with gap advice, then Nsight Compute on the top kernels (by default 5, until they cover 80% of GPU time) with guided analysis and a roofline. Profiler overhead is recorded and kept out of benchmarks.
Why can’t Kernalin read hardware counters on my cloud GPU?
Many cloud GPU pods deny access to performance counters (for example ERR_NVGPUCTRPERM on the H100 pod used on 27 September 2026). Kernalin checks first and refuses a counter job up front with the reason, rather than returning empty numbers.
Benchmarking
Which runtimes can Kernalin benchmark?
PyTorch eager, torch.compile, ONNX Runtime and TensorRT, in fp32, fp16, bf16 and fp64, at the batch sizes you choose: up to 48 configurations per run. Every runtime’s output is checked against the reference, and one that diverges is excluded from the ranking.
Can Kernalin alert us when performance regresses?
Yes. Scheduled benchmarks run automatically and email a report, a regression check runs after every stored run, and alerts go to Slack, Microsoft Teams, email or an HTTPS webhook.
CI gate
How does the CI regression gate work?
The rnx-benchmark GitHub Action benchmarks the pull request and uploads the run; Kernalin compares it with the newest baseline on the base branch in the same environment and returns PASS, WARN or BLOCK. A real slowdown or a broken runtime fails the check. If Kernalin is unreachable the check warns instead of failing, and an intended slowdown can be accepted as the new baseline.
Will the CI gate make my builds flaky?
It is designed not to: verdicts are statistical rather than a fixed percentage, comparisons are only made within the same environment, thresholds for shared GitHub-hosted runners can be loosened in a config file, report-only mode exists, and an outage never fails your build.
Can Kernalin add the CI workflow for me?
Coming soon. "Add CI gate" will open a pull request that adds .github/workflows/kernalin.yml for you; today you add the rnx-benchmark action to a workflow yourself (about ten lines of YAML).
Hardware & cost
How does Kernalin compare GPUs on cost?
Compare hardware puts 2 to 6 runs of the same workload side by side on latency, throughput and cost per million items, using the hourly price set for each machine. It only calls one machine faster when the confidence intervals do not overlap, and exports CSV or Markdown.
Does Kernalin show live GPU utilization?
Cost monitoring and idle-GPU alerts are live but still awaiting verification on real GPU fleets. A richer utilization view (SM, Tensor Core and DRAM activity from NVML/DCGM, and GPU health monitors) is coming soon.
LLM serving
Does Kernalin benchmark LLM serving?
Coming soon. vLLM serving benchmarks (TTFT, TPOT, inter-token and end-to-end latency, goodput under an SLA, VRAM and the batching knee) are built and waiting for GPU verification. vLLM only for now; no results are published yet.
Conversions
Can Kernalin convert PyTorch to Triton?
Yes, for 11 ops: relu, gelu, silu, sigmoid, tanh, abs, add, mul, fused silu_mul, fused add_relu and row softmax. Every generated kernel runs through the correctness gate; on an H100 the converted gelu and silu_mul matched PyTorch to a max absolute error of 4.5e-8 and 2.2e-8.
Does Kernalin support TensorRT and FP8?
TensorRT runs in benchmarks today on GPU runners, held to the same correctness check as every runtime. A precision matrix across FP32, TF32, FP16, BF16, INT8 and FP8 is coming soon (GPU verification pending); INT8 and FP8 need NVIDIA ModelOpt on the runner.
Benchmark suites
Has Kernalin published KernelBench or speed-of-light results?
No. Benchmark suites (KernelBench-Verified, SOL-ExecBench), the speed-of-light score and public proof pages are built and coming soon, but no suite result has been recorded yet, so none is cited. Proof pages will refuse to publish unless the run used a GPU with locked clocks and a flushed L2.
Teams & security
Does Kernalin support SSO?
Yes. OIDC (with PKCE) and SAML 2.0 with DNS-verified domains and a break-glass admin sign-in are live. A Microsoft Entra ID live check is still pending, and MFA for password accounts is not available yet.
Is it safe to upload our kernels?
Uploads are scanned statically and never executed on intake. Jobs run in a sandbox with no network, a read-only root and an unprivileged user, and every result coming back from a runner is validated as untrusted. Kernel source is sent to Anthropic’s Claude (via OpenRouter) only when you use Explain or Optimize. Details are in the privacy policy.
Integrations
Can I use Kernalin from Claude Code, Cursor or Codex?
Coming soon. The kernalin-mcp server gives coding agents 9 read tools and 3 write tools (upload, benchmark, open a PR) with workspace agent tokens, and every agent action is audited under your name.
Point Kernalin at your slowest kernel.
Sign-up is open and self-serve. Connect GitHub or upload a file, and get a profile, a benchmark and verified rewrites, without changing how you ship.