Skip to content

Changelog

What shipped, and when.

From the real sprint history. Live means it is in the app at kerneln.com/app today. Coming soon means built and tested, waiting for the next release.

    • Live Phone and narrow-window layout fixes, plus 14 QA fixes.
    • Coming soon Cmd-K, the "Add CI gate" bot pull request, the staged optimize flow, Ways to save and quiet notifications.
    • Coming soon Deep profile: op profile → Nsight Systems → Nsight Compute (GPU verification pending).
    • Coming soon TensorRT precision matrix, FP32 to FP8 (GPU verification pending).
    • Coming soon Benchmark suites, speed-of-light score and public proof pages; multi-agent optimizer (GPU verification pending).

    • Coming soon MCP server and agent skills for Claude Code, Cursor and Codex.
    • Coming soon GitHub App with per-repository install.
    • Coming soon Verifier to the KernelBench-Verified standard, with a public anti-cheat rules page.
    • Coming soon New five-area navigation with an alerts inbox and triage.
    • Coming soon vLLM serving benchmarks, first cut (GPU verification pending).
    • Coming soon Real GPU utilization and seven GPU health monitors (GPU verification pending).

    • Coming soon Change points on main and auto-bisect to the culprit commit.

    • Live Security hardening release: rebuilt sandbox handoff, account-takeover and scanner fixes, validation of every runner result.
    • Live Interleaved A/B timing for honest speed measurement, and a CI gate that never fails because Kernalin is down.
    • Live Safer optimize-to-PR and GitHub connection handling; 7-day restore for deleted workspaces.
    • Live Nightly encrypted backups with restore tests; staged releases with rollback.
    • Live Runtime plugins, runner capability labels and a profiler import framework.
    • Coming soon Plain-language verdicts, one-click benchmark, Home feed and a page per workload.
    • Coming soon Imports for Nsight Systems, Nsight Compute reports and ONNX Runtime profiles.
    • Coming soon Prometheus metrics, a Grafana dashboard, utilization alerts and the Stack advisor.
    • Coming soon PR comment with code-owner acceptance, soft-fail mode and thresholds in kernalin.yml.

    • Live Optimized kernels sent back to GitHub as pull requests; compare a run with the previous one.

    • Live GPU cost monitoring, idle-GPU detection and cost-spike alerts (GPU verification pending).
    • Live Warp-divergence analysis from Nsight Compute counters.
    • Live Single sign-on (OIDC and SAML) with verified domains.
    • Live GitHub integration: link repositories, branches and files, with auto-sync on every commit.

    • Live Scheduled benchmarks with history and emailed reports.
    • Live Compare one workload across GPUs and CPUs, with cost per million items.

    • Live Organization workspaces with email invitations; optimization projects.
    • Live Email verification and password reset.
    • Live Roles (viewer, engineer, admin), resource-level access and an audit log.
    • Live Performance alerts to Slack, Microsoft Teams and email, with thresholds.

    • Live First end-to-end run on an NVIDIA H100: real TensorRT and ONNX Runtime runtimes, and a RunPod GPU runner setup.

    • Live Claude via OpenRouter for rewrites, with a daily rewrite allowance per workspace.

    • Live Kernalin web app opens: email sign-in, workspaces with private data, dashboards over run history.
    • Live Upload CUDA, Triton, PyTorch or ONNX workloads through a static security scan; jobs run in an isolated sandbox.
    • Live Benchmarks across PyTorch eager, torch.compile, TensorRT and ONNX Runtime, with Bayesian verdicts and a ranking.
    • Live Profiling sessions, kernel heatmaps and memory bottleneck analysis.
    • Live Optimization suggestions, plus AI kernel rewrites that are kept only when verified correct and faster.
    • Live PyTorch → Triton conversion.
    • Live CI integration: API tokens, the GitHub Actions benchmark gate and a CI runs page; regression checks after every run.

Dates are commit or merge dates. Before 10 October 2026 the release branch was itself what ran at kerneln.com/app, so earlier dates approximate when items went live.

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.

Get started Talk to us