hpc-labs.ai

An AI-enabled optimizer for whole code bases. It explores combinations of optimizations (compiler and runtime settings, loop and vector work, library swaps, data layout, algorithmic changes, Python-to-C ports), measures every candidate on real hardware, and keeps a change only when it is verified.

Correctness is the licence to operate: each change is checked against your own program's outputs, byte for byte, or against a tolerance you declare (numeric error, PSNR / SSIM, a task metric). We call that rule the contract: a change that does not meet it is rejected, whatever its speed.

• Bit-accurate by default. The refactored program writes outputs identical to the original unless you allow a stated numeric contract: a tolerance you declare, such as a numeric error, PSNR / SSIM or a task metric.
• Measured, not estimated. Original and candidate run alternately, with a significance test; a gain that is noise is reported as no gain.
• Your code stays on your machine. The command line tool runs locally; nothing is uploaded unless you turn a cloud model on and confirm.
• Python, Java, C and C++ today, for x86 and Arm (Linux, and Windows on x64 and Arm64).

In our own example suite (20 small projects, one run each on an Arm DGX Spark, release 0.3.7; later releases carry the same engine, compiled), 18 ran faster with the output verified, from about 1.1x to 50x. The numbers per example and per release are published with each release.

Covers: Python (code that Python 3.12 can parse), C and C++ (built with GCC; tested with GCC 13) and Java (JDK 17). Libraries: NumPy and OpenCV (tested with OpenCV 4.13 and 5.0.0). Systems: Windows 11 on x64 (installer and zip; Arm64 packages are built but not yet run on Arm hardware) and Linux or DGX on Arm64 (tested) and x86_64 (built, not yet run on x86 hardware). macOS on Apple silicon: tried by hand on one Mac (the Python and C++ samples passed); there is no installer or package yet.

• Covers: everything in ParetoStack-soft, plus AWS Lambda functions in Python (Python 3.12 tested; x86_64 and arm64), Google Cloud Functions (2nd generation, HTTP trigger, Python 3.12 tested) and a read-only probe of Linux cloud hosts (EC2, Azure, Google). It needs the aws or gcloud command line tool, signed in to a test account or project.
• Check and probe. cloud check tells whether the aws, az and gcloud tools are installed and signed in; cloud probe inspects a cloud host (processor, instance type, how Python, NumPy and OpenCV were built) and lists likely gains, labelled as not measured.
• AWS Lambda. Finds the cheapest memory and processor setting that gives identical answers, optimizes the function's own Python and, as an option, ports a numeric hot loop to C. Every step is measured on throwaway copies in a test account, with a cost ceiling, and the copies are deleted afterwards; the real function is never changed.
• Offline. lambda-optimize --zip and gcf-optimize --zip optimize a local deployment package and verify it on recorded events on your machine, byte for byte; nothing is sent anywhere.
• Google Cloud Functions. Finds the cheapest memory and CPU setting of a deployed function (gcf-tune).
• Coming soon: Azure Functions, and live verification and the C port for Google Cloud Functions.
• Planned, not built yet: database access optimization. The N+1 round-trip problem that monitoring tools only detect: a query inside a loop becomes one batched query, row-by-row work on the client becomes one server-side aggregation, a connection per call becomes a pool, and SELECT * becomes a column list with pagination. First targets: PostgreSQL and MySQL (psycopg, SQLAlchemy), DynamoDB (boto3), Redis and MongoDB. Each change is checked row for row against a local test database, so it needs no cloud account.

• Covers (the alpha): Verilog, SystemVerilog and VHDL designs through the open-source Yosys, nextpnr and GHDL, on models of Lattice iCE40 and ECP5 FPGAs, an ASIC area and delay estimate on the open SkyWater cells, and the re-routing of small KiCad circuit boards. You install the open-source tools yourself; the results are tool models, not signoff and not a replacement for vendor tools. PCB and chip-package design as a separate, higher-priced line comes later.
• The same loop. A change to a hardware design is kept only when it is verified, as in software.
• Equivalence contracts. The contract is equivalence with the original design, in place of identical program output.
• Later product lines. ParetoStack-cad (CAD software) and ParetoStack-bio are planned names; nothing is built.
• Measured. Frequency, area and power replace run time as what is measured.
• Status. Alpha, 0.5.0a2: free, for evaluation, no warranty. Tried on Linux (also under Windows WSL2) and on an Apple-silicon Mac; not yet on native Windows.