Circumventing browser worker memory ceilings via an optional lightweight workstation daemon. Retain the clean browser-native notebook interface while scaling to 128 GB+ workstation RAM, multi-threaded Rayon parallelism, and direct memory-mapped file access.
While ForgeCurves executes numerical routines with compiled WebAssembly efficiency, standard in-browser Web Workers operate under unavoidable host restrictions:
Standard wasm32 addresses memory via 32-bit unsigned integers, imposing an absolute theoretical ceiling of 4 GB ($2^{32}$ bytes) per linear memory instance.
Modern browser engines (V8, Blink, Gecko) allocate worker memory from per-tab quotas. Chrome, Edge, and Firefox typically cap a single worker or WASM instance between 1.4 GB and 2.0 GB of RAM.
Allocating millions of temporary JavaScript objects before transferring them to linear memory causes severe garbage collection pauses and occasional out-of-memory tab crashes.
A dense $50,000 \times 2,000$ matrix of 64-bit doubles requires 800 MB. Factorizations (SVD, QR) or cross-set distances require transient allocations exceeding 2 GB, crashing browser workers.
Because WebAssembly bytecode is portable across host environments, ForgeCurves solves the memory ceiling through an optional downloadable background daemon (ForgeHyperCurves / forgeHypercurves):
Users remain entirely within the familiar ForgeCurves browser notebook interface at forgecurves.com. No desktop app installation or window switching required.
Cell execution requests transparently route over a fast local loopback WebSocket connection (ws://127.0.0.1:8765) directly to the companion runner.
The companion daemon accesses the workstation's physical RAM (16 GB, 32 GB, 64 GB, 128 GB+) and high-speed NVMe swap, bypassing browser memory quotas completely.
If the daemon is offline, ForgeCurves automatically falls back to the in-browser Web Worker engine, ensuring 100% functionality without mandatory software installation.
ForgeHyperCurves provides two interchangeable execution engines sharing the identical communication protocol:
Embeds a standalone WebAssembly runtime (Bytecode Alliance Wasmtime) executing the exact identical octave_wasm_math.wasm binary used in the browser. Guarantees 100% bitwise parity while expanding memory to physical RAM with 64-bit indexing (memory64).
Pre-compiled standalone executable compiled from clean Rust without browser wrappers. Unlocks 64-bit native pointers, hardware vectorization (AVX-512 / ARM Neon), Rayon multi-threaded CPU parallelization, and mmap memory-mapped file access for 200 GB+ datasets.
Comparing the standard in-browser Web Worker against the two ForgeHyperCurves companion modes:
| Architecture Dimension | In-Browser Web Worker (Default) | Companion Mode 1 (Wasmtime WASM) | Companion Mode 2 (Native Multi-Threaded) |
|---|---|---|---|
| Memory Ceiling | ∼1.5 – 2.0 GB (Browser Tab Cap) | Up to Full System RAM (64 GB+) | Full RAM + NVMe Swap (128 GB+) |
| Execution Engine | V8 / SpiderMonkey Worker | Standalone Wasmtime JIT | Native x86_64 / ARM64 Machine Code |
| Hardware Vectorization | WASM SIMD128 (128-bit) | WASM SIMD128 / memory64 | Native AVX2, AVX-512, ARM Neon |
| Multi-Threading | Single Worker Thread | Multi-threaded WASM instances | Rayon Thread Pool (All CPU Cores) |
| Local Filesystem | Virtual OPFS / IndexedDB | Local FS over JSON-RPC | Direct mmap (Datasets > RAM) |
| User Installation | Zero Install (Pure Web) | Download 5 MB Executable | Download 5 MB Executable |
| Bitwise Parity | Reference standard | 100% Bitwise Identical | IEEE 754 Floating-Point Parity |
The companion daemon communicates using standard JSON-RPC 2.0 messages matching the existing OctaveKernelAdapter interface: