Supported Workloads

Pyvorin delivers the best results on data-heavy, CPU-bound Python. Speedups are workload-specific and depend on input size, type stability, and compilation tier.

ETL & Data Processing

Workload	Typical Speedup*	Notes
Filter / map pipelines	10–1,000×	List comprehensions and simple lambdas compile well.
Group-by aggregations	20–500×	Hash-based grouping on primitive types is highly optimised.
Rolling window calculations	15–300×	Sliding operations over numeric arrays benefit from loop fusion.
CSV transforms	10–50–	Parsing and columnar extraction on large files.
JSON transforms	5–40×	Key extraction and nested filtering.
Log parsing	10–100×	Regex-lite string splitting and tokenisation.

String Processing

Workload	Typical Speedup*	Notes
Tokenisation	100–500×	Whitespace and delimiter splitting in tight loops.
Pattern matching	50–300×	Fixed-string search and simple character-class scans.

Finance Kernels

Workload	Typical Speedup*	Notes
Black-Scholes	50–2,000×	Closed-form option pricing on large portfolios.
Value at Risk (VaR)	20–200×	Historical simulation and percentile extraction.
Monte Carlo simulation	10–100×	Path generation depends on random-number backend.
Present Value (PV)	30–500×	Cash-flow discounting loops.

What Makes a Good Workload

Type-stable loops: Variables keep the same type (e.g., int or float) across iterations.
Primitive collections: Lists and dicts of scalars compile better than deeply nested heterogeneous structures.
Pure computation: Heavy arithmetic, reductions, and transformations with minimal I/O inside hot paths.
Large inputs: Compilation overhead is amortised over millions of operations, not dozens.

What to Avoid

Async / await and coroutine-heavy code
Dynamic code generation (eval, exec)
GPU training or CUDA-dependent libraries
Heavy NumPy / Pandas internals (ndarray methods are not yet fully lowered)
Network-bound services where Python overhead is negligible

See Unsupported & Fallback for details on how Pyvorin handles these constructs.

* Speedup ranges are indicative only, derived from internal benchmarking on x86_64 Linux. Your results will vary based on hardware, input size, and compiler version.