🚀 Performance increased at least 2× across all benchmarks
We've revisited and optimized every core performance-critical path in Purem. Through a series of low-level enhancements and architecture-specific optimizations, the system now demonstrates at least double the speed compared to previous versions – with real-world benchmarks confirming these improvements on all supported platforms.
The New Standard
Optimized for x86-64, Intel, and ARM (Apple Silicon M1–M4) servers.
About
This repository contains real-world benchmarks for the Purem core technology, designed for maximum performance testing and local validation.
Purem now offers a fully licensed and notarized sandbox binary for Apple Silicon M1 – M4 machines.
Unlimited
Unrestricted
No signup
No payment
No cloud dependencies
You can freely test and benchmark Purem locally – the same architecture, the same performance, ready to run directly on your Mac.
Key Highlights
⚡ Full support for x86-64 and ARM64 (Apple M1–M4) architectures.
⚡ Production-grade benchmark scripts included.
⚡ Optimized Softmax computation using NumPy + Numba (maximum CPU utilization).
⚡ Parallelized and vectorized execution for massive data sizes (millions to billions of elements).
⚡ Accurate cost estimation for cloud and serverless environments (AWS Lambda, Azure Functions, etc.).
⚡ Focus on real hardware performance, not synthetic scores.
Getting Started
Requirements
Python 3.11+
purem
numpy
numba
torch
pandas
scipy
pytest
matplotlib
MacBook or Mac server with Apple Silicon (M1–M4) – or any x86-64 server with supported Purem license.
About the Makefile
🧠 This project includes a Makefile for internal tasks like testing.
👉 You can run it if you like – it won't affect the project or cause any issues.
However, to perform full tests and benchmarks properly, please follow the official usage instructions. Just use the project as normal.
Generates full performance plots under src/purem_benchmarks/ showing a runtime vs input size.
📊 Benchmark Highlights
Here’s a quick look at how Purem compares to existing Python ML libraries:
Acceleration
Acceleration (Purem vs NumPy, PyTorch, Numba)
OPS
Operations per second on large inputs (log scale)
Stability
Mean numerical error across large input sizes – lower values indicate higher numerical precision under load.
Licensing
Apple Silicon (M1–M4):
The sandbox binary is permanently free to use – no registration required.
You can fully develop, benchmark, and integrate locally.
x86-64 Servers (Intel/AMD):
A separate licensed version is available.
Please refer to the Purem by Work TIF website for more details.
Important Dates
🗓️ Sandbox Version Public Release: April 28th, 2025
Why Benchmark with Purem?
By using the free sandbox, you can:
Validate performance claims locally.
Benchmark your own models with production-grade accuracy.
Optimize cloud costs based on real runtime data.
Prepare real-world production integrations with no vendor lock-in.
