elegantly

/

purem_sandbox

Like 1

Hi everyone,

🛠 Context and Motivation

Today, local CPU performance for AI models is often treated as secondary compared to GPU acceleration.
We believe this needs to change.
With the right architecture, CPUs can match – and sometimes even outperform – traditional GPU solutions, especially in the context of edge computing, early prototyping, and resource-efficient inference.

Purem is designed for:

• ML researchers
• AI startups
• Edge AI applications
• Cost-sensitive deployments
• Local LLM fine-tuning and validation

Benchmarks – https://github.com/worktif/purem_sandbox/actions/runs/14724603653/job/41324557920

📊 Benchmark Highlights

Here’s a quick look at how Purem compares to existing Python ML libraries:

Acceleration (Purem vs NumPy, PyTorch, Numba)

Operations per second on large inputs (log scale)

Mean numerical error across large input sizes – lower values indicate higher numerical precision under load.

We are opening the free version permanently, enabling unrestricted benchmarking and real-world local testing by a non-commercial use.

Purem's softmax kernel reaches ~6500 ops/sec on Apple M2 CPUs, based on real large-scale benchmarks.
For comparison, industry solutions recalculated for CPU:
• OpenAI FlashAttention achieves ~800–5000 ops/sec on CPU.
• Meta's Xformers (PyTorch 2.0) achieves ~1300–1500 ops/sec on CPU.

Purem brings industrial-grade AI performance without requiring GPUs, specialized cloud services, or complex hardware setups – running natively on consumer CPUs optimizations.

📊 Benchmark Highlights

• Fully local execution
• No cloud dependencies
• O(N) linear-time performance
• Transparent memory management
• Deterministic speed without drift

Purem enables developers and researchers to test, validate, and deploy AI workflows locally – with real production-grade speed.

Would love to hear your feedback, insights, and suggestions! 🚀

(If you’re interested, we also included detailed benchmark data, architectural notes, and upcoming development plans at https://worktif.com).

🛡️ TL;DR

• Purem: Supercomputer-grade AI kernel running locally.
• Softmax on CPU reaching ~6500 ops/sec.
• Free demo goes live April 28.
• Ready to unlock AI potential for everyone.