1024 Project Hardware Application for Magnetron

[MarioSieg]

1024 Project Hardware Application

主要用途 (Main Purpose)

上游集成 / 社区适配

Porting and optimizing a high-performance machine learning framework (Magnetron) to LoongArch, including SIMD kernel implementation and CPU inference optimization.

---

预期效果 (Expected Outcome)

• Port Magnetron (C-based ML framework) to LoongArch
• Implement LASX SIMD backend for tensor operations
• Run transformer inference models (e.g. Qwen family) on Loongson CPUs
• Benchmark performance vs x86 (AVX512) and ARM (NEON)
• Identify performance bottlenecks and contribute optimizations
• Publish results and share findings with the LoongArch community

---

申请设备 (Requested Hardware)

Preferred:

• 主机板卡-3A6000双网卡

Alternative:

• 整机-3A6000NUC

---

身份信息 (Identity)

Mario Sieg

• Author & Maintainer of Magnetron ML Framework
• Software Engineer (CPU/CUDA Kernel Development) at Prime Intellect
• Student at TU Berlin (Computer Science, Mathematics)

---

联系方式 (Contact)

• Email: mario.sieg.64@gmail.com
• GitHub: https://github.com/MarioSieg
• X: https://x.com/mario_neo

---

备注 (Additional Information)

Magnetron is a lightweight, high-performance machine learning framework written in C with a Python interface. It features a custom tensor engine, SIMD vectorization (AVX2 / AVX-512 / NEON), and a strong focus on performance and minimal dependencies.

The framework already supports transformer inference (e.g. Qwen models) on CPU. Current development focuses on improving performance scaling, kernel fusion, and architecture-specific optimizations.

By porting Magnetron to LoongArch, I aim to:

• provide real-world ML inference benchmarks on Loongson CPUs
• implement optimized LASX SIMD kernels
• evaluate LoongArch as a platform for modern AI workloads
• improve compatibility and performance of ML software on LoongArch

I am willing to:

• share benchmark results publicly
• contribute fixes and optimizations
• participate in community discussions or technical reports

This work directly supports the LoongArch ecosystem by enabling efficient machine learning workloads on Loongson hardware and improving developer adoption.

[manxing-github]

Hi Mario,

Thank you for your interest. We are currently working on preparing the hardware and will contact you via email once it's ready to ship out.

[MarioSieg]

Hello,

thank you very much for the update - that sounds amazing!
Looking forward to working with the hardware.