Maxwell：一种介于代码与硬件之间的新模型
Maxwell: A new kind of model that sits between code and hardware
doubleAI 发布 Maxwell，一种可在运行时分析代码与硬件、改写二进制并以形式化证明保证与原代码等价的新型优化模型。
原文给出Whisper与cuGraph等场景的实测加速比，并强调优化与原代码形式化等价，为评估这类动态kernel改写工具提供具体参照。
by doubleAI
Maxwell understands and improves code while it runs, reasoning about hardware and data, not just code. Every improvement is formally proven equivalent to the original code. Read the manifesto
- 38×faster data processing pipelines
- 4.25×faster AI inference kernels
What is a code execution model?
A short animated film, with narration and captions, introducing Maxwell: most code never sees the hardware it runs on or the data it processes. Maxwell watches a program while it runs, rewrites the binary underneath to fit that exact machine and data, and formally proves every improvement equivalent to the original. It runs on CPUs and GPUs, speeding up data pipelines and AI inference kernels.
Featured customer
LTX-2.5 generates 4K video in half the time
LTX built LTX-2.5 for efficiency from day one, compiled end to end, sequence-parallel across GPUs, decoding on a hand-written attention kernel. A pipeline built and optimized by top human experts. Maxwell made their 4K generation nearly twice as fast.
- 1.87×faster 4K video generation
- 134sper 4K clip on two B200s, down from 250s
Maxwell’s Demonstration
See Maxwell in action optimizing both GPU-based and CPU-based code. Once Maxwell has enough runtime data, it swaps in formally proven optimizations tailored to that exact machine while the program keeps running.
Two terminals run the same Python script on the same RTX 4090: 180 transcriptions with Whisper-small from a cold start. Without Maxwell, on the left, it takes 106.23 seconds. With Maxwell, on the right, it completes in 49.82 seconds. Both produce identical transcripts. The recording plays at 4× speed.
Two terminals run the same Python script on the same machine: NetworkX betweenness centrality, 15 runs. Without Maxwell, on the left, it takes 61.44 seconds. With Maxwell, on the right, it completes in 17.50 seconds. Both print identical results. The recording plays at 4× speed.
Research behind Maxwell
-
Maxwell approaches Speed of Light on Blackwell
After one day of search, Maxwell beats NVIDIA’s PyTorch baselines on 90% of SOL-ExecBench’s Blackwell kernels.
(opens on doubleai.com) -
doubleAI’s Maxwell: Surpassing Expert-Written Kernels At Scale
Independently generated and verified a hyper-optimized drop-in version of NVIDIA cuGraph, 3.6× faster on average.
(opens on doubleai.com) -
Blazing Fast pandas Without Changing a Line
Maxwell completely rewrote pandas’ kernels from scratch, specialized for modern hardware.
(opens on doubleai.com)
来源：Hacker News · heymaxwell.ai