doubleAI introduces Maxwell, a code execution model that optimizes running code

Maxwell optimizes code while it runs

doubleAI, a Tel Aviv, Israel-based startup founded in 2023, introduced Maxwell on Monday. The code execution model evaluates, optimizes and rewrites binary code during execution. It is named for Maxwell's Demon, a thought experiment on entropy imagined by James Clerk Maxwell in 1867, and it preprocesses algorithms and machine instructions to maximize performance on available hardware.

Maxwell is similar in concept to a JIT compiler, but it is more flexible and formally verifies the equivalence of replacement code. Source code is not touched. Co-founder Gal Beniamini, who holds a PhD in theoretical computer science, said a compiler cannot do high-level reasoning such as changing an algorithm. Maxwell can see the exact hardware, the shapes and types of data flowing through a program, the environment state, what else is running and which bottlenecks appear under which conditions.

Reported speedups and security checks

doubleAI said Maxwell has made data pipelines more than 38x faster and AI inference kernels more than 4.5x faster. LTX-2.5, an open world model that generates 4K video, managed to do so almost twice as fast with Maxwell. The optimizations reduced data exchanged between GPUs and improved kernel selection to drop unnecessary token padding.

Beniamini said Maxwell does best on algorithmic code that can be tailored to hardware in a non-trivial manner. He added that Python code and PyTorch graphs can still get far-reaching optimizations because they are often far from optimal. Maxwell can also address security by proving that memory errors do not exist. He said the company found cases in Python itself where Python was not memory safe and Maxwell proved that to be the case.

Deployment and trust

Maxwell's tracing and profiling have a maximal overhead of 2 percent, and zero when it is not tracing. Beniamini said the optimizations outweigh that cost, and doubleAI also applies Maxwell to the runtime itself to get back a much larger than 2 percent flat speedup across the board. Developers install the runtime and run their code, with the Maxwell model hosted by doubleAI, an arrangement that requires some trust.

Amnon Shashua, co-founder and CEO of doubleAI, said AI has changed how software is written and is now changing how it runs. He described code execution models as agents on behalf of hardware rather than code agents on behalf of the human user, calling Maxwell the first AI that works for the machine, not the programmer.

Sources: The Register · doubleAI