PassiveLogic’s Compiler Sets New Bar for World’s Fastest AI Training
PassiveLogic has announced it has set a new record in AI training, says the company. Its compiler runs hundreds of times faster than Google’s TensorFlow and Meta’s PyTorch, the company claims.
In benchmark comparisons training a heterogeneous neural net, PassiveLogic’s differentiable Swift compiler outperformed leading AI frameworks, reporting speeds 322x faster than TensorFlow and 238x faster than PyTorch for equivalent models that run on the same host processor.
Simply put, a compiler takes source code written by humans, for which there are many programming languages, and transforms it into a form that can be run directly by computers. "In our case, we also have the compiler generate mathematical derivatives of our source code, a critical piece of how we optimize our digital twins and control paths," said Brad Larson, one of the members of the Compiler Team at PassiveLogic.
Other Members or the team, which attended the LLVM Developers’ Meeting in Santa Clara, CA on October 10-12, as shown above, are (left to right): Kshitij Jain, Viranchee Lotia, Manasij Mukherjee, and Brad Larson. LLVM is a collection of modular and reusable compiler and toolchain technologies.
Collaborating with Apple, PassiveLogic selected the Swift language to advance differentiable computing, focusing on AI performance in edge and industrial applications. This speed milestone highlights the work of PassiveLogic’s Swift compiler team—the largest of its kind outside of Apple and the largest in the world devoted to differentiable programming. This team is focused on advancing differentiable programming and enabling generative autonomous systems.
In contrast to the AI models that exist today written in frameworks like TensorFlow or PyTorch, PassiveLogic's framework in Swift leverages language-integrated differentiable computing that not only merges systems programming and AI, but enables a new generation of heterogeneous networks and typed interfaces that can be clicked together. This expands the limits of AI far beyond chatbots, democratizing AI and enabling full autonomy for systems.
“People are surprised when they hear a building controls company has an entire compiler team,” said CEO Troy Harvey. “But given the complexity of buildings, you need incredibly advanced AI that can handle the computational load to find a building’s most efficient control path. Our work in differentiable Swift enables the reversible computing that powers our AI to make real time control decisions at the edge.”
While PassiveLogic is leading differentiable computing in the industry, the company is committed to open source, collaborating with Apple and the community to bring these innovations to the rest of the market. PassiveLogic is focused on expanding Swift to the industrial market where its modern features can replace legacy languages like C++. As a result of its efforts, the team has isolated, identified, and reported 33 issues to the Swift language project and provided 27 patches themselves in 2023 alone. Some of these issues identified were longstanding language bugs.
“Differentiable Swift changes the game for edge-based AI and how we build applications beyond conventional deep learning,” said Harvey.
Since 2021, TechBuzz has reported on PassiveLogic's innovative technology, its fundraising activities (a $34 million Series B, a $15 million venture round from NVIDIA's venture arm Nventures), and vision from its CEO, Troy Harvey, in an introductory profile article marking the company's cumulative $16 million in funding that it had raised thus far.