Meta’s Custom AI Silicon Strategy Takes Shape
Meta has begun rolling out its first-generation AI inference accelerator, MTIA v1, alongside its publicly available Llama 2 large language model. Behind these launches is a broader push toward custom silicon, with the company’s in-house chip design team working to optimize compute efficiency for its infrastructure. The effort is led by Olivia Wu, a silicon industry veteran with three decades of experience, who is responsible for taking Meta’s ML accelerator from design through deployment.
Why Build Chips In-House?
Wu’s team handles the design and development of Meta’s machine learning accelerator, coordinating closely with architects, verification engineers, and the software teams who will ultimately run workloads on the hardware. The decision to bring silicon design in-house is rooted in a basic mismatch: AI is evolving faster than the traditional silicon development cycle.
A typical chip development cycle spans 18 months to two years, sometimes stretching to four. That means Wu and her team are designing hardware for software that hasn’t been written yet. The silicon has to handle today’s AI workloads and a range of yet-unseen future demands. The workaround, Wu says, is tight coupling between the hardware and software teams. Because both sit inside Meta, chip architects have visibility into the specific AI workload trends and feature requirements that will land on their hardware. That information is baked into the silicon from the start, rather than retrofitted after the fact.
Wu joined Meta in 2018 after seeing a post from Chief AI Scientist Yann LeCun about building AI silicon in-house. At the time, few companies were designing their own AI accelerators, and those that did were mostly focused on the chip itself, not the surrounding software ecosystem. Meta’s angle was different: bringing silicon engineers in to work side-by-side with software developers, rethinking the full system stack to gain efficiency by optimizing across the hardware-software boundary.
Coming from a startup background, Wu describes the transition as surprisingly smooth. The small early team gave her hands-on exposure across all aspects of ASIC development, and Meta’s culture of whiteboard sessions with hardware, software, and cross-functional teams provided insight into the company’s critical workloads that was hard to get elsewhere.
The MTIA v1 Milestone
MTIA v1 is Meta’s first-generation ML accelerator, fitted for its deep learning recommendation models — a core component across Facebook, Instagram, WhatsApp, Meta Quest, Horizon Worlds, and Ray-Ban Stories. Meta isn’t stopping purchasing silicon from external partners, but designing its own chips allows it to tailor hardware to its specific workloads and take control of the entire stack: silicon, system, software, and application.
Building that first chip meant constructing Meta’s entire design and verification flow from scratch, plus the silicon development infrastructure itself. It was a heavy lift early on, but Wu says it has paid off long-term. The team still operates with a startup-like ethos — wearing many hats and jumping into unfamiliar roles as needed.
What Comes After the First Chip
The immediate challenge is moving beyond recommendation systems. Meta’s AI ambitions now include GenAI and video processing, each with different hardware requirements. That diversity opens up opportunities to build tailored products rather than a one-size-fits-all accelerator.
The MTIA development process has generated learnings that Wu says will feed directly into future products. Meanwhile, the team is preserving the developer experience that PyTorch eager-mode development offers by continuing to support PyTorch 2.0, which pushes optimization down to the compiler level. Feedback from Meta’s AI software teams is shaping the feature set of future AI silicon.
For next-generation MTIA chips, Wu points to system bottlenecks — memory and inter-chip communication in particular — as key areas of focus, so that scaling doesn’t hit unanticipated walls.
A Culture Built for Engineers
Wu’s advice to engineers from underrepresented groups is straightforward: participate actively, speak up so people know what you can do, and seek mentors — they don’t have to share your background. Meta’s infrastructure team runs mentoring programs for women and underrepresented engineers, offering both group coaching and one-on-one sessions.
For engineers, researchers, and developers considering Meta now, Wu highlights the learning environment. Meta provides resources for people to get familiar with the entire stack — silicon, firmware, compiler, application, and the data-center-scale system design — even if they have no prior exposure to parts of it. The massive deployment scale adds a dimension of engineering challenge that makes the work both difficult and worth doing.
