Meta unveils its next-gen MTIA custom AI chip

While everyone else is fighting for Nvidia GPUs, Meta is busy rolling out its own custom silicon.

Meta just pulled back the curtain on its next-generation MTIA chip, an in-house inference accelerator built specifically to power the recommendation algorithms behind Facebook and Instagram. The new processor is already live in production across 16 data center regions.

Tailored for feeds, not just chatbots

Most AI hype centers on massive language models, but Meta's actual bread and butter is the ranking engine serving ads and Reels to billions of users every second.

These recommendation workloads need massive memory bandwidth and ultra-low latency rather than raw brute-force compute. Meta upgraded from TSMC's 7nm process to 5nm, pushing clock speeds from 800 MHz to 1.35 GHz and lifting power consumption from 25W to 90W.

The specs reflect those exact bottlenecks. Dense compute jumped 3.5x and sparse compute jumped 7x compared to v1. Meta also doubled on-chip SRAM to 256MB with a massive 2.7 TB/s bandwidth, while boosting off-chip LPDDR5 memory to 128GB.

To house them, Meta built custom rack systems packing up to 72 accelerators connected via PCIe Gen5, with optional RDMA networking to scale even further.

The PyTorch superpower

Silicon is only as good as the software driving it, and being the creators of PyTorch gives Meta an unfair advantage here.

The team built a dedicated Triton-MTIA compiler backend that plugs directly into PyTorch 2.0. Because the system is backward-compatible with code written for MTIA v1, engineers were able to migrate production traffic in just a few days.

In fact, it took less than nine months to go from first silicon to live deployment across 16 data center regions. Across four internal evaluation models, Meta reported a 3x jump in chip performance and a 6x throughput boost at the system rack level.

Why it matters

Don't expect to buy an MTIA card off the shelf. But if you use Meta's apps, this silicon is quietly driving the ranking engines that keep you scrolling, helping Meta deliver sharper targeting with lower latency.

For the broader industry, this is a masterclass in workload optimization. Instead of burning thousands of scarce, power-hungry Nvidia GPUs on ad ranking, Meta offloads those daily repetitive tasks to its hyper-efficient custom chips.

We think this pragmatic approach gives Meta breathing room. By slashing the cost of powering its core ad engine, Meta can redirect precious Nvidia hardware and datacenter power toward training next-gen frontier models. Custom ASICs for internal inference alongside commercial GPUs for frontier training is quickly becoming the ultimate hyperscaler playbook.