News
IndustryTSMC’s grip on advanced chipmaking will face competitionIndustryCadence joins European Chip Design Platform to support semiconductor startups and SMEsMaking chipsMarc Hijink on ASML’s Culture, Supply Chain, and GeopoliticsProcessorsIntel's post-Nova Lake CPUs take shape as Razor Lake desktop and laptop variants surfaceAI chipsHow Google’s New AI Chip Software Helps It Battle NvidiaIndiaMore IC Packaging Plants Emerge In IndiaIndustryUS Tightens China AI Chip Access as Nvidia, TSMC and Semiconductor Stocks SlideDesignAxelera AI moves into HPC with TitaniaMaking chipsChinese envoy says chip equipment export curbs threaten EU-China trade thawIndiaSEMICON India 2026 Day 2 Sees 25 MoUs, Strategic Collaborations Across Semiconductor Value ChainIndustryTSMC’s grip on advanced chipmaking will face competitionIndustryCadence joins European Chip Design Platform to support semiconductor startups and SMEsMaking chipsMarc Hijink on ASML’s Culture, Supply Chain, and GeopoliticsProcessorsIntel's post-Nova Lake CPUs take shape as Razor Lake desktop and laptop variants surfaceAI chipsHow Google’s New AI Chip Software Helps It Battle NvidiaIndiaMore IC Packaging Plants Emerge In IndiaIndustryUS Tightens China AI Chip Access as Nvidia, TSMC and Semiconductor Stocks SlideDesignAxelera AI moves into HPC with TitaniaMaking chipsChinese envoy says chip equipment export curbs threaten EU-China trade thawIndiaSEMICON India 2026 Day 2 Sees 25 MoUs, Strategic Collaborations Across Semiconductor Value Chain
EveryChip
MTIA v1

MTIA v1

Meta’s first custom AI inference chip for recommendations

Designed by Meta in 2020, the MTIA v1 is a custom accelerator for deep learning recommendation models. Built on TSMC 7nm technology, it targets efficient inference workloads within PyTorch environments, offering an alternative to general-purpose GPUs for specific social media ranking tasks.

Share on WhatsApp

  • Designed by Meta in 2020 for AI inference
  • Built on TSMC 7nm process with 25 W TDP
  • Features 64 RISC-V based processing elements
  • Supports up to 128 GB LPDDR5 memory
  • Optimised for deep learning recommendation models

What it is

The MTIA v1 is a specialised application-specific integrated circuit (ASIC) created by Meta to handle artificial intelligence inference tasks. Specifically, it targets deep learning recommendation models used across Meta’s services, such as content feeds and advertising ranking. Announced as part of a 2020 design effort, this chip represents Meta’s move away from relying solely on graphics processing units for these specific workloads. By co-designing the hardware with its software stack, the company aims to improve efficiency and performance per watt for its massive scale operations.

How it came about

Meta identified that standard graphics processors were not always optimal for their unique recommendation workloads at scale. In response, the team designed the first-generation MTIA ASIC in 2020. The goal was to create a solution tailored to next-generation recommendation model requirements. This initiative involved integrating the accelerator directly into the PyTorch framework to maintain developer efficiency while optimising the underlying silicon for specific computational patterns found in social media ranking algorithms.

Inside the chip

The chip features a grid of 64 processing elements arranged in an 8x8 configuration. Each element contains two processor cores based on the RISC-V instruction set, customised for control and compute tasks. These cores are supported by fixed-function units for matrix multiplication and other critical operations. A mesh network connects the elements to each other and to memory resources. The system includes 128 MB of shared on-chip SRAM for low-latency access and supports up to 128 GB of off-chip LPDDR5 memory. A dedicated control subsystem manages firmware and job orchestration.

How it was made

TSMC manufactured the MTIA v1 using its 7nm process node. The silicon operates at a clock speed of 800 MHz. In terms of performance, it delivers 102.4 TOPS at INT8 precision and 51.2 TFLOPS at FP16 precision. The thermal design power is rated at 25 W. The chips are mounted on small dual M.2 boards, which facilitates easier aggregation into servers. These boards connect to host CPUs via PCIe Gen4 x8 links, with the total board consumption reaching up to 35 W.

Where it went

These accelerators are deployed in data centres running Meta’s internal workloads. They are housed in servers following the Open Compute Project Yosemite V3 specification. Each server can hold 12 accelerators, connected through a hierarchy of PCIe switches to allow parallel workload distribution without involving the host CPU for every communication. The primary application is running deep learning recommendation models for content understanding, feeds, and ads ranking, where efficiency and memory bandwidth are critical.

Versions and family

The MTIA v1 is the first generation in the Meta Training and Inference Accelerator family. The current focus remains on optimising the software stack for medium and high-complexity models.

Why it matters

This chip marks Meta’s entry into custom silicon for AI inference, aiming to reduce total cost of ownership through better performance per watt. Evaluations showed it handled low and medium-complexity recommendation models more efficiently than comparable neural network processors and GPUs. However, the company acknowledged that optimisation for high-complexity models was still ongoing. The project demonstrates a shift towards co-designed hardware and software solutions to address the specific scaling challenges of memory and interconnect bandwidth in large-scale AI deployment.

Questions people ask about MTIA v1

What is the primary purpose of the MTIA v1?

The MTIA v1 is designed to accelerate deep learning recommendation models. It handles tasks like content ranking and ad targeting for Meta’s services, offering a more efficient alternative to general-purpose GPUs for these specific workloads.

Which instruction set architecture does the chip use?

The processing elements within the MTIA v1 use cores based on the RISC-V open instruction set architecture. These cores are heavily customised to perform the necessary compute and control tasks required by Meta’s AI models.

How much memory can the MTIA v1 access?

The chip includes 128 MB of on-chip SRAM for fast data access. It also supports off-chip memory resources using LPDDR5, which can scale up to 128 GB, allowing it to handle large embedding tables and data sets.

What is the power consumption of the accelerator?

The silicon itself has a thermal design power of 25 W. When mounted on the dual M.2 boards used for server integration, the total power consumption is as little as 35 W, making it energy-efficient for data centre deployment.

Good reading and sources

    Where this page comes from

    There is no English Wikipedia article on MTIA v1 yet, so the facts on this page come from the sources listed above: ai.meta.com/blog/meta-training-inference-accelerator-AI-MTIA/, nextplatform.com/compute/2026/04/08/contemplating-metas-homegrown-mtia. Some are the subject's own pages, so its claims about itself are reported as its claims. EveryChip wrote the text in its own words, with AI assistance, and compares it with the sources to make sure no passage is copied.

    It may contain mistakes, and figures change over time. Nothing here is investment or buying advice. If something is wrong, tell us. Last refreshed 2026-11-14.

    The weekly digest

    Every Monday: the chip of the day, new pages on EveryChip, and the conferences and anniversaries of the week ahead.