Meta MTIA v2
Custom AI inference accelerator for Meta ranking models
Meta designed the MTIA v2 to accelerate its internal AI workloads, specifically ranking and recommendation systems. Released in 2024, this custom silicon doubles the compute and memory bandwidth of its predecessor while optimising for high efficiency at low batch sizes.
- Built on TSMC 5 nm process with 2.35 billion gates
- Features an 8x8 grid of processing elements
- Consumes 90 watts per accelerator at 1.35 GHz
- Delivers 354 TFLOPS/s sparse FP16/BF16 performance
- Uses LPDDR5 memory with 204.8 GB/s bandwidth
What it is
The Meta MTIA v2 is a custom-built artificial intelligence accelerator developed by Meta. It serves as the second generation of the company’s in-house silicon programme, following the initial MTIA v1. The chip targets specific internal workloads, primarily the deep learning models used for ranking content and serving advertisements across Meta products. Unlike general-purpose graphics processors, this design prioritises memory capacity and bandwidth efficiency for inference tasks where batch sizes may remain small. The company states that the architecture balances compute power with substantial static random-access memory (SRAM) to maintain high hardware utilisation. This approach allows the system to handle both simple and complex recommendation models effectively within Meta’s data centres.
How it came about
Meta began its journey into custom silicon with the first-generation MTIA, aiming to improve compute efficiency for its unique software needs. The v2 represents a significant step forward in this long-term venture, addressing the growing sophistication of generative AI and advanced research models. The development team focused on co-designing the hardware alongside the software stack to ensure seamless integration. By building on the foundation laid by the previous version, engineers were able to deploy the new silicon rapidly. The company reports moving from first silicon to running production models in sixteen regions in under nine months, leveraging existing software pathways to accelerate adoption.
Inside the chip
The core of the MTIA v2 features an eight-by-eight grid of processing elements. These units deliver dense and sparse compute performance improvements over the prior generation. To feed this grid, the design triples local storage and doubles on-chip SRAM capacity and bandwidth. An improved network on chip facilitates low-latency coordination between elements. The chip includes a control core and decompression engine on the periphery, surrounding the central compute matrix. It operates at a clock speed of 1.35 GHz, a notable increase from the 800 MHz of its predecessor. The architecture emphasises high SRAM capacity relative to typical graphics processors, enabling efficient operation even when processing limited batches of data concurrently.
How it was made
TSMC manufactures the MTIA v2 using a five-nanometre process node. The die size measures approximately 421 square millimetres and contains around 2.35 billion gates. Each accelerator consumes 90 watts of power, which allows for air-cooled deployment in standard server chassis. The chips are mounted on boards that hold two accelerators each, connecting to the host system via PCIe Gen5 interfaces. This packaging strategy supports high density within rack-based systems, with each chassis holding up to twenty-four accelerators. The design choice of LPDDR5 memory provides a favourable ratio of memory capacity to power consumption compared to high-bandwidth memory alternatives.
Where it went
Meta deploys these accelerators in large-scale rack systems to power its ranking and recommendation engines. The hardware handles both low-complexity and high-complexity models that drive user experiences across the company’s family of apps. A single rack can contain up to seventy-two accelerators, organised in three chassis. The system supports scaling out beyond the rack using remote direct memory access network interface cards if needed. By integrating closely with PyTorch 2.0 and the Triton compiler, the chip allows developers to write efficient kernels for these specific workloads. This setup enables Meta to achieve higher throughput and better performance per watt for its internal AI services.
Versions and family
The MTIA v2 succeeds the original MTIA v1, which was Meta’s first attempt at custom inference silicon. The new version more than doubles the compute and memory bandwidth of the previous solution. While the v1 operated at lower frequencies and power levels, the v2 increases clock speeds and thermal design power to deliver greater performance. The software stack remains compatible between generations, allowing code developed for the first chip to run on the second with minimal changes. This continuity helped the engineering team scale up kernel authoring and expand operator support quickly without rewriting the entire infrastructure.
Why it matters
This chip demonstrates Meta’s commitment to building domain-specific hardware for its massive AI operations. By controlling the full stack from silicon to software, the company claims greater efficiency than what is possible with commercial off-the-shelf graphics processors. The design highlights a different approach to memory hierarchy, favouring capacity and cost-efficiency over raw bandwidth peaks. Early results indicate a threefold performance improvement over the first generation across key models. At the platform level, the new system delivers six times the model serving throughput and a fifty percent gain in performance per watt, supporting the company’s expanding generative AI initiatives.
Questions people ask about Meta MTIA v2
What is the primary purpose of the Meta MTIA v2?
The MTIA v2 is a custom AI accelerator designed by Meta to power its internal ranking and recommendation models. It optimises inference workloads for ads and content feeds, focusing on efficiency for both simple and complex deep learning tasks.
How does the MTIA v2 differ from standard GPUs?
Unlike typical GPUs, the MTIA v2 prioritises high SRAM capacity and memory efficiency over raw bandwidth. It uses LPDDR5 memory instead of HBM, offering a better ratio of memory capacity to power consumption, which suits Meta’s specific low-batch-size inference needs.
What software tools support the MTIA v2?
The chip integrates fully with PyTorch 2.0, utilising TorchDynamo and TorchInductor. Meta also developed a Triton-MTIA compiler backend, allowing developers to write high-performance kernels in the Triton language for both ahead-of-time and just-in-time workflows.
Good reading and sources
There is no English Wikipedia article on Meta MTIA v2 yet, so the facts on this page come from the sources listed above: ai.meta.com/blog/next-generation-meta-training-inference-accelerator-A, servethehome.com/new-meta-ai-accelerator-mtia-2-revealed/, nextplatform.com/ai/2024/04/11/with-mtia-v2-chip-meta-can-do-ai-infere. 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.