Do these 3 things before closing this tab:
1Repair Windows errors before they cause bigger problems2Fix the driver behind crashes, sound loss and screen glitches3Clear out junk files and repair common Windows errorsMeta is using its custom MTIA chips for training as well as inference: the company says MTIA 300 is already in production for training ranking and recommendation models. But the wider roadmap is not a dedicated push to replace GPUs for all AI training. Meta says its next MTIA generations are focused mainly on generative-AI inference, while it continues to buy chips from outside suppliers.
What chip is Meta building?
Meta’s custom accelerator family is called MTIA, short for Meta Training and Inference Accelerator. It is designed for Meta’s own data-center workloads, not sold as a consumer or retail chip. Meta says it developed the family in 2023 and has deployed hundreds of thousands of MTIA chips for inference across organic content and advertising in its apps. Meta’s March 2026 announcement describes MTIA as part of a custom, full-stack system rather than a chip standing alone.
The practical point is that Meta is tailoring processors and systems to the workloads behind its services. That does not mean the chips are general-purpose replacements for GPUs, or that outside companies can buy and install them.
Which MTIA generations are for training?
| Generation | Status reported by Meta | Stated workload |
|---|---|---|
| MTIA 300 | Already in production as of Meta’s March 2026 announcement | Training for ranking and recommendations |
| MTIA 400 | In development in March 2026 | GenAI inference is the primary near-term focus |
| MTIA 450 | In development in March 2026 | Optimized first for GenAI inference; also designed to support other workloads, including GenAI training |
| MTIA 500 | In development in March 2026 | Optimized first for GenAI inference; also designed to support other workloads, including GenAI training |
Meta said it expects to develop and deploy four new generations over the following two years. The company’s announcement does not give a specific production date for each future generation. It presents MTIA 400, 450, and 500 as part of a roadmap whose main near-term use is GenAI inference, not as a series of chips dedicated primarily to training large generative models. Meta’s roadmap announcement
#1 Best Overall
- ✅Powered by 26 Tera-Operations Per Second (TOPS) Hailo-8 AI Processor. 2.5W typical power consumption
- ✅Scalable, enabling simultaneous processing of multi-streams & multi-models
- ✅Enabling real-time, low latency and high-efficiency AI inferencing on the edge devices
- ✅Supports TensorFlow, TensorFlow Lite, ONNX, Keras, Pytorch frameworks
- ✅Supports Linux and Windows. Supports the temperature range of -40°C to 85°C
Training and inference are different jobs
Training is the process of adjusting a model using data so it can learn patterns or improve its behavior. Inference is what happens when a trained model is used to produce a result—for example, ranking content or generating a response. These jobs place different demands on computing systems, so a chip optimized for one is not automatically the best choice for the other.
Meta’s roadmap reflects both needs. MTIA 300’s announced production role includes training ranking and recommendation systems. The later generations are described as inference-first, while the 450 and 500 can also handle workloads such as GenAI training. Meta has not said that these generations will take over all training workloads.
Rank #2
- High-Performance AI Processing: The MX3 is designed to handle the most demanding AI computer vision workloads, delivering exceptional performance and efficiency.
- Flexible Integration: The MX3 can be easily integrated into your existing systems via its M.2 M-key form factor and support for Linux operating systems.
- Energy Efficient: The MX3 is designed to provide high performance while minimizing power consumption.
- Comprehensive Software Development Kit (SDK): The MX3 is supported by a comprehensive SDK that simplifies development and deployment.
- Hardware compatability: The MX3 is compatible with the PCI-SIG M.2 M-key 2280 Specification. It can be used with the Raspberry Pi 5 with a M-key 2280 HAT.
Is Meta replacing Nvidia GPUs with its own chips?
No wholesale replacement is established by Meta’s statements. On a 2025 follow-up call, Meta executive Chad Heaton said the company expected to keep buying silicon from leading suppliers while designing its own chips for workloads where off-the-shelf products are not optimal. Meta also said MTIA adoption for core ranking and recommendation inference began in the first half of 2024, with plans to expand use and replace some GPU-based servers as they reached the end of their useful lives. Meta’s Q1 2025 earnings-call transcript
The distinction is portfolio versus isolation: custom chips can serve selected workloads, while merchant silicon remains part of Meta’s infrastructure. Meta has described its own chips as potentially more compute-efficient and cost-efficient for their intended jobs, but it has not published a quantified savings figure in the cited announcement. That is a company claim, not an independently measured comparison.
Why software and data-center systems matter
An accelerator is useful only if software can run work on it and the surrounding servers can deploy it effectively. Meta says MTIA is built around PyTorch, vLLM, Triton, and Open Compute Project standards. It also says a modular design lets new chips fit into existing rack systems, reducing the need to redesign the whole infrastructure for each generation. Meta’s description of its MTIA system
Meta says this reusable approach supports a release cadence of every six months or less, compared with a typical industry cadence of one to two years. Both figures are Meta’s stated comparisons, not an independent industry benchmark.
Rank #4
- ✅Powered by 26 Tera-Operations Per Second (TOPS) Hailo-8 AI Processor. 2.5W typical power consumption
- ✅Scalable, enabling simultaneous processing of multi-streams & multi-models
- ✅Enabling real-time, low latency and high-efficiency AI inferencing on the edge devices
- ✅Supports TensorFlow, TensorFlow Lite, ONNX, Keras, Pytorch frameworks
- ✅Supports Linux and Windows. Supports the temperature range of -40°C to 85°C
A July 2026 preprint by researchers describing Triton on MTIA-2i offers a limited view of software deployment: the authors report production use across approximately 60 model types, covering 50% of layers and 47% of non-GEMM execution time for those models. Those figures apply to the specified models and execution category; they are not an overall MTIA performance benchmark. “Triton for MTIA: Bridging the Programming Model Gaps for Custom AI Accelerators”
Independent reader supportYour contribution helps us test, update, and keep practical guides available for everyone.What is known about Meta’s Iris chip?
In July 2026, Reuters reported that Meta planned to begin manufacturing a chip code-named Iris in September 2026, citing an internal memo it had reviewed. Reuters said Broadcom was helping with design and Taiwan Semiconductor Manufacturing Company (TSMC) would fabricate the processor. Meta declined to comment, according to the report. The report establishes a planned start, not that manufacturing actually began. Reuters’ July 2026 report
What’s actually slowing this PC down?
Pick the symptom - the matching free tool is one click away.
Best Value
- DEEPX DX-M1M NPU: Powered by the DEEPX DX-M1M neural processing unit, purpose-built for efficient on-device AI inference workloads.
- COMPACT M.2 2242 FORM FACTOR: Fits the standard M.2 2242 slot, making it easy to integrate into embedded systems, edge devices, and compact computing platforms.
- EDGE AI ACCELERATION: Designed to accelerate deep learning inference at the edge, enabling real-time AI applications without relying on cloud connectivity.
- RADXA AICORE MODULE: The Radxa AICore DX-M1M delivers a plug-and-play AI compute solution ideal for robotics, smart cameras, and industrial automation.
- WARRANTY AND ORIGIN: Backed by a 1-year manufacturer warranty and crafted with quality components for reliable long-term performance in demanding environments.
Reuters had earlier reported a small deployment test of Meta’s first in-house AI training chip in March 2025, with broader production contingent on the test going well. That was an early reported test milestone; it should not be confused with Meta’s later statement that MTIA 300 is already in production for ranking and recommendation training. Reuters’ March 2025 report
Quick Recap
What Meta’s chip strategy means
- Training has a concrete place in the program: Meta says MTIA 300 is in production for ranking and recommendation training.
- Inference remains the central near-term target: Meta says the 400, 450, and 500 generations are being developed primarily for GenAI inference.
- Meta is not claiming to abandon outside chips: its stated approach combines in-house accelerators with purchases from leading silicon suppliers.
- Deployment is a system challenge: software compatibility, racks, and data-center integration are part of the plan alongside chip design.
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
