Guozhen AIGlobal AI field notes and model intelligence

Realtime AI News

Meta Orders Custom AMD MI450 Accelerators With 144GB HBM4 Memory

Meta has placed orders for custom AMD MI450 AI accelerators featuring 144GB of HBM4 memory, roughly one-third of the standard memory configuration. The custom order signals that large tech companies are increasingly seeking specialized AI chips optimized for their specific workloads.

Published
Meta向AMD定制MI450加速器,搭载144GB HBM4内存
Image source: ir.amd.com

According to MLQ.ai, Meta has submitted orders for custom AMD MI450 AI accelerators, marking the latest sign of large technology companies pursuing tailored AI silicon.

The custom MI450 accelerators feature 144GB of HBM4 high-bandwidth memory, approximately one-third of the standard memory configuration. This memory profile suggests Meta may have optimized the chips for specific inference or training workloads rather than maximizing general-purpose memory capacity.

The choice of HBM4 memory is notable. As the next-generation high-bandwidth memory standard, HBM4 offers significant bandwidth and power efficiency improvements over current HBM3e. AMD's adoption of HBM4 in the MI450 demonstrates its aggressive roadmap in high-performance AI computing.

Meta has previously collaborated with AMD on silicon, including deploying AMD EPYC processors in its data centers. This custom MI450 order deepens that partnership and reflects Meta's strategy of diversifying its AI infrastructure investments.

For AMD, landing a custom order from a hyperscale customer like Meta represents significant market validation for its AI product line. While NVIDIA still dominates the AI chip market, AMD is gaining ground through customization and differentiation.

The order also carries competitive implications for NVIDIA. When a major customer like Meta begins customizing AMD chips, it signals the AI infrastructure supply chain is diversifying away from reliance on a single general-purpose accelerator vendor.

The MI450 accelerators are expected to be deployed in Meta's AI inference and training clusters, though specific deployment timelines have not been announced.

Why it matters

Meta's custom AMD MI450 order is a strong signal of AI chip supply chain diversification, intensifying AMD's competition with NVIDIA in the data center AI accelerator market.

MetaAMDMI450HBM4AI Hardware
Back to realtime news

Nearby Updates

All

07/23, 19:55

Google Stock Falls Amid Questions Over AI Leadership and Gemini 4 Roadmap

Alphabet shares fell as investors questioned Google's AI leadership position and the clarity of its Gemini 4 roadmap. The market is growing concerned that Google may be falling behind competitors like OpenAI and Anthropic in the AI race.

07/23, 16:23

iFlytek Spins Off 'Yaofang Intelligent' to Build Robot Brains, Argues VLA Is Not the Final Answer

Chinese AI giant iFlytek has registered a new robotics AI subsidiary called Yaofang Intelligent, led by distinguished scientist Pan Jia, to tackle the robot brain problem. The company argues that the mainstream VLA approach is not the endgame for embodied AI and proposes adding 'body cognition' to world models.

07/23, 14:58

Alibaba Cloud's Zhenwu M890 Supernode Adapts Qwen3.8, Goes Live on Bailian for Inference

Alibaba's Zhenwu M890 supernode has successfully adapted Qwen3.8, Alibaba's flagship model with 2.4 trillion parameters, and is now available on the Bailian platform for inference services. The supernode interconnects 64 Zhenwu M890 chips via ICN Switch 1.0 at 800GB/s, making it the first supernode in China to run a model exceeding 2 trillion parameters.

07/23, 14:38

China releases first AI agent interconnection standard system, creating digital 'IDs' for AI agents

China has released its first AI agent interconnection standard system, establishing unified identity authentication and communication protocols for AI agents from different vendors. The standard aims to solve interoperability challenges that have been a bottleneck for enterprise AI agent deployment.