AMD Instinct MI450 for Meta Cuts Memory to 144GB HBM4

Meta is shifting its data center strategy by adopting custom AMD Instinct MI450-series AI accelerators tailored specifically for recommendation systems. This move matters because it prioritizes cost efficiency and power savings over raw compute versatility, a trade-off that could reshape how social platforms scale their AI infrastructure. Buyers in the social media and recommendation engine space should watch this closely, as it signals a move away from general-purpose training chips toward specialized inference hardware.

Custom AI accelerators reduce memory and compute for recommendation systems

The custom hardware, identified as the Instinct MI455X for Meta, relies on the AMD Instinct MI450-series architecture. This design uses three times less memory than the standard Instinct MI455X model. The standard MI455X ships with 432 GB of HBM4 memory and 19.6 TB/s of bandwidth. The Meta-specific variant cuts this down to 144 GB of HBM4 memory using six 8-Hi packages.

Specifications

• Custom Memory: 144GB HBM4
• Standard Memory: 432GB HBM4
• Standard Bandwidth: 19.6 TB/s
• Target Workload: Recommendation systems

SemiAnalysis reports that this cut-down design significantly reduces compute power compared to the full model. The smaller memory footprint and reduced compute capabilities lower the Bill of Materials (BOM) costs for Meta. This specialized configuration allows the accelerators to consume significantly less power when running recommendation workloads. The trade-off is clear: the chips are optimized for specific tasks rather than general-purpose AI training.

This specialization limits the versatility of the MI450-series chips for other workloads. The custom units are less suitable for large language model training and inference compared to standard MI455X or Nvidia hardware. The design focuses entirely on the bandwidth-per-dollar ratio for recommendation systems. This approach aligns with Meta's previous agreement with AMD for a 6 GW supply deal over five years.

We have been tracking Instinct MI450-series AI Accelerator closely — see our earlier coverage on AMD Instinct MI450 GPU Sampling Begins,. The custom accelerators represent a targeted shift in Meta's hardware procurement strategy. This move highlights the growing importance of specialized AI chips for specific social platform workloads.