Meta Unveils Four Superior Custom AI Chips

🔑 Enhanced Key Takeaways

• •Meta's custom chips are part of the MTIA (Meta Training and Inference Accelerator) family, with the next-generation version featuring 128 MB on-chip memory, 64 GB off-chip LPDDR5, and clock speeds up to 1.35 GHz at 90 watts per chip[2].
• •Meta has a multi-year partnership with AMD to deploy up to 6 gigawatts of custom Instinct MI450 GPUs and 6th Gen EPYC CPUs (Venice/Venano), with first shipments in 2H 2026, complementing its in-house silicon efforts[3][5].
• •The MTIA chips are deployed in a rack-scale system holding up to 72 accelerators across three chassis, designed for co-optimized hardware-software stacks handling low-to-high complexity ranking and recommendation models[2].
• •Meta continues heavy reliance on Nvidia for training with over 1.3 million GPUs planned by end-2025, while shifting some inference workloads to AMD MI300X and future MI455X for cost-efficiency[6].

🛠️ Technical Deep Dive

• •Next-gen MTIA: TSMC 5nm process, clocked at 1.35 GHz (up from 800 MHz), power at 90W (up from 25W), with 128 MB on-chip memory and 64 GB off-chip LPDDR5[2].
• •Memory bandwidth: Local 400 GB/s per PE, on-chip 800 GB/s, off-chip LPDDR5 176 GB/s; compute at 2.76 TFLOPS/s (FP32)[2].
• •Rack system: Up to 72 accelerators (3 chassis x 12 boards x 2 chips), optimized for 10x-100x varying model sizes in ranking/recommendation workloads[2].
• •Earlier MTIA variant: 256 MB on-chip memory, 128 GB off-chip LPDDR5, local memory 384 KB per PE with 1 TB/s bandwidth per PE[2].

🔮 Future ImplicationsAI analysis grounded in cited sources

⏳ Timeline