Meta MTIA 300
Meta's first MTIA chip built for large-scale training, with on-chip networking
N/A
Datacenter
216GB
HBM3E
72
cores
Meta detailed the MTIA 300 on August 24, 2026, its first MTIA chip aimed at large training workloads rather than just inference. It pairs a 12x6 grid of processing elements with 216GB of HBM3E and two network chiplets carrying twelve custom 800 Gbps RDMA NICs, giving each chip 1.2 TB/s of I/O bandwidth so training jobs can scale across racks without relying on separate networking hardware.
// similar hardware in Datacenter
NVIDIA B200
NVIDIA Blackwell flagship datacenter GPU for frontier model training
NVIDIA · Datacenter · N/A
NVIDIA H100 SXM
Hopper-architecture datacenter GPU for large-scale AI training
NVIDIA · Datacenter · ~$30000
NVIDIA H200
High-memory Hopper refresh for large-model inference
NVIDIA · Datacenter · ~$32000
AMD Instinct MI300X
AMD flagship datacenter accelerator for large-model AI training
AMD · Datacenter · ~$15000
AMD Instinct MI325X
High-memory CDNA 3 refresh for large-model inference
AMD · Datacenter · ~$16000
NVIDIA A100 80GB
Ampere-generation datacenter GPU, still widely deployed
NVIDIA · Datacenter · ~$15000