Catalog/MTIA v1

Meta MTIA v1

Meta Training & Inference Accelerator Architecture

Active

Launched May 2023

Core Specifications

VendorMeta
ArchitectureMeta Training & Inference Accelerator
Form Factor—
VRAM128 GB
Memory Bandwidth—
TDP400 W

Compute Performance

PrecisionTFLOPs

Performance Benchmarks

recommendation

ConfigurationPrecisionPerformanceSource
DLRM recommendation models (primary use case)—125,000 inferences_per_secondView

llm inference

ConfigurationPrecisionPerformanceSource
LLaMA 3 8B, optimized for Meta production—4,200 tokens_per_secondView

Quick Stats

Actions

Compare with OthersOfficial Specs

Similar XPUs

View other Meta GPUs or compare across vendors

All Meta XPUs →