Specifications
Family
MTIA 300, MTIA 400, MTIA 450 and MTIA 500 — four successive generations
Release Cadence
Roughly six months between generations
Silicon Partner
Broadcom, under a multi-year multi-generation partnership running through 2029
Deployment Scale
Initial commitment exceeds 1 GW, the first phase of a multi-gigawatt rollout
MTIA 300 Status
Already in production for ranking and recommendation (R&R) training
MTIA 400 Status
In lab testing ahead of datacentre deployment
MTIA 450 Status
GenAI inference part, mass deployment targeted early 2027
MTIA 500 Status
Adds more memory running at higher speed, mass deployment later in 2027
HBM Bandwidth Scaling
4.5x from MTIA 300 through MTIA 500
Compute Scaling
25x compute FLOPs from MTIA 300 through MTIA 500
Workload Coverage
From R&R inference to R&R training, general GenAI workloads and GenAI inference
Infrastructure Reuse
All four generations share the same basic infrastructure so chips can be swapped as they are upgraded
Design Philosophy
Inference-first, built natively on industry standards for frictionless adoption
Co-design
Each generation is co-designed with Meta's software stack and guided by the trajectory of future models
Lineage
Follows the earlier MTIA 100 and MTIA 200 generations
Announced
April 2026 alongside the extended Broadcom partnership
Overview
Meta's MTIA programme has moved to a rapid, iterative cadence rather than a single flagship chip per generation. MTIA 300, 400, 450 and 500 were announced together, covering four successive generations on roughly a six-month cadence, all developed with Broadcom under a partnership now extended through 2029.
The spread of capability is substantial: from MTIA 300 to MTIA 500, HBM bandwidth increases 4.5x and compute FLOPs increase 25x. The workload map expands in parallel — MTIA 300 is already in production for ranking and recommendation training, MTIA 400 is in lab testing ahead of datacentre deployment, MTIA 450 targets GenAI inference with mass deployment in early 2027, and MTIA 500 adds more memory at higher speed for deployment later in 2027.
The strategic point is infrastructure reuse. All four generations share the same basic system infrastructure, so Meta can swap accelerators as they are upgraded without redesigning the surrounding server, rack or software. Combined with an initial commitment exceeding 1 GW, this is one of the largest custom-silicon programmes outside the hyperscale cloud providers. QS Compute supplies AI accelerators, inference servers and rack-scale AI infrastructure — request a quote.
Key Benefits
Rapid cadence: four generations on a six-month release rhythm. 25x compute, 4.5x bandwidth: measured scaling from MTIA 300 to MTIA 500. Swap-in infrastructure: one platform across four silicon generations. Proven in production: MTIA 300 already serving ranking and recommendation training workloads.
Applications
Large-scale recommendation and ranking inference, ranking and recommendation training, general generative AI workloads, GenAI inference at social-platform scale, and datacentre acceleration targeting multi-gigawatt deployments.
Request a Quote — META MTIA 300 / 400 / 450 / 500 — FOUR-GENERATION AI ACCELERATOR FAMILY
QS Compute — global B2B supply of AI computing hardware, edge AI systems and accelerators. Volume pricing, 15-day sample lead time.
Get Your Quote →