
Meta Training and Inference Accelerator (MTIA v2)
Co-designed for Meta's massive recommendation, ranking, and inference fleets across 16+ global regions, Next-Gen MTIA v2 packs an 8×8 processing-element grid on a 421 mm² TSMC 5nm die, delivering 708 sparse INT8 TFLOPS (354 dense), 256 MB of 2.7 TB/s on-chip SRAM, and 128 GB LPDDR5 at a frugal 90W TDP. MTIA v2 rounds out our top 10 as the textbook blueprint for domain-specific hyperscale efficiency—tripling compute over MTIA v1 and achieving 6x platform serving throughput while freeing Meta's H100/Blackwell clusters for frontier Llama pre-training. Its internal-only deployment and 204.8 GB/s off-chip LPDDR5 keep it at #10.
Peer Contenders Ranked Alongside Meta Training and Inference Accelerator (MTIA v2)
| Entity / Contender | Category / Type | Shared Rankings | 1-Tap Showdown |
|---|---|---|---|
| Meta Training and Inference Accelerator (MTIA v2) Gadget · Official Profile | Gadget | 4 Active Lists | Subject Benchmark |
| NVIDIA Blackwell B200 & GB200 NVL72 | GADGET | 4 Shared Lists | |
| Microsoft Azure Maia 200 | GADGET | 4 Shared Lists | |
| NVIDIA H200 Tensor Core GPU | GADGET | 4 Shared Lists | |
| AWS Trainium2 | GADGET | 4 Shared Lists | |
| Intel Gaudi 3 AI Accelerator | GADGET | 4 Shared Lists |
Gadget · Official Profile
Vote on Rankings Featuring Meta Training and Inference Accelerator (MTIA v2)
Contender: Meta Next-Gen MTIA v2
Contender: Meta Next-Gen MTIA v2
Contender: Meta Next-Gen MTIA v2
Contender: Meta Next-Gen MTIA v2
Official Stats & Specifications
Should Meta Training and Inference Accelerator (MTIA v2) rank higher than #10 in NVIDIA Blackwell vs. Google TPU v6e? Top 10 AI Datacenter Chips & Accelerators?
Ranked #10 on "NVIDIA Blackwell vs. Google TPU v6e? Top 10 AI Datacenter Chips & Accelerators in 2026". Cast your vote to back or challenge this spot.
Be the first to vote on this ranking.