Meta、数十億ユーザーの推論コスト削減に向けて4世代の独自AIチップを発表
本文の状態
日本語全文を表示中
詳細モードで約1分の本文を読めます。
同じ出来事の情報源
この情報源を基点に整理
The Decoder
Metaが推論処理に特化した4世代の独自AIチップを発表し、NvidiaやAMDなどのGPUメーカーへの依存度低減を推進している。
Source Article
元記事を日本語で読む
本文に関係しない購読案内、埋め込み通知、サイト内プロモーションは除いています。
Metaは、推論に特化した4つの新世代カスタムAIチップを発表しました。これにより、NVIDIAやAMDといったGPUメーカーへの依存度を低減させる取り組みを進めています。
この記事「Meta unveils four generations of custom AI chips to cut inference costs for billions of users」は、The Decoderで最初に公開されました。
原文を表示
Meta has unveiled four new generations of custom AI chips—MTIA 300, 400, 450, and 500—designed to make AI cheaper to run across its platforms.
The chips are being developed in partnership with Broadcom and are built to make AI applications more cost-effective for the billions of users on Meta's platforms. Meta says it's following a roughly six-month development cycle per chip generation. From MTIA 300 to 500, memory bandwidth (HBM) increases by a factor of 4.5 and computing power jumps 25x.
MTIA 300 is optimized for ranking and recommendation models (R&R) and is already in production, according to Meta. MTIA 400 is the first generation that Meta says can compete with leading commercial products on raw performance. A rack of 72 chips forms a single scale-up domain. MTIA 400 has completed lab testing and is currently being rolled out to data centers.
MTIA 450 and 500 target generative AI inference
MTIA 450 and 500 are specifically optimized for generative AI inference. MTIA 450 doubles the HBM bandwidth compared to MTIA 400, outperforming existing commercial products, according to Meta. The chips support low-precision data formats like MX4 and MX8, which cut the computing power needed for inference without significantly hurting model quality. MTIA 500 adds another 50 percent HBM bandwidth and up to 80 percent more HBM capacity. Both chips are scheduled for mass production in 2027.
Metric
MTIA 300
MTIA 400
MTIA 450
MTIA 500
Workload Focus
R&R Training
General
GenAI Inference
GenAI Inference
Module TDP
800 W
1200 W
1400 W
1700 W
HBM Bandwidth
6.1 TB/s
9.2 TB/s
18.4 TB/s
27.6 TB/s
HBM Capacity
216 GB
288 GB
288 GB
384-512 GB
MX4 Performance
-
12 PFLOPs
21 PFLOPs
30 PFLOPs
FP8/MX8 Performance
1.2 PFLOPs
6 PFLOPs
7 PFLOPs
10 PFLOPs
BF16 Performance
0.6 PFLOPs
3 PFLOPs
3.5 PFLOPs
5 PFLOPs
Scale-up domain size
16
72
72
72
Scale-up network
(unidirectional bandwidth*)
1 TB/s
1.2 TB/s
1.2 TB/s
1.2 TB/s
Scale-out network
(unidirectional bandwidth*)
200 GB/s**
100 GB/s
100 GB/s
100 GB/s
On the software side, Meta built the chips around industry standards like PyTorch, vLLM, and Triton. Developers can port existing models to MTIA without special adaptations and run them on GPUs and MTIA at the same time. More technical details are available on Meta's blog.
Meta also continues to work with AMD and Nvidia for GPUs. In early February 2026, Meta announced a billion-dollar deal with AMD to provide up to six gigawatts of AMD Instinct GPU computing power for Meta's AI workloads.
AI News Without the Hype – Curated by Humans
Subscribe to THE DECODER for ad-free reading, a weekly AI newsletter, our exclusive "AI Radar" frontier report six times a year, full archive access, and access to our comment section.
Subscribe now
今日のまとめ
AIデイリーブリーフで今日の重要ニュースをまとめ読み