Meta Iris晶片評測：自研AI晶片9月量產甩開輝達 | Meta Iris Chip Review: In-House AI Silicon Ships in September

8月 31, 2026

By Kit 小克 | AI Tool Observer | 2026-09-01

🇹🇼 Meta Iris晶片評測：自研AI晶片9月量產甩開輝達

Meta Iris晶片將於2026年9月正式量產，這是Meta旗下MTIA（Meta Training and Inference Accelerator）自研AI晶片計畫的第一顆量產成品，目標是把運算容量在2027年前翻倍衝上14GW，同時降低對輝達（Nvidia）與AMD的依賴。這不是First Company要自研晶片，但Meta這次動作的規模和時間點，讓「大廠自研AI晶片」這個話題再度成為焦點。

Iris晶片是什麼？

Iris是Meta今年3月公布的MTIA晶片藍圖中的第一代量產機型，該藍圖一次揭露了MTIA 300、400、450、500四個世代的規劃。晶片設計由博通（Broadcom）協助，實際製造交給台積電（TSMC）代工。根據內部備忘錄，測試花了六週時間，過程中沒有發現重大問題，這也是Meta決定按原計畫在9月開始量產的原因。

為什麼Meta要自己做晶片

成本控制：輝達GPU單價高、供應吃緊，自研晶片長期能壓低單位運算成本
供應鏈自主：不必完全仰賴輝達或AMD的產能排程
針對性優化：MTIA可以針對Meta自家的推薦系統、廣告演算法、Llama系列模型做客製化調校，而非通用型GPU

對整體AI硬體市場的意義

Meta的目標是2026年底前達到7GW運算容量，2027年翻倍到14GW。這個規模级的算力擴張，如果部分能靠自研晶片撐起來，代表Meta每年在輝達身上省下的採購費用會是天文數字。這也解釋了為什麼近期AI晶片市場上，除了輝達的Blackwell/Rubin路線圖，還能看到Intel在Hot Chips 2026發表Crescent Island加速器、AMD MI350系列、以及各家自研ASIC百花齊放——大家都在找輝達之外的第二條路。

不過要注意，自研晶片不代表Meta會完全甩開輝達。從Google TPU、Amazon Trainium的經驗來看，自研晶片通常是用在特定、可預測的內部工作負載（比如推薦系統推論），訓練最前沿的大模型仍然高機率繼續用輝達GPU。Iris晶片短期內比較像是「省錢工具」，而非「輝達替代品」。

值得觀察的後續

Iris量產後的良率與實際部署規模，是否真的能支撐7GW目標
後續MTIA 400/450/500世代的效能是否能追上訓練級GPU
其他大廠（微軟、Google、亞馬遜）是否加速自研晶片投入

好不好用，試了才知道。

🇺🇸 Meta Iris Chip Review: In-House AI Silicon Ships in September

Meta's Iris chip is set to enter mass production in September 2026, marking the first production-ready silicon from Meta's in-house MTIA (Meta Training and Inference Accelerator) program. The goal: double Meta's compute capacity to 14 gigawatts by 2027 while cutting reliance on Nvidia and AMD. Custom AI silicon isn't new, but the scale and timing of Meta's move puts the "Big Tech builds its own AI chips" story back in the spotlight.

What Is the Iris Chip?

Iris is the first production chip from the MTIA roadmap Meta unveiled back in March, which laid out four chip generations at once: MTIA 300, 400, 450, and 500. Broadcom co-designed the chip, and manufacturing is handled by TSMC. According to an internal memo, testing took six weeks and turned up no major issues — which is why Meta is sticking to its original September production start.

Why Meta Is Building Its Own Silicon

Cost control: Nvidia GPUs are expensive and supply-constrained; custom silicon lowers per-unit compute cost over time
Supply chain independence: less dependence on Nvidia or AMD's production schedules
Targeted optimization: MTIA chips can be tuned specifically for Meta's own recommendation systems, ad algorithms, and Llama-family models rather than being general-purpose GPUs

What This Means for the AI Hardware Market

Meta is targeting 7GW of compute capacity by the end of 2026, doubling to 14GW in 2027. If even part of that expansion runs on in-house silicon instead of Nvidia GPUs, the annual savings on procurement would be enormous. That's also part of why the AI chip landscape has gotten crowded lately — alongside Nvidia's Blackwell/Rubin roadmap, Intel showed off its Crescent Island accelerator at Hot Chips 2026, AMD is pushing its MI350 series, and custom ASICs are popping up across the industry. Everyone is looking for a second option beyond Nvidia.

That said, custom silicon doesn't mean Meta is walking away from Nvidia entirely. Judging from Google's TPU and Amazon's Trainium playbooks, in-house chips typically handle specific, predictable internal workloads — like recommendation-system inference — while frontier model training still largely runs on Nvidia GPUs. In the near term, Iris looks more like a cost-saving tool than a true Nvidia replacement.

What to Watch Next

Whether Iris's actual yield and deployment scale can support the 7GW target
Whether later MTIA 400/450/500 generations can catch up to training-grade GPU performance
Whether other giants (Microsoft, Google, Amazon) accelerate their own custom silicon efforts in response

Good or not, you won't know until you try it.

Sources / 資料來源

延伸閱讀 / Related Articles

AI 工具觀察站 — 每日精選 AI Agent 與工具趨勢
AI Tool Observer — Daily curated AI Agent & tool trends

搜尋此網誌

AI小貼士