Google · v7p / v8 / v9

Google Tensor Processing Unit

Google's custom AI accelerator line. From Cloud Next 2026 onward Google bifurcates the eighth generation into Sunfish (8t, training, Broadcom-led) and Zebrafish (8i, inference, MediaTek-led) — both on TSMC N2 / CoWoS, targeted for 2027 H2. The MediaTek role expands beyond pure inference: per 工商時報 2026-04-27, MediaTek delivers I/O Die + back-end design service on the 8t training chip as well, using TSMC N3P + CoWoS-S. Roadmap continues to v9 Humufish, expected to move packaging to Intel EMIB to relieve CoWoS bottlenecks.

主要 variants

點任一張卡片進入該 variant 的完整 BOM 與供應商分析。

4 個 variant

TPU v7p Ironwood

2025 H2 → 2026 ramp

Currently-shipping Broadcom-designed training TPU; 192 GB HBM3e per chip, 4.6 PFLOPS FP8, 9216-chip superpod [8].

Liquid-cooled600 WTSMC N3

2026 FYmedium

~3500k 台 (3000k–4000k) · ASP ~0.0M USD

查看 BOM →

TPU 8t (Sunfish, training)

2027 H2

Eighth-gen training chip; Broadcom-led architecture on TSMC N2; 2 compute dies + 1 I/O chiplet + 8 stacks of 12-high HBM3e (≈30% more BW than v7p) [1]. MediaTek contributes I/O Die + back-end design service per 工商時報 2026-04-27, with that portion on TSMC N3P + CoWoS-S [3].

Liquid-cooledTSMC N2 (compute) / N3P (I/O via MediaTek)

2027 H2low

~500k 台 (200k–800k) · ASP ~0.0M USD

查看 BOM →

TPU 8i (Zebrafish, inference)

2027 H2

Eighth-gen inference chip; MediaTek-led on TSMC N2; 1 compute die + 1 I/O die + 6 stacks of HBM3e; targets 20–30% lower inference cost than 8t [1]. Previously referenced as v8e in pre-Cloud Next coverage.

Liquid-cooledTSMC N2

2027 H2low

~1500k 台 (500k–2500k) · ASP ~0.0M USD

查看 BOM →

TPU v9 (Humufish, roadmap)

2029 (rumor)

Ninth-gen TPU; codename Humufish; expected to migrate packaging to Intel EMIB to relieve CoWoS-S capacity bottlenecks [4][9]. Broadcom supply contract through 2031 [7] anchors v9 timing around 2029. Specs not publicly disclosed.

Liquid-cooled

2029 FYrumor

~2000k 台 (500k–5000k) · ASP ~0.0M USD

查看 BOM →

Google · v7p / v8 / v9

Google Tensor Processing Unit

主要 variants

點任一張卡片進入該 variant 的完整 BOM 與供應商分析。

4 個 variant

TPU v7p Ironwood

2025 H2 → 2026 ramp

Currently-shipping Broadcom-designed training TPU; 192 GB HBM3e per chip, 4.6 PFLOPS FP8, 9216-chip superpod [8].

Liquid-cooled600 WTSMC N3

2026 FYmedium

~3500k 台 (3000k–4000k) · ASP ~0.0M USD

查看 BOM →

TPU 8t (Sunfish, training)

2027 H2

Liquid-cooledTSMC N2 (compute) / N3P (I/O via MediaTek)

2027 H2low

~500k 台 (200k–800k) · ASP ~0.0M USD

查看 BOM →

TPU 8i (Zebrafish, inference)

2027 H2

Liquid-cooledTSMC N2

2027 H2low

~1500k 台 (500k–2500k) · ASP ~0.0M USD

查看 BOM →

TPU v9 (Humufish, roadmap)

2029 (rumor)

Liquid-cooled

2029 FYrumor

~2000k 台 (500k–5000k) · ASP ~0.0M USD

查看 BOM →