A Google TPU (Tensor Processing Unit) is an AI accelerator that Google designs for its own models and rents to other companies through Google Cloud. As of October 2026, Google Cloud documents TPU v5e, v5p, v6e (Trillium) and TPU7x (Ironwood), and lists two eighth-generation chips, TPU 8t for training and TPU 8i for inference, as coming soon. Google says TPUs are “the engine behind Gemini” and its products including Search, Photos and Maps.

On this page

How a TPU differs from a GPU

A TPU is built around matrix-multiply units (MXUs), the hardware that does the bulk of the arithmetic in neural networks. Google’s documentation says each MXU is a grid of multiply-accumulators, 256 × 256 on TPU v6e and TPU7x and 128 × 128 on earlier versions, and that each MXU performs 16K multiply-accumulate operations per cycle.

The other half of the design is scale. Google groups TPUs into pods: chips joined by Google’s own inter-chip interconnect (ICI), so that thousands of them can work on one model. Pod size is one of the clearest differences between generations. TPU v5e and v6e pods hold 256 chips; v5p and TPU7x pods hold 8,960 and 9,216.

TPU generations at a glance

Figures are per chip, as Google states them in its Cloud TPU documentation and announcements. Note the precision in each row: FP4, FP8, BF16 and INT8 figures are not directly comparable.

Generation Peak compute per chip HBM per chip HBM bandwidth per chip Chips per pod Status (Google Cloud)
TPU v5e 197 TFLOPs (BF16); 393 TOPs (INT8) 16 GB 800 GiBps 256 Documented
TPU v5p 459 TFLOPs (BF16); 459 TFLOPs (FP8) 95 GiB 2,765 GBps 8,960 Documented
TPU v6e (Trillium) 918 TFLOPs (BF16); 1,836 TOPs (INT8) 32 GB 1,638 GBps 256 Generally available
TPU7x (Ironwood) 2,307 TFLOPs (BF16); 4,614 TFLOPs (FP8) 192 GiB 7,380 GBps 9,216 Generally available
TPU 8t 12.6 PFLOPs (FP4) 216 GB 6,528 GB/s 9,600 per superpod Coming soon
TPU 8i 10.1 PFLOPs (FP4) 288 GB 8,601 GB/s Up to 1,024 active chips per pod Coming soon

Google’s documentation also covers TPU v2, v3 and v4. Accelerator specs from every vendor are collected in our AI chip catalog.

v5e and v5p

The fifth generation came as two products of very different size. Google describes TPU v5e as supporting “single and multi-host training and single-host inference,” with 16 GB of HBM per chip, eight chips per host and 256-chip pods. TPU v5p carries 95 GiB of HBM per chip at 2,765 GBps and scales to pods of 8,960 chips.

Trillium (v6e)

Trillium, the sixth generation, keeps the 256-chip pod of v5e but raises per-chip compute to 918 BF16 TFLOPs and HBM to 32 GB. Google says it is “optimized for transformer, text-to-image, and convolutional neural network (CNN) training, fine-tuning, and serving,” and that it delivers 4.7x higher peak compute per chip than v5e with 67% better energy efficiency. Inter-chip bandwidth is 800 GBps per chip.

Ironwood (TPU7x)

Google calls TPU7x “the first release within the Ironwood family” and, in its documentation, “the latest TPU available on Google Cloud.” Google describes Ironwood as built for large-scale training, reasoning and inference. Each chip has 192 GiB of HBM at 7,380 GBps and peak compute of 2,307 BF16 TFLOPs or 4,614 FP8 TFLOPs. Pods hold 9,216 liquid-cooled chips, which Google says provide 42.5 exaflops, with 4x better performance per chip than Trillium. Inter-chip bandwidth is 1,200 GBps per chip.

TPU 8t and TPU 8i

On April 22, 2026, Google announced its eighth generation as two chips rather than one. TPU 8t is for training: Google lists 12.6 FP4 petaflops, 216 GB of HBM, 128 MB of on-chip SRAM and superpods of 9,600 chips in a 3D torus, with what it calls “nearly 3x the compute performance per pod over the previous generation.” TPU 8i is for post-training and inference: 10.1 FP4 petaflops, 288 GB of HBM at 8,601 GB/s, 384 MB of on-chip SRAM and a network Google calls Boardfly, with up to 1,152 chips fully connected.

Google’s comparisons with Ironwood: up to 2.7x better training performance per dollar for TPU 8t, up to 80% better inference performance per dollar for TPU 8i, and up to 2x better performance per watt for both. Both use Google’s Arm-based Axion CPUs as hosts. In April, Google said both chips “will be generally available later this year.” As of October 2026, its TPU product page lists them as coming soon and invites interested customers to request more information.

Who can use Google TPUs

Outside companies use TPUs as a Google Cloud service:

  • Google Cloud customers rent TPU capacity through Compute Engine and Google Kubernetes Engine (GKE). Google’s documentation covers both routes for TPU7x and mentions reservations as an access option. GKE deployments can scale to 130,000 nodes, according to Google.
  • Frameworks: Google says Cloud TPUs support TensorFlow, PyTorch and JAX.
  • Google itself: TPUs train and serve Gemini and run Search, Photos and Maps, according to Google.
  • AI labs at scale: on October 23, 2025, Anthropic announced it would expand its use of Google Cloud TPUs to “up to one million TPUs,” bringing “well over a gigawatt of capacity online in 2026.” Our AI companies directory tracks the labs and their compute partners.

Large AI campuses and their announced capacity are listed in the AI data center tracker.

Sources