TPU v4
Google AI accelerator chip with 3D mesh networking
The TPU v4 is an artificial intelligence accelerator designed by Google for machine learning workloads. It features two TensorCores per chip, 32 GiB of high-bandwidth memory, and a reconfigurable 3D mesh interconnect for large-scale training pods.
- Each chip delivers 275 teraflops of peak compute
- Features 32 GiB HBM2 with 1200 GBps bandwidth
- Supports reconfigurable 3D mesh and twisted torus topologies
- Consumes between 90 and 192 watts depending on load
- Pods can scale up to 4096 interconnected chips
What it is
The TPU v4 is a specialised integrated circuit built by Google to accelerate machine learning tasks. Each chip contains two TensorCores, which are processing units dedicated to matrix mathematics. These cores handle operations in bfloat16 or int8 formats, delivering up to 275 teraflops of peak compute power. The chip includes 32 GiB of high-bandwidth memory (HBM2) to store model weights and data during training. It connects to other chips via a flexible network topology, allowing thousands of units to work together as a single supercomputer pod.
How it came about
Google developed the TPU v4 as a successor to the TPU v3, aiming to improve performance and memory efficiency for large language models. The design introduces a unified memory space across the chip, enabling better coordination between its two TensorCores. Google has historically kept its hardware details private until newer generations launch, but the v4 architecture was documented to support developers using Google Kubernetes Engine. The chip serves as the foundation for training massive AI models before the arrival of the v5p generation.
Inside the chip
The architecture centres on two TensorCores per die, each equipped with four matrix-multiply units, a vector unit, and a scalar unit. This layout doubles the number of multiplication units compared to previous designs. A key feature is the inter-chip interconnect, which uses six links per chip to form a three-dimensional mesh. Larger slices of chips can wrap around to create a 3D torus, or even a twisted torus, which increases bandwidth for global communication patterns. The system supports Non-Uniform Memory Access optimisations, allowing software to bind processes to specific memory nodes for faster throughput.
How it was made
Specific manufacturing details such as the foundry or process node size are not publicly listed in the provided documentation. The chip integrates 32 GiB of HBM2 memory with a bandwidth of 1200 GBps. Power consumption varies between 90 watts at minimum and 192 watts at maximum, with a mean draw of 170 watts. The physical package includes an x16 PCIe Gen3 interface for direct connection to host servers. Google reserves part of the physical memory for hardware reliability, so user-allocatable space is slightly less than the total capacity.
Where it went
These accelerators are deployed exclusively within Google Cloud infrastructure. They are managed via the Cloud TPU API or Google Kubernetes Engine. Users typically access them as part of a Pod, which can contain up to 4096 chips linked together. The hardware is suited for training large language models, recommendation systems, and image classification tasks. Twisted torus topologies are particularly effective for workloads that require heavy global communication, such as TPU embedding layers. Quota for these resources in specific zones requires manual approval from Google.
Versions and family
The TPU v4 sits between the TPU v3 and the newer TPU v5p in Google's product line. While the v4 supports standard 3D mesh and torus configurations, it also introduces twisted torus topologies for specific slice shapes like 4x4x8. The subsequent TPU v5p offers roughly 2.3 times better performance per dollar and clusters into larger pods of 8960 chips. The v4 remains supported for existing workloads, though Google recommends migrating to newer versions for Compute Engine usage as the Cloud TPU API enters maintenance mode.
Why it matters
The TPU v4 enabled Google to scale its machine learning capabilities significantly, supporting the training of complex models before the Gemini era. Its reconfigurable optical interconnects allow for flexible supercomputer arrangements without rewiring physical hardware. By offering high bisection bandwidth through twisted topologies, it improved efficiency for data-parallel workloads. The chip represents Google's strategy of keeping advanced AI hardware within its own cloud ecosystem, rather than selling it as a standalone component for on-premise use.
Questions people ask about TPU v4
How much memory does a single TPU v4 chip have?
Each TPU v4 chip contains 32 GiB of HBM2 memory. The bandwidth for this memory is 1200 GBps. Note that some physical memory is reserved for hardware reliability, so the user-allocatable amount reported by tools may be lower.
What is the power consumption of the TPU v4?
The measured power draw ranges from a minimum of 90 watts to a maximum of 192 watts. The mean power consumption during operation is approximately 170 watts. This efficiency supports dense packing in data centre pods.
Can I buy a TPU v4 for my own server?
No, the TPU v4 is not available for purchase as a standalone component. It is exclusively accessible through Google Cloud services, such as Google Kubernetes Engine or the Cloud TPU API, for users with approved quota.
What is a twisted torus topology?
It is a network configuration where wraparound connections shift coordinates, creating a symmetric topology. This design increases bisection bandwidth by up to 70 percent compared to non-twisted versions, benefiting workloads with global communication needs.
Good reading and sources
There is no English Wikipedia article on TPU v4 yet, so the facts on this page come from the sources listed above: cloud.google.com/tpu/docs/v4, servethehome.com/google-tpu-v5p-ai-chip-launches-alongside-gemini/. Some are the subject's own pages, so its claims about itself are reported as its claims. EveryChip wrote the text in its own words, with AI assistance, and compares it with the sources to make sure no passage is copied.
It may contain mistakes, and figures change over time. Nothing here is investment or buying advice. If something is wrong, tell us. Last refreshed 2026-11-14.