# TPU vs GPU: how Google's chips and NVIDIA's differ

> A TPU is Google's custom AI chip and a GPU is a general parallel processor. Here is how they compare after Google's TPU 8tx and 8i Max launch.

URL: https://claudegpt.si/news/tpu-vs-gpu-differences  
Published: 2026-11-09T14:55:35.849Z  
Author: ClaudeGPT Si Newsroom  
Section: Reviews  
Topics: Google, TPU, Nvidia, GPU, AI chips  
Publisher: ClaudeGPT Si

## Key takeaways

- A TPU is an application-specific chip Google built for matrix math, available through Google Cloud, while NVIDIA GPUs run on every major cloud.
- On November 9, 2026 Google added TPU 8tx and TPU 8i Max to the TPU 8t and 8i it introduced in April, with no price or date for the new two.
- Every performance number for both chips is a vendor claim from different tests, so the two cannot be compared directly.

A TPU is a chip Google designed only for machine learning, while a GPU is a general-purpose parallel processor that NVIDIA and others have adapted to AI. In the TPU vs GPU choice, TPUs are specialized hardware you rent from Google Cloud, and NVIDIA GPUs are the default almost everywhere else. On November 9, 2026, Google [added two more systems to its eighth-generation TPU line](https://cloud.google.com/blog/topics/ai-infrastructure/tpu-8tx-and-tpu-8i-max), TPU 8tx and TPU 8i Max, which makes the comparison a live one. Our [news story on the launch](/news/google-tpu-8tx-8i-max) has the announcement itself.

This page compares the two on design, scale, software and availability, using Google Cloud's documentation and NVIDIA's own pages. Every performance figure below is a vendor claim. Neither company's numbers are directly comparable to the other's, because they come from different tests.

## TPU vs GPU at a glance

| | Google TPU | NVIDIA GPU (Blackwell) |
| --- | --- | --- |
| What it is | An application-specific chip (ASIC) built mainly for matrix operations, per Google's docs | A parallel processor that runs many kinds of workloads, with Tensor Cores for AI |
| Core compute unit | Matrix-multiply units (MXUs), systolic arrays of multiply-accumulators | Tensor Cores, plus general CUDA cores |
| Who makes it | Google | NVIDIA |
| Where you get it | Google Cloud (Compute Engine, GKE, Vertex AI), and in customers' own data centers | Every major cloud, plus servers you buy |
| Main software | JAX and PyTorch, compiled through XLA | CUDA, with PyTorch and other frameworks on top |
| Scale-up unit | A Pod, a group of chips linked by inter-chip interconnect (ICI) | The NVL72 rack: 72 GPUs in one NVLink domain |
| Newest generation | TPU 8t, 8i (April 2026), 8tx, 8i Max (November 2026) | Blackwell and Blackwell Ultra; Vera Rubin announced |

## What a TPU is

Google's documentation describes a Tensor Processing Unit as an application-specific integrated circuit built mainly for the matrix operations that neural networks run on. Each chip has one or more TensorCores. A TensorCore holds matrix-multiply units, a vector unit for things like activations and softmax, and a scalar unit for control flow.

The matrix-multiply unit is a systolic array: a grid of multiply-accumulators that passes data from cell to cell. In Google's docs, TPU v6e and TPU7x use 256 x 256 arrays, and earlier versions use 128 x 128. Some versions also carry SparseCores, which speed up the embedding lookups that recommendation models depend on.

TPUs are sold as slices of a Pod, a group of chips on a dedicated network. Chips inside a slice talk over ICI. Larger slices use optical circuit switches to link groups of 64 chips into a 3D torus.

## What a GPU is

A GPU was built for graphics and turned out to suit AI, because both need huge amounts of parallel arithmetic. NVIDIA's data center GPUs now add dedicated hardware for AI. NVIDIA says its Blackwell GPUs have 208 billion transistors, two dies joined by a 10 TB/s link, a second-generation Transformer Engine, and support for FP4, a four-bit number format that cuts memory use.

The unit NVIDIA pushes hardest is the rack. Its GB200 NVL72 puts 72 Blackwell GPUs and 36 Grace CPUs in one liquid-cooled rack. NVIDIA lists 130 TB/s of NVLink bandwidth across the rack and 13.4 TB of HBM3E memory. NVIDIA's own claims against H100 include 30x faster real-time inference on trillion-parameter models and 4x faster training of a 1.8-trillion-parameter model. Both come with narrow test conditions on its page.

## Software: CUDA vs JAX, PyTorch and XLA

This is where the two differ most for a team deciding what to run.

NVIDIA describes CUDA as the foundation for GPU computing. It comes with a compiler, libraries and developer tools, and frameworks such as PyTorch run on it. Most AI code has been written and tested on GPUs first.

Google's docs name PyTorch and JAX as the frameworks to use on TPUs. In the November 9 post, Google says it open-sourced TorchTPU in October, so PyTorch code and Hugging Face models run natively on TPUs, with TPU backends for vLLM and SGLang. It also announced an open-source XLA Serving Engine that serves models across TPUs, GPUs and CPUs. Google gives no migration effort estimate, so the cost of moving an existing GPU codebase is something to test, not assume.

## Scale and networking

Both camps sell scale as the point. Google's April 22 Next '26 post says a TPU 8t superpod packs 9,600 chips, 121 exaflops of compute and two petabytes of shared memory, and that clusters can reach more than a million TPU chips, orchestrated by Pathways and JAX. NVIDIA's NVL72 is a 72-GPU domain, and its Blackwell page says fifth-generation NVLink scales to 576 GPUs. Google also says its Virgo network supports NVIDIA-based A5X instances at up to 80,000 GPUs in one data center.

The domains are not the same thing, and "exaflops" depends on the number format used. Treat these as descriptions of design philosophy, not a scoreboard.

## What Google's 8th-generation TPUs change

Google now sells four eighth-generation systems. It introduced two at Next '26 in April and two on November 9.

| System | Aimed at | What Google says |
| --- | --- | --- |
| TPU 8t | Large-scale training | Up to 2.7x better training performance than the previous generation (November post); 9,600 chips per superpod (April post) |
| TPU 8i | Low-latency inference and reinforcement learning | 384 MB of on-chip SRAM, 288 GB of HBM, 19.2 Tb/s ICI bandwidth, 80% better performance per dollar for inference than the prior generation |
| TPU 8tx | Secure, confidential inference and small-scale training | Same silicon as 8t with custom AMD CPU hosts; up to 2x lower latency than 8t |
| TPU 8i Max | Huge context windows, dense racks | Two 8i chips in one package: double the compute, SRAM and HBM bandwidth of 8i |

Google gives no absolute specs for 8tx or 8i Max, such as memory size or FLOPs, and no price. Availability is also open. In April Google said 8t and 8i would be available to Cloud customers "soon," and the November post gives no date for the new systems.

The split matters for the comparison. A GPU is one product that you point at training or inference. Google now sells separate TPU systems for each, and for confidential computing too. NVIDIA, for its part, still addresses confidential computing on the GPU itself: its Blackwell page calls it the first TEE-I/O capable GPU.

## Cost and availability

No honest price comparison is possible from published pages. Google's eighth-generation prices have not been published, and GPU prices vary by cloud, region and commitment. The one cost claim from Google is relative: 80% better inference performance per dollar for TPU 8i against the previous TPU generation, not against any GPU.

Availability is a practical difference. TPUs are available through Google Cloud and, per Google, deployed directly in some customers' data centers. NVIDIA GPUs run on every major cloud and in on-premises servers. Google also says it will offer NVIDIA's Vera Rubin NVL72 systems, so the same customer can use both. Google says nearly 100 customers use TPUs.

## Which to choose

- If your code is built on CUDA or you need to move between clouds, GPUs are the safer default.
- If you already run JAX or PyTorch on Google Cloud, or you train and serve at very large scale, TPUs are worth a benchmark on your own workload.
- If you need confidential inference for regulated data, compare TPU 8tx and NVIDIA's confidential computing once 8tx has a date and a price.

Independent results are the missing piece. Google cites SemiAnalysis evaluations and InferenceX and AgentX marks for its stack, and NVIDIA's multipliers come from its own tests. Run your model on both before committing money.

This comparison is based on the products' documentation and the vendors' published claims, not on our own testing.

## Frequently asked questions

### What is the difference between a TPU and a GPU?

A TPU is an application-specific chip Google designed mainly for the matrix operations in neural networks. A GPU is a more general parallel processor that NVIDIA has added AI hardware to, such as Tensor Cores.

### Is a TPU faster than a GPU?

No published source settles it. Google and NVIDIA each report gains over their own earlier chips, such as 2.7x training performance for TPU 8t and 30x inference for GB200 NVL72 against H100, under different test conditions.

### Can you buy a TPU like a GPU?

Google sells TPU access through Google Cloud, and says it also deploys TPUs directly in some customers' own data centers. NVIDIA GPUs are sold through all major clouds and as servers.

### Do TPUs run PyTorch?

Google's docs name PyTorch and JAX as the frameworks for TPUs. Google says its open-source TorchTPU lets PyTorch code and Hugging Face models run natively on TPUs.

## Sources

- [Our newest TPUs: 8tx and 8i Max, built for the agentic era](https://cloud.google.com/blog/topics/ai-infrastructure/tpu-8tx-and-tpu-8i-max), Google Cloud
- [AI infrastructure at Next '26](https://cloud.google.com/blog/products/compute/ai-infrastructure-at-next26), Google Cloud
- [TPU system architecture](https://docs.cloud.google.com/tpu/docs/system-architecture-tpu-vm), Google Cloud
- [NVIDIA GB200 NVL72](https://www.nvidia.com/en-us/data-center/gb200-nvl72/), NVIDIA
- [NVIDIA Blackwell architecture](https://www.nvidia.com/en-us/data-center/technologies/blackwell-architecture/), NVIDIA
- [CUDA Zone](https://developer.nvidia.com/cuda-zone), NVIDIA
