- XPU is a blanket term for different kinds of AI accelerators
- Unlike general purpose chips, accelerators are designed to serve specific types of AI workloads
- Public cloud players have their own accelerators but other XPUs are generally available
If you know one thing about AI, it’s probably that the technology relies heavily on chips called graphics processing units (GPUs) made by the likes of Nvidia and AMD. But such a complex technology – and its different use cases – requires more than just one kind of chip to run. That’s where XPUs come in.
The word “XPU” is a blanket term for a whole range of AI accelerators, which are chips designed to deliver enhanced performance for different kinds of workloads. And there are whole bunch of different kinds of XPUs, including generally available chips and custom options available only through certain cloud providers.
For instance, there are Google’s tensor processing units (TPUs). The first of these were designed for speeding AI training workloads, but Google has more recently focused on honing inferencing capabilities in its latest iteration.
Then there’s the Language Processing Unit (LPU) from rapidly-growing player Groq. The LPU is purpose-built for large language model inferencing workloads. As AI transitions from primarily training to more inference workloads (and as Groq continues to raise eye-popping sums), LPUs are garnering more attention. But they’ve been around since 2019.
Some AI accelerators don’t necessarily have an XPU-style name. Other AI accelerators include Amazon’s Trainium; Microsoft’s Maia 100 (production of which has reportedly been delayed to 2026); IBM’s Spyre, Intel’s Gaudi, Huawei’s Ascend series and forthcoming chips from collaborations between OpenAI and Broadcom and Rebellions and Marvell.