.svg)
Graphcore builds the Intelligence Processing Unit (IPU), a massively parallel processor designed from the ground up for machine intelligence workloads. Unlike CPUs and GPUs, the IPU uses a MIMD (Multiple Instruction, Multiple Data) architecture with 1,472 independent IPU-Tiles, each containing its own IPU-Core and In-Processor-Memory. This design enables fine-grained, high-performance parallel computing suited to the irregular, sparse data patterns characteristic of modern AI and machine learning models.
The Bow IPU is Graphcore's latest processor generation and the first in the world to use Wafer-on-Wafer 3D stacking technology in partnership with TSMC. Each Bow IPU delivers up to 350 teraFLOPS of AI compute, representing a 40% performance improvement and up to 16% greater power efficiency compared to the previous generation. Key features include 900MB of In-Processor-Memory per IPU, 65.4TB/s memory bandwidth, 11TB/s all-to-all IPU-Exchange, and 10x IPU-Links providing 320GB/s chip-to-chip bandwidth.
Graphcore co-designs its hardware with the Poplar SDK, providing a software environment that works with popular machine learning frameworks. IPU hardware is accessible to developers and enterprises through cloud service partners, allowing teams to build, train, and deploy models using familiar tools. Pricing for data centre products is available on request through Graphcore's sales team.
- Accelerating training of large-scale machine learning models on dedicated IPU hardware
- Running inference workloads with high throughput and power efficiency on IPU processors
- Building and deploying deep learning models via cloud service providers offering IPU access
- Developing and testing ML applications using the Poplar SDK with popular ML frameworks
- Scaling AI compute workloads across multiple IPUs using IPU-Links chip-to-chip connectivity
- Achieving high-performance parallel processing for fine-grained AI workloads in data centres
- Accessing IPU compute in the cloud without on-premise hardware through cloud partner integrations
- Enabling semiconductor and AI research teams to benchmark new model architectures on IPU
- Deploying AI compute at scale for enterprise organizations requiring custom pricing and installation
- Exploring next-generation AI hardware capabilities through Graphcore's developer resources and documentation
