The Graphcore team has been active innovating at all levels including their manufacturing methods to deliver new third generation IPUs, with increased core count and faster processing which are now the most advanced AI systems for performance, power and scalability.
These new IPUs deliver 40% higher performance with a 16% reduction in power consumption through a new wafer-on-wafer manufacturing technique. This is a world first use of 3D semiconductor technology which delivers much higher processor bandwidth and power efficiency, in the same IPU form factor.
Why does this matter for AI? Model parameters and AI dataset are increasing in size, and models with a billion parameters are now common. All of this requires more memory and capacity at the compute level to reduce I/O cycles and time. Graphcore’s ability to handle larger models results in faster training and inference time and improved model accuracy.
Graphcore’s Bow systems deliver the best performance per $ when compared to alternative GPU based systems, often by a factor of 10x.
Bow is named after the area in London where Graphcore is located, and the new Bow IPU is available in all Pod configurations from 16 to 1024, with the Pod256 delivering more than 89 PetaFLOPS of AI compute while the Bow Pod1024 unleashes 350 PetaFLOPS!
See the details of the Bow Pod systems here on the XENON website.
Performance Gains for Bow IPUs
With the 40% increase in processing power, customers are seeing performance gains of 1.3x across all types of AI workloads – from image classification, object detection, GNN, NLP and Speech.
This impressive boost to processing power has come with a reduction in power consumption, making the Bow systems even more efficient.
How to Get Started
Bow Systems come in powerful 1RU form factor which houses 4x IPUs and delivers 1.4 PetaFLOPS. You can start with just one of these Bow-2000 units, and scale up to interconnected modules of 4 Bow-2000’s with a dedicated host server controller. Bow Pod systems are certified with a number of OEM server partners and XENON supplies all varieties and will match the host server performance requirements to your applications and Pod set-up.
XENON works with customers to optimise the associated infrastructure to ensure maximum performance from the Bow Systems – this includes:
- High speed data storage, and data management to ensure the AI dataset is accessible and protected
- High speed networking
Contact XENON today to find out how Graphcore’s new Bow systems can accelerate your AI workloads.
Review the Bow Pod system specifications here.
Talk to a Solutions Architect
