Our IPU lets innovators make new breakthroughs in machine intelligence

Explore Products

IPU-POD₁₆

Pre-configured with a 4 petaFLOP AI system, IPU-POD₁₆ is where you experience the power and flexibility of larger IPU systems.

Learn More

IPU-POD₆₄

16 petaFLOPS of AI-compute for both training and inference workloads, the IPU-POD₆₄ is designed for AI at scale.

Learn More

IPU-POD₂₅₆

With a powerful 64 petaFLOPS of AI compute, our flagship IPU-POD256 is designed to provide cutting-edge performance for machine intelligence at datacenter scale.

Learn More

Graphcloud

A secure IPU cloud service to add state of the art AI compute on demand - no on-premise infrastructure deployment required.

Request Access

WORLD-CLASS PERFORMANCE

State of the art performance for natural language processing and computer vision

Best for Natural Language Processing

BERT-Large training in 5.88 minutes. These are the blistering fast results you get on an IPU-POD₁₂₈ in the latest MLPerf 1.1 benchmark results in the open category for production hardware, delivering real benefits to our customers.

MLPerf v1.1 Training Results | MLPerf ID: 1.1-2088, 1.1-2089, 1.1-2087

Best for Computer Vision

Our latest MLPerf submission for ResNet-50 show the IPU-POD₁₆ platform outperforming the latest DGX A100 platform, training ResNet in 28.3 minutes.

IPU systems really shine with next-generation CV models like EfficientNet. With the IPU-POD's impressive scaling capabilities, you can train EfficientNet in 1.7 hours on an IPU-POD₂₅₆ while an IPU-POD₁₆ trains in 20.7 hours - 3.5x faster than the 70.5 hours published result for DGX A100.

IPU-POD SDK2.4 Results

DGX A100 (A100-SXM4-80GB) | Published results

Best for Large Models

IPU-POD systems can support today’s large AI models, indicated by the impressive scaling of GPT-class models.

The performance capabilities of the larger IPU-POD systems, fully managed by the Poplar SDK, allows customers to experiment and to build new types of large models.

IPU-POD systems can scale to support the largest brain-scale models with trillions of parameters.

SDK2.3 Throughput Results

More Benchmarks

Our Customers

We benchmarked with a wide range of interesting state-of-the-art networks and we see the IPUs are at least multiple times faster than the common GPUs. To give a number, I think it would be at least 10x faster