Bow Pod16

Product · Hardware / device by Graphcore

Accelerating large-scale machine intelligence in the cloud.

Bow Pod16 is a high-performance AI compute system designed for large-scale machine learning training and inference. It provides businesses with cloud-native access to Graphcore's IPU technology for accelerating complex models like GPT and Vision Transformers.

Languages
English, Korean, Japanese, Chinese

Category: AI Infrastructure

System Type
IPU-POD AI Compute System
Supported Frameworks
PyTorch, TensorFlow, Poplar SDK
Cloud Platforms
Tencent Cloud, Kingsoft Cloud
Workload Types
Training, Fine-tuning, Inference
Target Applications
NLP, Computer Vision, GNN, ViT, GPT
  • Optimized for training and inference of large-scale AI models
  • Available via Tencent Cloud IPU Preview service
  • Supports GPT, Vision Transformers, and Graph Neural Networks
  • Part of the Bow Pod family of AI compute systems

Overview

Bow Pod16 is a high-performance AI compute system developed by Graphcore for businesses and AI researchers. It is designed to accelerate demanding machine learning workloads, including the training and inference of large-scale models like GPT and Vision Transformers. The system is available as a cloud-native service through partners such as Tencent Cloud, allowing organizations to scale their AI infrastructure without investing in on-premises hardware. Bow Pod16 leverages Graphcore's Intelligence Processing Unit (IPU) technology and the Poplar software stack to provide high efficiency for sparse computing and complex neural networks. It serves industries ranging from finance and healthcare to meteorology by providing the compute power needed for next-generation AI applications.

Compute and Performance

  • AI Workload Acceleration: The system is engineered to accelerate large-scale machine learning workloads for both training and inference.
  • Broad Model Support: Supports innovative AI models including Vision Transformers (ViT), Generative Pre-trained Transformers (GPT), and Graph Neural Networks (GNN).
  • Domain Versatility: Optimized for Natural Language Processing (NLP), computer vision, recommendation systems, and probability tasks.
  • Sparse Computing Support: Designed to handle sparse computing workloads efficiently using Graphcore's Intelligence Processing Unit technology.

Deployment and Integration

  • Cloud-Native Access: Available through Tencent Cloud's IPU Preview service for rapid deployment without local data centers.
  • Enterprise Cloud Integration: Integrates with Tencent Cloud's Cloud Hosting Cluster (CHC) for storage and clustering solutions.
  • Kubernetes Support: Utilizes Tencent Kubernetes Engine (TKE) for containerized workload management.
  • Cloud Storage Compatibility: Supports Cloud Object Storage (COS) and Cloud File Storage (CFS) for data-intensive applications.

Use cases

  • AI researchers use Bow Pod16 to pre-train and fine-tune large language models like GPT with high computational efficiency.
  • Enterprise developers deploy computer vision applications such as image recognition on cloud-hosted IPU systems for faster inference.

Specifications

System Type
IPU-POD AI Compute System
Supported Frameworks
PyTorch, TensorFlow, Poplar SDK
Cloud Platforms
Tencent Cloud, Kingsoft Cloud
Workload Types
Training, Fine-tuning, Inference
Target Applications
NLP, Computer Vision, GNN, ViT, GPT

How it works

  1. Cloud Access Setup: Users can access Bow Pod16 through the Tencent Cloud IPU Preview platform to evaluate and deploy AI workloads.
  2. Model Mapping: Developers use the Poplar SDK to compile and map their machine learning models (such as BERT or GPT) to the IPU hardware.
  3. Execution and Scaling: Workloads are executed across the IPU-POD system, utilizing high-bandwidth memory and parallel processing for acceleration.

FAQ

How can I access the Bow Pod16?

The Bow Pod16 is available through cloud-native access on Tencent Cloud as part of their IPU Preview service.

What AI models are supported on Bow Pod16?

It supports a wide range of models including GPT (Generative Pre-trained Transformers), Vision Transformers (ViT), Graph Neural Networks (GNN), and BERT.

Can Bow Pod16 be used for both training and inference?

Yes, the system is designed to support both the training of large-scale models and production-scale inference.

What software is used to manage Bow Pod16?

The system is supported by Graphcore's Poplar software stack, which provides the tools and libraries needed to map machine learning models to IPU hardware.

Added by AntSpace from public sources (graphcore.ai).

Related

More products by Graphcore