AMAZON ANNOUNCED COMING EC2 INSTANCES BASED ON GAUDI

New Amazon EC2 Instance based on Gaudi will bring a new level of efficiency to AI training customers

Read Blog
AWS logo

On December 1, AWS announced Gaudi-based EC2 Instances with customer availability targeted to first half of 2021

AWS Product info
SynapseAI

Our aim is to make developing workloads on Gaudi easy, whether you're developing from scratch or migrating existing workloads.

Enablement whitepaper

Gaudi Training Products

EFFICIENT SCALE IS AT THE FOUNDATION OF GAUDI'S ARCHITECTURE

Gaudi:
The only AI processor to provide the
game-changing advantages of integrated,
on-chip RoCE v2

Habana integrates ten 100 GbE ports of RoCE RDMA over Converged Ethernet–into every Gaudi processor to deliver unmatched advantages to customers to efficiently scale AI Training from one processor to 1000s for data parallel and model parallel systems.

Industry-standard RoCEv2 RDMA
Eliminates proprietary lock-in
On-chip integration reduces bottlenecks
More on Gaudi's integrated RoCE

A NEW WAY TO SCALE WITH GAUDI:
THE HLS-1H SERVER

The HLS-1H provides unprecedented cross-sectional bandwidth with RoCEv2 RDMA. Featuring four GAUDI HL-205 mezzanine cards, each having ten 100GbE ports dedicated to scale-out, the HLS-1H delivers massive external scale out of 40x 100 GbE.


Scaling Training Details

HABANA SYNAPSE AI SOFTWARE SUITE FOR TRAINING

SynapseAIHabana’s SynapseAI software suite is designed to facilitate high-performance DL training on Gaudi accelerators. The software suite includes Habana’s graph compiler and runtime, communication libraries, TPC kernel library, firmware and drivers. SynapseAI is integrated with TensorFlow and PyTorch frameworks, and performance-optimized for Gaudi

TPC KERNEL LIBRARY & SDK

  • Programmable TPC with SDK
  • Extensive TPC kernel libraries support user customization
  • TPC tools (compiler, simulator, debugger)  for custom kernel development

TCP Habana

COMPILER, RUNTIME & OPTIMIZER

  • Seamlessly integrates with TensorFlow and PyTorch
  • Can be interfaced with C or Python API
  • Generates optimized binary code that implements the given model topology on Gaudi.
  • Enables pipelining, layer fusing

SynapseAI Habana

PERFORMANCE MANAGEMENT & DEPLOYMENT TOOLS

  • Deployment validation toolset: test and monitor functionality and performance
  • Libraries and tools for run-time detection and reporting
  • Run-time and orchestration plug-ins

Software Tools icon
SynapseAI Enablement Whitepaper