Training and inference
have different requirements
Training and inference workloads have
different compute, memory, scale-out, latency
and power requirements
Training processor for inference
Delivering two dedicated product lines

Using a training processor for inference is costly
and results in inefficient use of power.
As a result, Habana delivers
two dedicated product lines.

Inference: Habana Goya
AI Habana processors
  • PCIe Card: PCIe Gen4 X 16 lanes
  • Form Factors: Dual-Slot or Single Slot
  • Memory: 4/8/16GB memory with ECC
  • Mixed-precision Numeric

Download datasheet >>

Training: Habana Gaudi
Available 2019
Learn about Habana Goya
inference platform
Abstraction to drive
developer adoption
Habana’s unified software stack seamlessly integrates with all popular frameworks
Abstraction to drive developer adoption
habana Goya logo logo habana gaudi
Inference at scale with
Habana SynapseAI
Inference optimizer and runtime toolkit

Habana SynapseAI compiler executes
ahead-of-time (AOT) compilations.
It is used to optimize the graph and create
a working plan for Goya, including:

Graph compiler
Graph
compiler
Quantization optimizer
Quantization
optimizer
AI performance library
AI performance
library
Quantization optimizer
Runtime
API
SynapseAI supports models trained by any processor: GPU, TPU, CPU and Habana Gaudi
Advanced profiling
and customization options

Habana’s Neural-Network Profiler allow power users to identify bottlenecks at compile-time and optimize deep-learning models. Habana’s TPC’s Software Development Tools allow advanced users to add their own custom performance libraries

Neural-Network Profiler
Full trace capability
Graphical views
Performance analysis
Real time
Development Tools
LLVM based
C Compiler
IDE: Debugger
/ Simulator
demonstration AI processors performance
demonstration AI processors performance
demonstration AI processors performance
llvm - compiler infrastructure