Document Updates¶

The main updates to the documents are summarised below. In addition there will have been many incremental improvements and corrections.

Contents

October 2021 ¶

For full details of all the changes in this release, see the SDK 2.3 Release Notes.

New TensorFlow 1 Pipelining Tutorial.
The TensorFlow 2 Graph recompilation tutorial was removed, as similar content is available in the technical note on optimising for the IPU in TensorFlow. This now covers both TensorFlow 1 and TensorFlow 2.
The Distributed MNIST with Horovod Training Demo was removed because it uses a feature (HostAllReduce) which is no longer supported. For distributed training, use PopRun and PopDist instead. The example is compatible with SDK 2.1 and SDK 2.2.

Targeting the IPU from TensorFlow 2 has been revised to reflect the move to TensorFlow 2.4
The PopVision Analysis Library (libpva) User Guide and PopVision Trace Instrumentation Library documents have been split out from the Poplar & PopLibs User Guide
The Vertex/Assembler Programming Guide has been merged into the Poplar and PopLibs User Guide

For full details of all the changes in this release, see the SDK 2.2 Release Notes.

A new white paper on Graphcore’s AI-Float technology with hardware support for mixed-precision and stochastic rounding.

For full details of all the changes in this release, see the SDK 2.1 Release Notes.

The Graphcore GitHub repositories have been split into a section for tutorials and examples to help you get started, and another with sample applications for the IPU.
New tutorial on using the PopVision analysis library (libpva)
New example of using PopDist for distributed training
New technical note published on using the availableMemoryProportion option to optimise memory use: Optimising Temporary Memory Usage for Convolutions and Matmuls on the IPU
Updated TensorFlow 1 tutorial on half precision training:
- an improved section on loss scaling
- a new section on how to avoid numerical issues (overflow, underflow, unstable operations)
- and an additional example using the IPUEstimator
Updated PyTorch tutorial on half precision training:
- a new section on numerical stability (loss scaling, stochastic rounding)
- a section on debugging floating-point exceptions
- a section on PopTorch tracing
Updated Poplar profiling tutorial (tutorial 4):
- code for the IPU Model and for running on hardware (previously only the IPU Model could be used)
- an improved section on out-of-memory errors
- an improved section on navigating a profile generated by PopVision Graph Analyser

Published a new technical note on the implementation and performance of BERT-Large on Graphcore IPU-POD systems: Pre-Training and Fine-Tuning BERT for the IPU
The V-IPU Administrator Guide has been updated with a new chapter on monitoring application metrics using Prometheus and InfluxDB.

A new technical note has been published with an overview of how to add custom operations in PyTorch, TensorFlow and PopART: Creating Custom Operations for the IPU.
The technical note Model parallelism with TensorFlow: sharding and pipelining was revised to be consistent with SDK 2.0, and with the addition of content on pipelining Keras models in TensorFlow 2.
The “Writing custom operations” chapter in the TensorFlow 1 and TensorFlow 2 user guides has been completely rewritten,
New tutorials have been added on using half and mixed-precision floating point in PopTorch and TensorFlow 1, and efficient data loading with PopTorch

Since the SDK release, the following documents have been updated:

The Graphcore Command Line Tools document has been updated to 1.0.51. This adds a description of the information returned by the gc-info --tile-overview option and documents the new gc-exchangewritetest command.
The Targeting the IPU from TensorFlow 1 has been updated with an improved explanation of the role of XLA.

For full details of all the changes in this release, see the SDK 2.0 Release Notes.

The main documentation changes are:

Additions to:
New: PopDist and PopRun: User Guide documents the tool for running distributed programs and the associated PopDist library.
New and updated functions in the Poplar and PopLibs API Reference:
- New libraries:
  poplar/DebugContext.hpp
  
  poplar/TensorRearranger.hpp
  
  poplin/Cholesky.hpp
  
  poplin/TriangularSolve.hpp
  
  popnn/CTCLoss.hpp
  
  popnn/LogSoftmax.hpp
  
  popnn/Rnn.hpp
  
  popops/CollectiveTypes.hpp
  
  popops/SequenceSlice.hpp
  
  popops/SortOrder.hpp
  
  popops/TopK.hpp
  
  popsparse/MatMulParams.hpp
  
  poputil/OptionParsing.hpp
  
  poputil/TensorMetaData.hpp
- Changes:
  Renamed poplar/CycleEstimateFunc.hpp to poplar/PerfEstimateFunc.hpp
  
  Renamed popops/Collectives.hpp to popops/TensorCollectives.hpp
The PopVision trace instrumentation (PVTI) and PopVision analysis (PVA) libraries are now included in the in Poplar docs.
- pvti in the Poplar User Guide
- pvti API
- pva
- pva-python

There is now a contents page listing the examples and tutorials available on GitHub.
The IPU-M2000 DA system has been renamed to Getting Started with IPU-POD4 DA and IPU-POD16 DA and updated.

This includes the release of Poplar SDK 1.4, the IPU-M2000, the IPU-POD reference design and the supporting software tools.