LogoLogo
Continuum WebsiteContinuum ApplicationsContinuum KnowledgeAxolotl Platform
Continuum Knowledge
Continuum Knowledge
  • Continuum
  • Data
    • Datasets
      • Pre Training Data
      • Types of Fine Tuning
      • Self Instruct Paper
      • Self-Alignment with Instruction Backtranslation
      • Systematic Evaluation of Instruction-Tuned Large Language Models on Open Datasets
      • Instruction Tuning
      • Instruction Fine Tuning - Alpagasus
      • Less is More For Alignment
      • Enhanced Supervised Fine Tuning
      • Visualising Data using t-SNE
      • UMAP: Uniform Manifold Approximation and Projection for Dimension Reduction
      • Training and Evaluation Datasets
      • What is perplexity?
  • MODELS
    • Foundation Models
      • The leaderboard
      • Foundation Models
      • LLama 2 - Analysis
      • Analysis of Llama 3
      • Llama 3.1 series
      • Google Gemini 1.5
      • Platypus: Quick, Cheap, and Powerful Refinement of LLMs
      • Mixtral of Experts
      • Mixture-of-Agents (MoA)
      • Phi 1.5
        • Refining the Art of AI Training: A Deep Dive into Phi 1.5's Innovative Approach
      • Phi 2.0
      • Phi-3 Technical Report
  • Training
    • The Fine Tuning Process
      • Why fine tune?
        • Does Fine-Tuning LLMs on New Knowledge Encourage Hallucinations?
        • Explanations in Fine Tuning
      • Tokenization
        • Tokenization Is More Than Compression
        • Tokenization - SentencePiece
        • Tokenization explore
        • Tokenizer Choice For LLM Training: Negligible or Crucial?
        • Getting the most out of your tokenizer for pre-training and domain adaptation
        • TokenMonster
      • Parameter Efficient Fine Tuning
        • P-Tuning
          • The Power of Scale for Parameter-Efficient Prompt Tuning
        • Prefix-Tuning: Optimizing Continuous Prompts for Generation
        • Harnessing the Power of PEFT: A Smarter Approach to Fine-tuning Pre-trained Models
        • What is Low-Rank Adaptation (LoRA) - explained by the inventor
        • Low Rank Adaptation (Lora)
        • Practical Tips for Fine-tuning LMs Using LoRA (Low-Rank Adaptation)
        • QLORA: Efficient Finetuning of Quantized LLMs
        • Bits and Bytes
        • The Magic behind Qlora
        • Practical Guide to LoRA: Tips and Tricks for Effective Model Adaptation
        • The quantization constant
        • QLORA: Efficient Finetuning of Quantized Language Models
        • QLORA and Fine-Tuning of Quantized Language Models (LMs)
        • ReLoRA: High-Rank Training Through Low-Rank Updates
        • SLoRA: Federated Parameter Efficient Fine-Tuning of Language Models
        • GaLora: Memory-Efficient LLM Training by Gradient Low-Rank Projection
      • Hyperparameters
        • Batch Size
        • Padding Tokens
        • Mixed precision training
        • FP8 Formats for Deep Learning
        • Floating Point Numbers
        • Batch Size and Model loss
        • Batch Normalisation
        • Rethinking Learning Rate Tuning in the Era of Language Models
        • Sample Packing
        • Gradient accumulation
        • A process for choosing the learning rate
        • Learning Rate Scheduler
        • Checkpoints
        • A Survey on Efficient Training of Transformers
        • Sequence Length Warmup
        • Understanding Training vs. Evaluation Data Splits
        • Cross-entropy loss
        • Weight Decay
        • Optimiser
        • Caching
      • Training Processes
        • Extending the context window
        • PyTorch Fully Sharded Data Parallel (FSDP)
        • Train Short, Test Long: Attention with Linear Biases Enables Input Length Extrapolation
        • YaRN: Efficient Context Window Extension of Large Language Models
        • Sliding Window Attention
        • LongRoPE
        • Reinforcement Learning
        • An introduction to reinforcement learning
        • Reinforcement Learning from Human Feedback (RLHF)
        • Direct Preference Optimization: Your Language Model is Secretly a Reward Model
  • INFERENCE
    • Why is inference important?
      • Grouped Query Attention
      • Key Value Cache
      • Flash Attention
      • Flash Attention 2
      • StreamingLLM
      • Paged Attention and vLLM
      • TensorRT-LLM
      • Torchscript
      • NVIDIA L40S GPU
      • Triton Inference Server - Introduction
      • Triton Inference Server
      • FiDO: Fusion-in-Decoder optimised for stronger performance and faster inference
      • Is PUE a useful measure of data centre performance?
      • SLORA
  • KNOWLEDGE
    • Vector Databases
      • A Comprehensive Survey on Vector Databases
      • Vector database management systems: Fundamental concepts, use-cases, and current challenges
      • Using the Output Embedding to Improve Language Models
      • Decoding Sentence-BERT
      • ColBERT: Efficient and Effective Passage Search via Contextualized Late Interaction over BERT
      • SimCSE: Simple Contrastive Learning of Sentence Embeddings
      • Questions Are All You Need to Train a Dense Passage Retriever
      • Improving Text Embeddings with Large Language Models
      • Massive Text Embedding Benchmark
      • RocketQAv2: A Joint Training Method for Dense Passage Retrieval and Passage Re-ranking
      • LLM2Vec: Large Language Models Are Secretly Powerful Text Encoders
      • Embedding and Fine-Tuning in Neural Language Models
      • Embedding Model Construction
      • Demystifying Embedding Spaces using Large Language Models
      • Fine-Tuning Llama for Multi-Stage Text Retrieval
      • Large Language Model Based Text Augmentation Enhanced Personality Detection Model
      • One Embedder, Any Task: Instruction-Finetuned Text Embeddings
      • Vector Databases are not the only solution
      • Knowledge Graphs
        • Harnessing Knowledge Graphs to Elevate AI: A Technical Exploration
        • Unifying Large Language Models and Knowledge Graphs: A Roadmap
      • Approximate Nearest Neighbor (ANN)
      • High Dimensional Data
      • Principal Component Analysis (PCA)
      • Vector Similarity Search - HNSW
      • FAISS (Facebook AI Similarity Search)
      • Unsupervised Dense Retrievers
    • Retrieval Augmented Generation
      • Retrieval-Augmented Generation for Large Language Models: A Survey
      • Fine-Tuning or Retrieval?
      • Revolutionising Information Retrieval: The Power of RAG in Language Models
      • A Survey on Retrieval-Augmented Text Generation
      • REALM: Retrieval-Augmented Language Model Pre-Training
      • Retrieve Anything To Augment Large Language Models
      • Generate Rather Than Retrieve: Large Language Models Are Strong Context Generators
      • Active Retrieval Augmented Generation
      • DSPy: LM Assertions: Enhancing Language Model Pipelines with Computational Constraints
      • DSPy: Compiling Declarative Language Model Calls
      • DSPy: In-Context Learning for Extreme Multi-Label Classification
      • Optimizing Instructions and Demonstrations for Multi-Stage Language Model Programs
      • HYDE: Revolutionising Search with Hypothetical Document Embeddings
      • Enhancing Recommender Systems with Large Language Model Reasoning Graphs
      • Retrieval Augmented Generation (RAG) versus fine tuning
      • RAFT: Adapting Language Model to Domain Specific RAG
      • Summarisation Methods and RAG
      • Lessons Learned on LLM RAG Solutions
      • Stanford: Retrieval Augmented Language Models
      • Overview of RAG Approaches with Vector Databases
      • Mastering Chunking in Retrieval-Augmented Generation (RAG) Systems
    • Semantic Routing
    • Resource Description Framework (RDF)
  • AGENTS
    • What is agency?
      • Rephrase and Respond: Let Large Language Models Ask Better Questions for Themselves
      • Types of Agents
      • The risk of AI agency
      • Understanding Personality in Large Language Models: A New Frontier in AI Psychology
      • AI Agents - Reasoning, Planning, and Tool Calling
      • Personality and Brand
      • Agent Interaction via APIs
      • Bridging Minds and Machines: The Legacy of Newell, Shaw, and Simon
      • A Survey on Language Model based Autonomous Agents
      • Large Language Models as Agents
      • AI Reasoning: A Deep Dive into Chain-of-Thought Prompting
      • Enhancing AI Reasoning with Self-Taught Reasoner (STaR)
      • Exploring the Frontier of AI: The "Tree of Thoughts" Framework
      • Toolformer: Revolutionising Language Models with API Integration - An Analysis
      • TaskMatrix.AI: Bridging Foundational AI Models with Specialised Systems for Enhanced Task Completion
      • Unleashing the Power of LLMs in API Integration: The Rise of Gorilla
      • Andrew Ng's presentation on AI agents
      • Making AI accessible with Andrej Karpathy and Stephanie Zhan
  • Regulation and Ethics
    • Regulation and Ethics
      • Privacy
      • Detecting AI Generated content
      • Navigating the IP Maze in AI: The Convergence of Blockchain, Web 3.0, and LLMs
      • Adverse Reactions to generative AI
      • Navigating the Ethical Minefield: The Challenge of Security in Large Language Models
      • Navigating the Uncharted Waters: The Risks of Autonomous AI in Military Decision-Making
  • DISRUPTION
    • Data Architecture
      • What is a data pipeline?
      • What is Reverse ETL?
      • Unstructured Data and Generatve AI
      • Resource Description Framework (RDF)
      • Integrating generative AI with the Semantic Web
    • Search
      • BM25 - Search Engine Ranking Function
      • BERT as a reranking engine
      • BERT and Google
      • Generative Engine Optimisation (GEO)
      • Billion-scale similarity search with GPUs
      • FOLLOWIR: Evaluating and Teaching Information Retrieval Models to Follow Instructions
      • Neural Collaborative Filtering
      • Federated Neural Collaborative Filtering
      • Latent Space versus Embedding Space
      • Improving Text Embeddings with Large Language Models
    • Recommendation Engines
      • On Interpretation and Measurement of Soft Attributes for Recommendation
      • A Survey on Large Language Models for Recommendation
      • Model driven recommendation systems
      • Recommender AI Agent: Integrating Large Language Models for Interactive Recommendations
      • Foundation Models for Recommender Systems
      • Exploring the Impact of Large Language Models on Recommender Systems: An Extensive Review
      • AI driven recommendations - harming autonomy?
    • Logging
      • A Taxonomy of Anomalies in Log Data
      • Deeplog
      • LogBERT: Log Anomaly Detection via BERT
      • Experience Report: Deep Learning-based System Log Analysis for Anomaly Detection
      • Log-based Anomaly Detection with Deep Learning: How Far Are We?
      • Deep Learning for Anomaly Detection in Log Data: A Survey
      • LogGPT
      • Adaptive Semantic Gate Networks (ASGNet) for log-based anomaly diagnosis
  • Infrastructure
    • The modern data centre
      • Enhancing Data Centre Efficiency: Strategies to Improve PUE
      • TCO of NVIDIA GPUs and falling barriers to entry
      • Maximising GPU Utilisation with Kubernetes and NVIDIA GPU Operator
      • Data Centres
      • Liquid Cooling
    • Servers and Chips
      • The NVIDIA H100 GPU
      • NVIDIA H100 NVL
      • Lambda Hyperplane 8-H100
      • NVIDIA DGX Servers
      • NVIDIA DGX-2
      • NVIDIA DGX H-100 System
      • NVLink Switch
      • Tensor Cores
      • NVIDIA Grace Hopper Superchip
      • NVIDIA Grace CPU Superchip
      • NVIDIA GB200 NVL72
      • Hopper versus Blackwell
      • HGX: High-Performance GPU Platforms
      • ARM Chips
      • ARM versus x86
      • RISC versus CISC
      • Introduction to RISC-V
    • Networking and Connectivity
      • Infiniband versus Ethernet
      • NVIDIA Quantum InfiniBand
      • PCIe (Peripheral Component Interconnect Express)
      • NVIDIA ConnectX InfiniBand adapters
      • NVMe (Non-Volatile Memory Express)
      • NVMe over Fabrics (NVMe-oF)
      • NVIDIA Spectrum-X
      • NVIDIA GPUDirect
      • Evaluating Modern GPU Interconnect
      • Scalable Hierarchical Aggregation and Reduction Protocol (SHARP)
      • Next-generation networking in AI environments
      • NVIDIA Collective Communications Library (NCCL)
    • Data and Memory
      • NVIDIA BlueField Data Processing Units (DPUs)
      • Remote Direct Memory Access (RDMA)
      • High Bandwidth Memory (HBM3)
      • Flash Memory
      • Model Requirements
      • Calculating GPU memory for serving LLMs
      • Transformer training costs
      • GPU Performance Optimisation
    • Libraries and Complements
      • NVIDIA Base Command
      • NVIDIA AI Enterprise
      • CUDA - NVIDIA GTC 2024 presentation
      • RAPIDs
      • RAFT
    • Vast Data Platform
      • Vast Datastore
      • Vast Database
      • Vast Data Engine
      • DASE (Disaggregated and Shared Everything)
      • Dremio and VAST Data
    • Storage
      • WEKA: A High-Performance Storage Solution for AI Workloads
      • Introduction to NVIDIA GPUDirect Storage (GDS)
        • GDS cuFile API
      • NVIDIA Magnum IO GPUDirect Storage (GDS)
      • Vectors in Memory
Powered by GitBook
LogoLogo

Continuum - Accelerated Artificial Intelligence

  • Continuum Website
  • Axolotl Platform

Copyright Continuum Labs - 2023

On this page
  • RAPIDS Visualisation Guide
  • Key features of RAPIDS
  • cuDF In-Depth
  • Performance benchmarks
  • RAPIDS Accelerator for Apache Spark
  • Installation

Was this helpful?

  1. Infrastructure
  2. Libraries and Complements

RAPIDs

RAPIDS is an open-source suite of GPU-accelerated libraries developed by NVIDIA that enables end-to-end data science and analytics pipelines to run entirely on GPUs.

It is designed to provide orders of magnitude speed-up compared to CPU-based alternatives for data loading, preprocessing, machine learning, graph analytics and more.

RAPIDS aims to bring GPU acceleration to the data science ecosystem without requiring significant code changes.

RAPIDS Visualisation Guide

The RAPIDS visualisation guide showcases how RAPIDS can accelerate interactive plotting of large datasets.

It highlights popular plotting libraries that are RAPIDS-accelerated:

  • HoloViews, hvPlot: Declarative APIs for building interactive plots and dashboards

  • Datashader: For rasterizing large datasets into images for display

  • Plotly, Bokeh: For interactive web-based visualisations

  • Seaborn: For static statistical graphics

These libraries can directly use cuDF dataframes instead of pandas for 5-100x faster plotting. The guide provides GPU-accelerated Interactive examples that dynamically plot millions of points.

It also covers general concepts like the importance of using a Jupyter server local to the GPU machine, handling large datasets that exceed GPU memory, and using Dask for distributed rendering.

Overall it demonstrates how to build end-to-end visual analytics pipelines that are GPU-accelerated.

Key features of RAPIDS

Accelerated data processing: RAPIDS includes libraries like cuDF, a GPU-accelerated DataFrame library that mimics the Pandas API, allowing users to process and manipulate large datasets efficiently on GPUs.

Machine learning and deep learning: cuML, a component of RAPIDS, provides GPU-accelerated implementations of popular machine learning algorithms, enabling faster training and inference.

Graph analytics: cuGraph is a library within RAPIDS that offers GPU-accelerated graph algorithms for analysing large-scale graph datasets.

Spatial analytics: cuSpatial is a library that provides GPU-accelerated spatial and trajectory data processing capabilities.

Data visualisation: cuxfilter is a RAPIDS library that enables the creation of interactive, GPU-accelerated dashboards for data visualization.

Seamless integration: RAPIDS integrates well with existing data science workflows and libraries, such as scikit-learn, NumPy, and Pandas, making it easier for data scientists to adopt GPU acceleration.

cuDF In-Depth

cuDF is a GPU DataFrame library that provides a pandas-like API for loading, filtering, aggregating and manipulating tabular data on GPUs. Key features of cuDF include:

Uses Apache Arrow columnar memory format on GPU for efficient processing:

  • Apache Arrow is a cross-language development platform for in-memory data.

  • It provides a standardised columnar memory format for efficient storage and processing of tabular data.

  • cuDF leverages the Apache Arrow format to store data in a columnar layout on the GPU.

  • Columnar data layout allows for better memory access patterns and enables SIMD (Single Instruction Multiple Data) operations on GPUs.

  • By using Apache Arrow, cuDF achieves efficient memory utilisation and high-performance data processing on GPUs.

Provides Python and C++ APIs, integrates with Numba, CuPy, Dask

  • cuDF offers both Python and C++ APIs for data manipulation and processing.

  • The Python API provides a user-friendly interface similar to pandas, allowing users to work with familiar DataFrame operations.

  • The C++ API enables low-level access to cuDF functionalities and allows for high-performance custom operations.

  • cuDF integrates seamlessly with other GPU-accelerated libraries such as Numba, CuPy, and Dask.

  • Numba is a just-in-time (JIT) compiler that can be used to accelerate custom Python functions on GPUs.

  • CuPy is a GPU-accelerated library for numerical computations that provides a NumPy-compatible interface.

  • Dask is a flexible library for parallel computing that can be used to scale cuDF operations across multiple GPUs or distributed systems.

Supports reading from a range of file formats

  • cuDF supports reading data from various file formats commonly used in data processing and analytics.

  • CSV (Comma-Separated Values) is a plain text format where each line represents a row, and values are separated by commas.

  • Parquet is a columnar storage format that provides efficient compression and encoding schemes for fast data retrieval.

  • ORC (Optimised Row Columnar) is another columnar storage format that offers high compression and efficient querying.

  • JSON (JavaScript Object Notation) is a lightweight data interchange format that is easy for humans to read and write.

  • cuDF can read data directly from these file formats into GPU memory, enabling fast data loading and processing.

Has a wide range of data types - numerics, strings, categoricals, timestamps

  • cuDF supports a wide range of data types to represent different kinds of data.

  • Numeric data types include integers (e.g., int8, int16, int32, int64) and floating-point numbers (e.g., float32, float64).

  • String data type represents textual data and allows for efficient string operations.

  • Categorical data type is used for storing data with a fixed number of unique values, providing efficient memory usage and fast comparisons.

  • Timestamp data type represents date and time values and enables time-based operations and analysis.

Provides Intra-dataframe operations

  • Intra-dataframe operations are performed within a single DataFrame.

  • Filter operation selects rows based on specified conditions, allowing for data subsets.

  • Project operation selects specific columns from the DataFrame, enabling column-wise transformations.

  • Aggregation operations perform computations across rows or columns, such as sum, average, count, etc.

  • Sort operation arranges the rows of the DataFrame based on one or more columns, in ascending or descending order.

  • These operations are crucial for data manipulation and analysis within a DataFrame.

Intra-dataframe operations

Intra-dataframe operations are for data manipulation and analysis within a single DataFrame.

These operations allow you to transform, filter, and aggregate data efficiently. \

  1. Filter operation:

    • The filter operation allows you to select a subset of rows from a DataFrame based on specified conditions.

    • It enables you to extract relevant data points that meet certain criteria, such as selecting rows where a column value exceeds a threshold or matches a specific pattern.

    • Filtering is useful for data cleansing, outlier detection, and focusing on specific subsets of interest.

    • Example: Selecting rows where the "age" column is greater than 18 to analyze adult population data.

  2. Project operation:

    • The project operation allows you to select specific columns from a DataFrame, creating a new DataFrame with only the chosen columns.

    • It enables you to focus on relevant features or variables for analysis and exclude unnecessary columns.

    • Projection is useful for dimensionality reduction, feature selection, and creating derived columns.

    • Example: Selecting only the "name" and "salary" columns from an employee DataFrame to analyze salary distribution.

  3. Aggregation operations:

    • Aggregation operations perform computations across rows or columns of a DataFrame to summarize the data.

    • Common aggregation operations include sum, average, count, minimum, maximum, etc.

    • Aggregations provide insights into the overall properties and trends of the data.

    • They are useful for statistical analysis, data summarization, and generating reports.

    • Example: Calculating the average sales price per product category to identify top-performing categories.

  4. Sort operation:

    • The sort operation arranges the rows of a DataFrame based on one or more columns, in ascending or descending order.

    • Sorting helps organize the data in a meaningful way and facilitates data exploration and analysis.

    • It is useful for ranking, identifying top or bottom values, and presenting data in a structured manner.

    • Example: Sorting a student DataFrame by the "grade" column in descending order to identify top-performing students.

These intra-dataframe operations are essential for data manipulation and analysis within a single DataFrame.

They allow you to filter relevant subsets, focus on specific columns, summarise data through aggregations, and organise the data through sorting.

By combining these operations, you can extract meaningful insights and prepare the data for further analysis or visualisation.

Provides Inter-dataframe operations like joins, unions, merges

  • Inter-dataframe operations involve combining or merging multiple DataFrames.

  • Join operation combines rows from two DataFrames based on a common key column, allowing for data integration from different sources.

  • Union operation concatenates two DataFrames vertically, stacking the rows of one DataFrame below the other.

  • Merge operation combines DataFrames horizontally, aligning rows based on common columns or indexes.

  • These operations facilitate data combination and integration from multiple DataFrames.

Inter-dataframe operations

Inter-dataframe operations involve combining or merging multiple DataFrames to integrate data from different sources or perform complex analyses.

These operations are important when working with data spread across multiple DataFrames or when you need to combine data based on common attributes.

  1. Join operation:

    • The join operation combines rows from two DataFrames based on a common key column.

    • It allows you to integrate data from different sources or DataFrames based on a shared attribute.

    • Joins are useful for combining related data, such as merging customer information with their purchase history.

    • Different types of joins (inner, left, right, outer) can be used depending on the desired output and handling of missing values.

    • Example: Joining a "customers" DataFrame with an "orders" DataFrame based on the "customer_id" column to analyze customer purchasing behavior.

  2. Union operation:

    • The union operation concatenates two DataFrames vertically, stacking the rows of one DataFrame below the other.

    • It is used when you have DataFrames with the same columns and want to combine their rows into a single DataFrame.

    • Union is useful for combining data from multiple sources or time periods into a unified dataset.

    • Example: Combining sales data from multiple years into a single DataFrame for trend analysis.

  3. Merge operation:

    • The merge operation combines DataFrames horizontally, aligning rows based on common columns or indexes.

    • It is similar to the join operation but provides more flexibility in terms of merging criteria and handling of missing values.

    • Merging is useful for combining data from different sources or DataFrames based on multiple columns or indexes.

    • Example: Merging a "products" DataFrame with a "sales" DataFrame based on the "product_id" and "date" columns to analyze product sales over time.

These inter-dataframe operations are essential for combining and integrating data from multiple DataFrames.

They allow you to join related data based on common attributes, concatenate DataFrames vertically for unified analysis, and merge DataFrames horizontally based on multiple criteria.

By leveraging these operations, you can build comprehensive datasets, perform complex analyses, and gain insights from data spread across different sources or DataFrames.

Inter-dataframe operations are particularly useful in scenarios where data is distributed across multiple files or databases, or when you need to combine data from different departments or systems within an organisation. They enable you to break down data silos and create a holistic view of the data for effective decision-making and analysis.

Implements Reductions operations

  • Reduction operations aggregate data across rows or columns to produce a single result.

  • Sum operation calculates the sum of values in a column or across rows.

  • Min and max operations find the minimum and maximum values in a column or across rows.

  • Unique operation returns the unique values in a column or DataFrame.

  • Nunique operation counts the number of unique values in a column or DataFrame.

  • These operations are useful for data summarization and statistical analysis.

Time-series operations

  • Time-series operations are specific to data that is ordered by time.

  • Moving window functions calculate metrics over a sliding window of data points, such as rolling average or rolling standard deviation.

  • Sampling operation selects data points at regular intervals or based on a specific frequency.

  • Shifting operation moves the data points forward or backward in time, allowing for lag or lead analysis.

  • These operations are essential for analyzing and manipulating time-series data.

Integrates with cuML

  • cuML is a GPU-accelerated machine learning library that is part of the RAPIDS ecosystem.

  • cuDF seamlessly integrates with cuML, allowing for end-to-end GPU-accelerated machine learning pipelines.

  • Data loaded and processed using cuDF can be directly passed to cuML for training and inference.

  • This integration enables fast and efficient machine learning workflows entirely on GPUs.

Supports multi-GPU and distributed processing via Dask-cuDF

  • Dask-cuDF is an extension of cuDF that enables multi-GPU and distributed processing.

  • It leverages Dask, a flexible parallel computing library, to scale cuDF operations across multiple GPUs or distributed systems.

  • Dask-cuDF allows for the processing of larger-than-memory datasets by partitioning the data across multiple GPUs.

  • It enables distributed computation and parallel execution of cuDF operations, providing scalability and faster processing times.

  • With Dask-cuDF, users can harness the power of multiple GPUs or distributed clusters for handling massive datasets.

These key functions of cuDF collectively provide a powerful and efficient framework for data processing, analysis, and machine learning on GPUs.

By leveraging the Apache Arrow memory format, offering a wide range of data types and operations, integrating with other GPU-accelerated libraries, and supporting multi-GPU and distributed processing, cuDF enables users to accelerate their data workflows and achieve high-performance computing on GPUs.

Performance benchmarks

show cuDF providing 10-100x speedup over pandas on a wide range of workloads. This allows data scientists to manipulate much larger datasets and iterate faster.

RAPIDS Accelerator for Apache Spark

A key innovation is the "RAPIDS Accelerator for Apache Spark" which can be used to accelerate Spark SQL and DataFrame operations transparently, without any application code changes.

The RAPIDS Accelerator for Apache Spark is a plugin that allows Spark SQL and DataFrame operations to be executed on NVIDIA GPUs for faster processing.

Overview

  • It allows existing Spark applications to leverage GPUs without code changes by replacing the backend for SQL and DataFrame operations with GPU-accelerated versions using the RAPIDS libraries like cuDF.

  • If an operation is not supported on GPU, it will automatically fall back to using the CPU version in Spark.

  • It provides an accelerated shuffle implementation that can do GPU-to-GPU transfers, bypassing the CPU for better performance. However, the GPU accelerated processing can be used independently of the accelerated shuffle.

Benefits

  • Significantly speeds up processing and reduces infrastructure costs compared to CPU-based Spark, especially for workloads shifting to AI/ML that strain traditional CPU frameworks. A mortgage dataset ETL was shown to get 3.5X better performance at 3.4X lower cost.

  • Existing Spark applications can be accelerated with no code changes simply by adding the RAPIDS plugin jar and enabling it with a config. Explains in the physical plan which operators run on GPU.

  • Provides a unified AI framework allowing GPU acceleration of the full pipeline from ETL to ML/DL in a single workflow.

Qualification Tool

  • The Qualification Tool helps identify which existing CPU-based Spark workloads are good candidates to be migrated to GPU.

  • It analyses Spark event logs from CPU runs and estimates potential speedup on GPU at an operator level based on benchmarks.

  • Combines speedup estimates with other heuristics to make recommendations on which applications to migrate.

  • Additionally provides optimized Spark configs for running on GPU and recommends GPU cluster sizing on cloud platforms.

  • Available as a pip package for both cloud (Dataproc, EMR, Databricks) and on-prem deployments.

So in summary, the RAPIDS Accelerator enables Spark workloads to transparently leverage GPUs for major speedups and cost savings, aided by the Qualification Tool to prioritise workloads for migration.

This empowers Spark as a unified framework for accelerated data analytics and AI.

Under the hood, it plugs into the Spark APIs and replaces CPU-based operations with GPU-accelerated cuDF operations where possible.

Installation

RAPIDS provides several installation methods depending on the environment and version:

  • Conda: The recommended method, allows creating ready-to-use RAPIDS environments. Miniconda or Anaconda can be used.

  • Docker: Containerised RAPIDS environment, useful for deploying in cloud environments. Requires Docker CE v19.03+ and nvidia-container-toolkit.

  • Pip: For existing Python environments, RAPIDS can be installed via pip. Requires exact matching of pip wheel to system CUDA toolkit version.

  • From Source: For development or custom builds.

The installation guide provides step-by-step instructions for each method.

It lists the hardware and software prerequisites (NVIDIA GPU Volta or later, CUDA 11.2+, Ubuntu/CentOS + GCC 9+).

Cloud instances from AWS, Azure, GCP etc. can also be used.

The installation is streamlined using the "Quick Start" tool which generates the exact conda/docker/pip commands based on your configuration choices.

Example

To install RAPIDS, you can follow the installation guide provided in the RAPIDS documentation. The installation process typically involves the following steps:

  1. Ensure that you have a compatible NVIDIA GPU and the necessary drivers installed.

  2. Install a supported version of Python and create a new virtual environment (optional but recommended).

  3. Use the conda package manager to install RAPIDS and its dependencies. The installation command may look like:

conda install -c rapidsai -c nvidia -c conda-forge rapids=0.19 python=3.8 cudatoolkit=11.0
  1. Verify the installation by importing RAPIDS libraries in a Python script or notebook.

Detailed installation instructions, including system requirements and troubleshooting steps, can be found in the RAPIDS documentation.

RAPIDS is designed to accelerate various stages of the data science pipeline, from data preprocessing and feature engineering to model training and inference.

It can significantly speed up data processing and machine learning tasks, especially when working with large datasets that can benefit from GPU acceleration.

PreviousCUDA - NVIDIA GTC 2024 presentationNextRAFT

Last updated 11 months ago

Was this helpful?

Page cover image