Back to Fastren

Graphcore

Freemium
artificial intelligencemachine learninghardwareipudeep learningai acceleratorsemiconductorhpc

Graphcore is a semiconductor company that develops Intelligence Processing Units (IPUs), purpose-built hardware and software systems designed to accelerate machine learning models for innovators in AI across various industries.


Graphcore creates specialized processors called IPUs (Intelligence Processing Units), a distinct hardware architecture compared to traditional CPUs and GPUs. The company serves AI researchers, data scientists, and enterprises building large-scale AI applications who require extreme performance and efficiency. Its unique value proposition lies in the IPU's novel architecture, featuring massive parallelism and substantial in-processor memory, which offers significant performance advantages for specific types of AI models. The accompanying Poplar software stack provides the necessary tools to program and optimize applications for this unique hardware. This combination enables developers to push the boundaries of AI research and deployment on a platform built specifically for that purpose.

Pros

  • Novel architecture designed specifically for AI graph structures, offering performance benefits on certain models.
  • Extremely high memory bandwidth from significant on-chip SRAM, reducing memory bottlenecks.
  • Designed for massive scalability from a single processor to thousands of IPUs working in concert.
  • Can deliver superior performance-per-watt for specific AI workloads compared to GPUs.
  • Comprehensive Poplar SDK with support for major ML frameworks like PyTorch and TensorFlow.

Cons

  • Requires using the proprietary Poplar SDK, which has a significant learning curve compared to the dominant CUDA ecosystem.
  • Performance benefits are most pronounced on specific model types and may not outperform top-tier GPUs universally.
  • Limited market share compared to NVIDIA, resulting in a smaller community and less third-party support.
  • Hardware is primarily accessible through a limited number of cloud partners or large-scale enterprise purchase.

Key features

  • Intelligence Processing Unit (IPU) Processor
  • Poplar Software Development Kit (SDK)
  • Support for PyTorch, TensorFlow, and ONNX
  • Bow IPU and Bow Pod large-scale systems
  • Massively Parallel MIMD Architecture
  • In-Processor Memory
  • IPU-Fabric for high-speed interconnectivity

Integrations

PyTorchTensorFlowONNXHugging FaceKubernetesDockerPaperspaceGcoreCirrascale

Target audience

AI researchers, machine learning engineers, data scientists, and large enterprises working on cutting-edge AI model development, high-performance computing (HPC), and large-scale model training.


Ratings & Reviews

0.0

Based on 0 reviews

Key Metrics

Founded

2016

Headquarters

Bristol, United Kingdom

Pricing Tiers

Cloud Access (Free Tier)

Provided via partners like Paperspace, offering free, time-limited access to IPUs within a notebook environment for experimentation, learning, and small-scale projects.

Free

Cloud Access (On-Demand)

Paid hourly or monthly access to dedicated IPU instances and PODs through cloud partners. Pricing varies by provider, region, and the size of the required resources.

Varies

Direct Purchase (Enterprise)

Purchase of Bow Pod systems and other hardware directly from Graphcore. This option is intended for large-scale enterprise, government, or research deployments with specific a TCO model.

Custom Quote


Frequently Asked Questions


Top Alternatives to Graphcore

NVIDIA (Tesla/Hopper GPUs)

NVIDIA is the market leader, and its GPUs with the CUDA platform offer the broadest ecosystem, largest community, and strongest performance across a wide range of common AI workloads.

Google TPU

Available exclusively on Google Cloud, TPUs are highly optimized for TensorFlow and JAX, making them a strong choice for users already embedded in the Google Cloud ecosystem.

Cerebras Systems

Cerebras offers a truly unique wafer-scale engine for organizations needing to train the absolute largest models that can't be easily distributed across smaller chips.

SambaNova Systems

SambaNova provides a full-stack, subscription-based AI platform aimed at enterprises that want a managed hardware and software solution for deploying a variety of models.

Ready to get started?

Join thousands of users and see how Graphcore can transform your workflow today.

Visit Graphcore