Habana Labs develops Intel® Gaudi® AI accelerators designed for high-performance deep learning training and inference, providing enterprises and cloud providers with efficient compute solutions. Their technology delivers up to 40% better price/performance on cloud instances, addressing the need for cost-effective and scalable AI infrastructure.
Funding
$75M raised to dateRaised to date based on public sources. This may differ from the amount the company actually raised and is based only on what is publicly available on the internet.
Founders
Product
Problem
Enterprises and cloud providers face increasing demands for high-performance, cost-effective, and scalable AI compute solutions to handle deep learning training and inference workloads, especially with the rise of generative AI. Existing solutions often lack the balance of performance, efficiency, and scalability required to meet these demands effectively.
Solution
Intel Gaudi AI accelerators provide a purpose-built solution for deep learning, offering a compelling alternative to traditional GPUs. The Gaudi architecture is designed to deliver competitive performance and efficiency for both training and inference, with a focus on price-performance. Gaudi accelerators integrate a large number of high-bandwidth Ethernet ports, enabling massive and flexible system scaling for demanding AI workloads. The Intel Gaudi Software suite simplifies model migration and development, allowing users to leverage existing frameworks and models with minimal code changes.
Target Audience
The primary target audience includes enterprises, cloud providers, and AI researchers who require high-performance, scalable, and cost-efficient AI compute infrastructure for deep learning training and inference.
Features
- Heterogeneous compute architecture with Matrix Multiplication Engines (MME) and Tensor Processor Cores (TPC)
- Support for various data types, including FP32, TF32, BF16, FP16, and FP8
- Integrated high-bandwidth memory (HBM) for fast data access
- High-speed Ethernet ports (200 GbE) for scalable multi-node configurations
- Optimized software stack with support for PyTorch and TensorFlow frameworks
- Habana Optimum Library for easy access to pre-trained models on Hugging Face
- Support for both air-cooled and liquid-cooled deployment options