Nvidia has launched the Blackwell platform, a new computing architecture designed to support the development and execution of generative artificial intelligence (AI) on large language models. The platform, which aims to reduce expenses and energy usage by up to 25 times compared to its predecessor, is expected to cater to various sectors including data processing, and engineering simulation, says Nvidia.
Key Features of the Blackwell Platform:
- High-Performance GPU: The GPU is equipped with 208 billion transistors and utilizes a 4NP TSMC manufacturing process. It includes two GPU dies linked by a 10 TB/second chip-to-chip connection, forming a single GPU unit.
- Second-Generation Transformer Engine: This engine is designed to enhance computational efficiency and model size handling, supporting new 4-bit floating point AI inference.
- High-Speed NVLink: The fifth iteration of Nvidia’s NVLink provides 1.8TB/s bidirectional throughput per GPU, facilitating communication among up to 576 GPUs for complex LLMs.
- Reliability and Serviceability Engine (RAS): This engine is focused on ensuring system reliability and uptime, incorporating AI-based preventative maintenance for diagnostics and reliability forecasting.
- Secure AI: The platform includes features for confidential computing, protecting AI models and customer data, and supports new encryption protocols for secure data handling.
- Decompression Engine: Dedicated to improving the performance of database queries, this engine is designed to accelerate data analytics and science by supporting the latest decompression formats.
Organizations such as Amazon Web Services, Dell Technologies, Google, Meta, Microsoft, OpenAI, Oracle, Tesla, and xAI are expected to adopt the Blackwell platform, says Nvidia.
Blackwell-powered products will be available from various partners, including cloud service providers and server manufacturers. Additionally, software makers in engineering simulation are set to use Blackwell processors to enhance their applications, according to Nvidia.
[Image courtesy: Nvidia]