Nvidia believes that every business that generative AI affects has the potential to be revolutionized; all that’s needed is the technology to match the challenge.
Aiming to empower the next generation of AI technologies for enterprises, the American chipmaker unveiled its advanced Blackwell computing platform at the firm’s annual GTC conference in San Jose, California, on March 18.
The new architecture is named after David Harold Blackwell, a mathematician at the University of California, Berkeley, specializing in statistics and game theory.
The firm claims the platform allows organizations to deploy real-time generative AI on trillion-parameter language models with 25 times less cost and energy than before.
“Generative AI is the defining technology of our time. Blackwell is the engine to power this new industrial revolution. Working with the most dynamic companies in the world, we will realize the promise of AI for every industry.” said Jensen Huang, CEO and founder of Nvidia, in a statement.
According to the firm, six revolutionary accelerated computing technologies found in the Blackwell GPU architecture will help NVIDIA take advantage of new business opportunities in data processing, engineering simulation, electronic design automation, computer-aided drug design, quantum computing, and generative artificial intelligence.
Advanced computing capacity
Powered by a custom-built 4NP TSMC process, Blackwell GPUs boast 208 billion transistors connected via 10TB/s chip-to-chip links.
Enhanced with micro-tensor scaling and advanced dynamic range management, Blackwell supports double the compute and model sizes, alongside 4-bit floating point AI inference. According to Nvidia, the latest NVLink offers 1.8TB/s bidirectional throughput per GPU, enabling seamless communication among up to 576 GPUs for complex LLMs.
Blackwell-powered GPUs feature a dedicated engine for reliability, availability, and serviceability, with added chip-level capabilities for AI-based preventative maintenance. This enhances system uptime and resilience for large-scale AI deployments, reducing operating costs.
Advanced confidential computing safeguards AI models and customer data while maintaining performance, including support for encryption protocols vital in sectors like healthcare and finance.
A dedicated decompression engine also accelerates database queries, ensuring top data analytics and science performance. GPU-accelerated data processing will become increasingly prevalent, revolutionizing industries spending billions annually.
Nvidia claims that its new Blackwell chip outperforms its predecessor by 2.5 times in FP8 for training and 5 times in FP4 for inference. It has a fifth-generation NVLink interface that can support up to 576 GPUs and is twice as fast as Hopper.
Additionally, two Blackwell NVIDIA B200 Tensor Core GPUs are connected to the NVIDIA Grace CPU via the NVIDIA GB200 Grace Blackwell Superchip via a 900GB/s ultra-low-power NVLink chip-to-chip interface.
Huang displayed a board containing the system. According to him, with this much computation packed into such a compact size, this computer is the first of its type. “They feel like it’s one big happy family working on one application together because this is memory coherent,” said Huang, during the event.
Nvidia’s new Quantum-X800 InfiniBand and Spectrum-X800 Ethernet platforms, also released during the event, can be used to connect GB200-powered systems for the best AI performance. These platforms offer sophisticated networking at up to 800Gb/s.
Nvidia’s Exaflop AI revolution
Nvidia created a brand-new chip called NVLink Switch in order to scale up Blackwell. With 1.8 terabytes per second, each may connect four NVLink interconnects and perform in-network reduction to reduce traffic.
The Nvidia GB200 NVL72, which Huang called “one giant GPU,” is a multi-node, liquid-cooled, rack-scale system that uses Blackwell to provide supercharged compute for trillion-parameter models.
It includes 36 Grace Blackwell Superchips, linked by fifth-generation NVLink, including 72 Blackwell GPUs and 36 Grace CPUs. NVIDIA BlueField-3 data processing units are also included in the GB200 NVL72 to support GPU compute elasticity in hyperscale AI clouds, composable storage, zero-trust security, and cloud network acceleration.
For LLM inference applications, the GB200 NVL72 offers up to a 30x performance boost and up to a 25x reduction in cost and energy consumption when compared to the same number of NVIDIA H100 Tensor Core GPUs.
To further expand, Nvidia also unveiled the Nvidia DGX SuperPOD, a next-generation AI supercomputer powered by Nvidia GB200 Grace Blackwell Superchips that can handle trillion-parameter models for superscale generative AI training and inference workloads with continuous uptime.
According to the firm, with Nvidia DG GB200 processors at its core, the new DGX SuperPOD has a revolutionary, extremely efficient liquid-cooled rack-scale design that can process 11.5 exaflops of AI supercomputing at FP4 precision and 240 terabytes of fast memory, with more capacity available with additional racks.
“In the future, data centers are going to be thought of … as AI factories. Their goal in life is to generate revenues, in this case, intelligence,” said Huang.
Nvidia claims that the world’s leading cloud service providers, innovative AI firms, system and server suppliers, local cloud service providers, and telecoms are set to embrace Blackwell for their businesses.
RECOMMENDED ARTICLES
0COMMENT
NEWSLETTER
The Blueprint Daily
Stay up-to-date on engineering, tech, space, and science news with The Blueprint.
Jijo Malayil Jijo is an automotive and business journalist based in India. Armed with a BA in History (Honors) from St. Stephen's College, Delhi University, and a PG diploma in Journalism from the Indian Institute of Mass Communication, Delhi, he has worked for news agencies, national newspapers, and automotive magazines. In his spare time, he likes to go off-roading, engage in political discourse, travel, and teach languages.
0