Nvidia unveiled its next-generation Blackwell graphics processing units (GPUs), which deliver 25x higher power consumption and lower cost for AI processing tasks.
The Nvidia GB200 Grace Blackwell Superchip – meaning it consists of multiple chips in the identical package – guarantees exceptional performance gains, offering as much as 30x performance improvement for LLM inference workloads in comparison with previous iterations.
Speaking at Nvidia Terms and Conditions 2024 In a keynote speech, Nvidia CEO Jensen Huang introduced Blackwell to 1000’s of engineers, saying it can usher in a transformative era in computing. Gaming products are expected to launch later.
During the keynote, Huang joked that the prototypes he was holding were price between $10 billion and $5 billion. The chips were a part of the Grace Blackwell system.
“For three many years, we’ve got pursued accelerated computing with the goal of enabling transformative breakthroughs resembling deep learning and AI,” said Huang. “Generative AI is the defining technology of our time. Blackwell GPUs are driving this recent industrial revolution. By collaborating with the world’s most dynamic corporations, we’ll realize the promise of AI for each industry.”
Giving details of the improvements, Nvidia said Blackwell-based computers will enable corporations all over the world to construct and run generative AI in real-time on large language models with trillions of parameters at 25 times lower cost and energy consumption than its predecessor, Hopper. Processing is scaled to AI models with as much as 10 trillion parameters.
These numbers are essential because Nvidia faces low-end competitors like chip designer Groq and high-end CPU vendors like Cerebras — not to say AMD and Intel. Groq is a Mountain View, California-based rival that focuses on chips for inference somewhat than AI training.
Named after mathematician David Harold Blackwell, the primary Black scientist inducted into the National Academy of Sciences, the Blackwell platform succeeds the Nvidia Hopper GPU architecture and sets recent standards in accelerated computing.
Originally intended for gaming graphics, GPUs are the essential engine of AI processing, and that has helped Nvidia achieve a market cap of $2.2 trillion. That's also why people like CNBC's Jim Cramer are broadcasting live from Nvidia GTC, which has a geeky past.
The platform introduces six breakthrough technologies that may enable groundbreaking advances in diverse areas, including data processing, engineering simulation, electronic design automation, computer-aided drug design, quantum computing and generative AI.
The strongest chip on the planet
Huang said Blackwell will likely be essentially the most powerful chip on the planet. The Blackwell architecture GPUs feature 208 billion transistors and are manufactured using a custom 4NP TSMC process with dual reticle limits, enabling high computing performance.
Blackwell encompasses a second generation transformer motor. Equipped with recent microtensor scaling support and advanced dynamic range management algorithms, the Transformer Engine doubles compute and model size with progressive 4-bit floating point AI inference capabilities.
Nvidia also introduced fifth-generation NVLink networking technology. The latest NVLink iteration improves the performance of multi-trillion-parameter AI models and delivers breakthrough bidirectional throughput per GPU, promoting seamless, high-speed communications.
NVLink delivers 1.8 TB/s bidirectional throughput per GPU, ensuring seamless, high-speed communications between as much as 576 GPUs for today's most complex LLMs.
And it has a RAS engine. To ensure reliability, availability and serviceability, Blackwell GPUs integrate dedicated engines and AI-based preventive maintenance capabilities to maximise system availability and minimize operating costs.
It also has a secure AI solution. Advanced confidential data processing capabilities protect AI models and customer data without compromising performance, targeting privacy-sensitive industries.
The dedicated decompression engine accelerates database queries by supporting the most recent formats and improves data evaluation and data science performance, revolutionizing data processing.
This superchip forms the cornerstone of the NVIDIA GB200 NVL72, a rack system with 1.4 exaflops of AI performance and 30 TB of fast storage.
With widespread adoption expected amongst major cloud providers, server manufacturers and leading AI corporations including Amazon, Google, Meta, Microsoft and OpenAI, the Blackwell platform is poised to revolutionize computing across industries.
Blackwell will goal computing customers in data processing, engineering simulation, electronic design automation, computer-aided drug design, quantum computing and generative AI – all recent industry opportunities for Nvidia.
The Nvidia GB200 Grace Blackwell Superchip connects two Nvidia B200 Tensor Core GPUs with the Nvidia Grace CPU via an ultra-low power, 900 GB/s chip-to-chip connection. The GB200 Superchip delivers as much as 30x performance improvement over the Nvidia H100 Tensor Core GPU for LLM inference workloads and reduces cost and energy consumption by as much as 25x.
The GB200 is a key component of the Nvidia GB200 NVL72, a liquid-cooled, multi-node rack system for essentially the most compute-intensive workloads. It combines 36 Grace Blackwell superchips, including 72 Blackwell GPUs and 36 Grace CPUs, connected via fifth-generation NVLink.
Additionally, GB200 NVL72 includes Nvidia BlueField-3 compute units to enable cloud network acceleration, composable storage, zero trust security, and GPU compute elasticity in hyperscale AI clouds. The platform acts as a single GPU with 1.4 exaflops of AI performance and 30TB of fast storage and is a constructing block for the most recent DGX SuperPOD.
Nvidia offers the HGX B200, a server board that connects eight B200 GPUs over high-speed interconnects to support the world's strongest x86-based generative AI platforms. HGX B200 supports network speeds of as much as 400 Gbps over the Nvidia Quantum-2 InfiniBand and Spectrum-X Ethernet networks
Platforms.
GB200 can even be available on the Nvidia DGX Cloud, an AI platform co-developed with leading cloud service providers that gives enterprise developers with dedicated access to the infrastructure and software vital to construct and deploy advanced generative AI models. AWS, Google Cloud and Oracle Cloud Infrastructure plan to host recent Nvidia Grace Blackwell-based instances later this 12 months.
Cisco, Dell Technologies, Hewlett Packard Enterprise, Lenovo and Supermicro are expected to produce a wide selection of servers based on Blackwell products, as are Aivres, ASRock Rack, ASUS, Eviden, Foxconn, Gigabyte, Inventec, Pegatron, QCT, Wistron and Wiwynn and ZT Systems.
