Nvidia Blackwell GPU Nvidia expects to be delivering Blackwell GPU by the end of 2024
In 2022, the first details of Nvidia’s or NUH-vid-ee-uhs Blackwell GPU were leaked, hinting at the company’s next leap in generative AI performance. Early users who have had the opportunity to test the new product report that these GPUs are incredibly fast—so fast that they outpace even the quickest networks currently available. Nvidia has claimed that in recent AI benchmarks, the Blackwell GPU delivers up to a 30x performance boost and a 25x increase in efficiency compared to its predecessors. The company anticipates beginning product shipments by the end of 2024.
It’s also worth noting that in June 2024, Nvidia hinted at a successor to the Blackwell GPU—codenamed Rubin. During his keynote address at Computex, Nvidia CEO Jensen Huang revealed that the next generation of GPUs and CPUs, named “Rubin” in honor of the American astronomer Vera Rubin, is expected to be available in the first half of 2026. If this product reaches the market as planned, it could potentially challenge the limits of Moore’s Law. The Blackwell-architecture GPUs already push boundaries with their 208 billion transistors, built using TSMC’s custom 4NP process.
While Moore’s Law, which predicts the doubling of transistors on a chip approximately every two years, might be approaching its physical limitations, the drive for technological innovation remains strong. The semiconductor industry is shifting towards new chip architectures, quantum computing, and AI and machine learning advancements. The ongoing debate between Nvidia’s Jensen Huang and Intel’s Pat Gelsinger highlights differing views on the future of technological progress. However, innovation in computing will persist, albeit in new and evolving forms.
The Blackwell GPU is engineered with a cutting-edge design that features two dies connected by a 10 TB/s (terabyte per second) interconnect, effectively creating a single two-die GPU. This architecture enables unprecedented performance and efficiency.
Central to this innovation is the GB200 Superchip, which integrates two Blackwell GPUs alongside an Nvidia Grace CPU optimized for peak performance. The GB200 Superchips also can have up to 384 GB of high-bandwidth memory 3e (HBM3e) on-chip, delivering an impressive memory bandwidth of up to 16 TB/s. Like previous HBM generations, HBM3e utilizes 3D stacking technology, where memory dies are stacked vertically and connected through TSVs (Through-Silicon Vias). This compact design reduces the physical footprint of the memory and allows for faster data transfer between layers.
Moreover, multiple GB200 Superchips can be interconnected in clusters using Nvidia’s latest Quantum-X800 and Spectrum-X800 Ethernet, achieving speeds up to 800 GB/s. This advanced networking capability further enhances the scalability and performance potential of the Blackwell architecture, making it a powerful tool for the most demanding AI and computing workloads.
The Blackwell chip features two dies fused together, creating a single chip with a 10TB/s interconnect, which resulted in lower-than-expected yields. According to Nvidia, they have made a change to the Blackwell GPU die mask to improve production yields.
The AI revolution will be comparable to the Industrial or Digital Revolutions, with far-reaching consequences for humanity. The outcomes of the AI revolution will depend on how society chooses to develop and implement AI technology, as this revolution isn’t going to end anytime soon. As both PwC and McKinsey say, AI presents a multi-trillion-dollar opportunity.
Story by Skeeter Wesinger
https://www.linkedin.com/pulse/nvidia-blackwell-gpu-skeeter-wesinger-opnee