Ahead of the Hot Chips 2024 tradeshow, Nvidia provided a glimpse into its Blackwell platform, showcasing servers being installed and set up. This was Nvidia’s way of reinforcing that Blackwell is still on track, despite reported delays. The company also highlighted its existing Hopper H200 solutions, FP4 LLM optimizations through the new Quasar Quantization System, warm water liquid cooling for data centers, and the use of AI to design advanced chips for future AI applications.
Nvidia’s roadmap for the upcoming years is becoming clearer. Blackwell Ultra is set to debut next year, followed by Vera CPUs and Rubin GPUs in 2026, and Vera Ultra in 2027. Nvidia had previously shared these plans at Computex in June, and AI continues to be a focal point for the company. Although there were reports of a three-month delay for Blackwell, Nvidia remained tight-lipped on the matter, instead opting to share images of Blackwell systems being installed and hardware components inside Blackwell GB200 racks and NVLink switches.
The company also released performance data for the existing H200 system, demonstrating up to a 1.5X improvement in inference workloads using a Llama 3.1 70B parameter model with NVSwitch. Blackwell will further double NVLink bandwidth, with the NVLink Switch Tray providing an aggregate 14.4 TB/s of bandwidth.
As data center power demands grow, Nvidia is collaborating with partners to enhance both performance and efficiency. One promising development is the use of warm water cooling, which not only reduces data center power usage by up to 28% but also offers the potential to repurpose heated water for additional cost savings. Nvidia’s latest advancements, including native FP4 support in Blackwell, ensure that its software maximizes the benefits of the new hardware features without compromising accuracy.
Nvidia also discussed its internal use of AI tools to design better chips, utilizing an LLM to accelerate design, debugging, analysis, and optimization processes. This approach was instrumental in developing the 208 billion transistor Blackwell B200 GPU and will be key in future projects like the Rubin GPUs.
Nvidia will be sharing more about the Blackwell architecture, generative AI in computer-aided engineering, and liquid cooling at the Hot Chips conference next week.
