Check out Electronic Design's coverage of GTC 2024.
What you’ll learn:
- What is the Blackwell GPU?
- Why is it optimized for 4-bit floating point?
NVIDIA's CEO, Jensen Huang, highlights the NVIDIA Blackwell GPU in his keynote speech (watch the video above) at this year's GTC AI Conference. Slated for release in 2024, the NVIDIA Blackwell GPU (see figure) represents a leap forward in graphics and computational technology, poised to significantly enhance artificial-intelligence (AI) computing. A performance factor of 2X to 4X over the previous series is guaranteed.
The name Blackwell is inspired by, and named after, the renowned mathematician David Harold Blackwell, reflecting NVIDIA's tradition of honoring notable scientists in its naming conventions.
The Blackwell GPU is expected to introduce a chiplet design for the first time in NVIDIA's lineup, a move that could drastically increase its computational power and efficiency. The system consists of a pair of die connected by a 10-TB/s, memory-coherent, chip-to-chip interconnect.
The Blackwell Tensor Core micro-tensor scaling support is optimized for large language models (LLMs) and mixture-of-experts (MoE) models. It supports 4-bit floating point (FP4), allowing for faster execution and smaller data requirements while maintaining a high level of accuracy.
Rumors suggest that it might employ the advanced TSMC 3-nm manufacturing process, indicating a significant boost in performance and energy efficiency compared to previous generations. The keynote spoke to the power savings vs. computational output compared to the previous Hopper series. It does appear to be the case.
With potential performance gains projected to be more than double that of its predecessors, the Blackwell B100 GPU is eagerly anticipated for its potential to push the boundaries of what's possible in AI research, data analysis, and complex computational tasks. Dozens of NVIDIA partners are eagerly awaiting the Blackwell release.
Check out more of our coverage of GTC 2024.