Three revolutionary technologies related to “Blackwell” will be introduced in NVIDIA's “Hot Chips 2024” lecture.
1. Dramatic improvements in cooling efficiency with hybrid liquid cooling solutions:
Improving the efficiency of cooling systems is essential for improving the performance of supercomputers and AI data centers. Blackwell is equipped with a hybrid liquid cooling solution that greatly evolves conventional cooling methods, and achieves unprecedented high cooling efficiency. It is expected that this technology will enable higher performance computing.
2. FP4 technology for 20 PFLOPS of AI processing performance:
By adopting revolutionary FP4 (4-bit floating point arithmetic) technology, Blackwell dramatically improves AI processing performance and achieves astonishing performance of 20 PFLOPS. The foundation for this is the “NVIDIA Quasar Quantization System” newly developed by NVIDIA. Since FP4 technology can drastically reduce the amount of computation while maintaining the accuracy of AI models, it has the potential to accelerate the spread of AI.
3. Using LLM to Streamline Chip Design:
In Blackwell's chip design, a revolutionary test is being carried out to improve the efficiency of design work by applying LLM (large-scale language model) to the coding of the hardware description language Verilog. This initiative is expected to greatly improve chip design productivity and accelerate the development of new products.
Improving the efficiency of cooling systems is essential for improving the performance of supercomputers and AI data centers. Blackwell is equipped with a hybrid liquid cooling solution that greatly evolves conventional cooling methods, and achieves unprecedented high cooling efficiency. It is expected that this technology will enable higher performance computing.
2. FP4 technology for 20 PFLOPS of AI processing performance:
By adopting revolutionary FP4 (4-bit floating point arithmetic) technology, Blackwell dramatically improves AI processing performance and achieves astonishing performance of 20 PFLOPS. The foundation for this is the “NVIDIA Quasar Quantization System” newly developed by NVIDIA. Since FP4 technology can drastically reduce the amount of computation while maintaining the accuracy of AI models, it has the potential to accelerate the spread of AI.
3. Using LLM to Streamline Chip Design:
In Blackwell's chip design, a revolutionary test is being carried out to improve the efficiency of design work by applying LLM (large-scale language model) to the coding of the hardware description language Verilog. This initiative is expected to greatly improve chip design productivity and accelerate the development of new products.
Disclaimer: Community is offered by Moomoo Technologies Inc. and is for educational purposes only.
Read more
Comment
Sign in to post a comment
Kimihiko OP : NVIDIA's hybrid liquid cooling solution is a cooling system that combines the advantages of both air cooling and liquid cooling. Conventional air cooling has limitations on cooling performance, and there is a problem that it is difficult to sufficiently dissipate heat from high-performance chips. On the other hand, although liquid cooling exhibits high cooling performance, there was a disadvantage that the system became complicated and the cost also increased.
In order to solve these problems, the hybrid liquid cooling solution uses high-performance liquid cooling for the part that directly comes into contact with the chip, and air cooling is adopted for the part that cools the entire system. Thus, system complexity and cost increases can be suppressed while greatly improving cooling performance.