Hi Adam,
I'm a few episodes behind, but episode 1673 has a few clips discussing the NVIDIA GB-200 superchip. The clip has somebody claiming that the new Blackwell chip will be great for scientific computing, especially CFD.
Scientific computing still almost exclusively uses FP64 (double precision floating point) for calculations. The Blackwell chip has FP64 performance of 90 TFLOPs, which is only a 25% increase over the H100 GPU. The memory bandwidth of the blackwell chip is double that of the H100, which can have a large impact on performance. However, that amount of increase is not unexpected for a new generation of the hardware, and isn't better than what the NVIDIA competitors are offering.
Maybe the person in the video was excited about the NVL72 system (consisting of GB200 superchips), but even that isn't particularly novel.
As somebody who has been deeply ingrained in optimizing performance of scientific computing codes for some of the largest and fastest High Performance Computing systems in the world, I can confidently say that those of us in HPC who actually understand how hardware impacts application performance are pretty underwhelmed by the utility of the newest NVIDIA products for scientific computing. NVIDIA seems to have focused on AI and basically ignored the needs of scientific computing.
Not to mention the fact that the cost of NVIDIA hardware is so high that the performance per dollar of NVIDIA for scientific computing isn't very good.
Many people still purchase new NVIDIA hardware for scientific computing, but that is because they have already ported their models to CUDA and don't want to invest the resources porting the software to other alternatives.
---
Some info about me to show I'm not some random blowhard.
- I have been doing performance optimization for scientific computing codes for 15 years.
- I have worked for and optimized scientific computing models for DoD, DOE, Academia, and industry.
- I led a team that was part of the design and delivery of the Frontier supercomputer at Oak Ridge National Lab (the #1 computer on the Top500 list, and the first to break the exascale barrier).
- I've optimized scientific computing models for many fields
- Weather forecasting
- Climate modeling
- Nuclear Physics
- CFD
- Acoustics
- And more...
- I currently work for a company that is trying to innovate new technology specifically for scientific computing at large scales
Please let me know if you have any questions.
Thanks,
Matt