Nvidia's recently unveiled Blackwell accelerators demonstrate a substantial leap in performance, achieving benchmarks up to 2.2x higher than the previous H100 models during real-world training workloads.
The DGX B200 systems utilized in the Nyx supercomputer showcase an impressive 2.27x increase in peak floating point performance across various precisions compared to the H100, highlighting the efficiency advancements of Blackwell.
What sets Blackwell apart isn't merely its FLOPS, but also the significant memory bandwidth advantage, offering up to 8 TBps, providing a notable enhancement in performance metrics in benchmarks.
Even with just 64 Blackwell GPUs tested in the GPT-3 benchmark, the performance achieved was equivalent to that of 256 H100 GPUs, emphasizing the efficiency and capabilities of the new architecture.
Collection
[
|
...
]