If your application requires memory-constrained inference, this card will be more cost-effective than the H100 and H200. The supported FP8 configuration allows direct deployment of FP8-optimized Deepseek V3 and Deepseek R1 models, unlike the A100 graphics card which is limited to FP16.