Core Viewpoint - Nvidia has faced criticism for neglecting the 64-bit performance needed for traditional modeling and simulation workloads, focusing instead on enhancing low-precision computing for AI applications. However, the company asserts it has not abandoned 64-bit computing and plans to improve its performance in future GPU generations [2][5][6]. Group 1: Performance Comparison - The transition from Nvidia's Hopper to Blackwell architecture did not yield significant improvements in FP64 performance, as highlighted by Jack Dongarra during the SC25 conference. He noted that the floating-point computing capability did not surpass the previous generation [2]. - Nvidia's Hopper H100 and H200 GPUs offer 340 teraflops of FP64 performance and 670 teraflops of FP64 Tensor Core performance, a substantial increase from the Ampere A100's 97 teraflops and 195 teraflops, respectively [2]. - The Blackwell architecture's B100 GPU was initially expected to have lower 64-bit computing capabilities, with FP64 and FP64 Tensor Core performance at only 30 teraflops. Ultimately, Nvidia released the B200 and GB200 Grace Blackwell "super chip," which still does not match the performance of the H200 [3][4]. Group 2: Market Response and Future Directions - Nvidia's Blackwell chips are primarily designed for low-precision AI workloads, responding to the increasing computational demands for training and running large language models. This focus has led to strong sales, making Nvidia the first company to surpass a market capitalization of $5 trillion [5]. - Despite the advancements in AI, the HPC community feels overlooked, as traditional FP64 computing remains essential for fields like materials science and climate modeling. The demand for high bandwidth memory is driven by AI needs rather than HPC requirements [5][6]. - Addison Snell from Intersect360 Research emphasized the critical importance of FP64 computing across various industries, including manufacturing, energy, finance, and healthcare. He noted that 64-bit computing should be considered a fundamental requirement for "scientific AI" [6]. Group 3: Future Enhancements - Nvidia's senior director, Dion Harris, acknowledged that while Blackwell's 64-bit performance is not as strong as Hopper's, the company remains committed to maintaining its leadership in this area. He mentioned the release of cuBLAS, a CUDA-X math library that can enhance FP64 matrix multiplication performance by 1.8 times [6][8]. - Harris indicated that Nvidia is working on improving the core underlying performance of future GPUs in 64-bit computing, although specific details were not disclosed. The HPC market is eager for significant enhancements in FP64 performance similar to those seen in previous generations [9].