Nvidia's GPU performs quadrillions of calculations per second.

Judge as of 2024-05-02, for the Nvidia GPU he was referring to (a current data-centre GPU such as H100/H200 or the announced Blackwell B200/GB200; if the context is unclear, Nvidia's flagship data-centre GPU at that date). 'Quadrillions' means at least 10^15 operations per second (petaFLOPS/petaOPS or more) at some precision Nvidia publishes. True: Nvidia's published specifications show the GPU reaching one or more petaFLOPS/petaOPS (for example in FP8, FP16 or with sparsity). Mostly true: the figure reaches the quadrillion range only at reduced precision or with sparsity, while dense FP32/FP64 throughput stays in the trillions. Mostly false: the referenced GPU reaches about 10^15 only for a different product, or only as a multi-GPU system figure presented as one GPU's. False: the GPU's published peak throughput at every precision is below 10^15 operations per second.

Set before the research began.

1. E1 NVIDIA Announces Hopper Architecture, the Next Generation of Accelerated Computing NVIDIA · 22 March 2022 Own record

   “NVIDIA's fourth-generation DGX™ system, DGX H100, features eight H100 GPUs to deliver 32 petaflops of AI performance at new FP8 precision”

In Nvidia's official Hopper announcement [E2], the DGX H100 system uses eight H100 GPUs to deliver 32 petaflops of AI performance at FP8 precision. That works out to about 4 petaflops per H100 GPU, which is above the 10^15 operations per second threshold at a precision Nvidia publishes. The criteria give FP8 as an example of a qualifying precision. The evidence does not show dense FP32 or FP64 throughput, so it cannot support the "mostly true" reduced-precision distinction. The only source is the company's own announcement.