Nvidia’s Hopper data center GPU

Nvidia introduced their long-anticipated Hopper GPU with startling compute results—the chip is actually more of a compute engine than a GPU per se. Nvidia is claiming a 6X improvement over the previous-generation Ampere. Of that improvement rate, 2X comes from getting the chip to do FP8 calculations for inferencing. Another 2X comes from improvements in Cuda, and 1.3X comes from increased clock frequencies, while 1.2X comes from adding SMs. Right, it doesn’t add up to 6X, but there’s some kind of multiplier effect in the calculations.  H100 compute improvement summary. (Source: Nvidia)   The generation-to-generation numbers are impressive and interesting.