A month ago Nvidia introduced the H100 accelerator on the new Hopper architecture. This is a solution with 80 GB of memory, 18,432 CUDA cores (of which 15,872 are active) and a performance of 60 TFLOPS.
Of course, this is, firstly, a specialized accelerator for servers and supercomputers, and secondly, this is the latest top-level solution. And such a product cannot cost a little – and it is not worth it.
In Japan, they ask for a novelty of 36,445 dollars, including taxes. It is difficult to say how expensive it is, given the specificity of the device. It is not yet possible to compare prices, since no one else has added a card to their assortment yet. And the reason is simple: Nvidia H100 has not yet entered the market – this will happen only in the second half of the year.
However, it is still possible to compare prices indirectly. The same store previously added an Instinct MI210 accelerator to the site, priced at $16,500. Yes, Nvidia’s solution is more productive, but not twice as much. In particular, the performance with single precision (FP32) of the new AMD is 45.3 TFLOPS, and the PCIe 5.0 H100 is 48 TFLOPS.