It used to be 32-bit GFLOPS was the metric to use to judge the performance (and, presumably, the price or value) of a supercomputer.
As computers evolved and got more powerful, the FLOPS calculation got extended to 64 bits, and speed measurements of integer processing also got added, so we had INT32, FP32, and FP64 performance measurements, and often different machines did better in one test than another, resulting in multiple first-place winners each year.
A few years ago (2006), the Green500 list, which ranks the most energy-efficient supercomputers, was introduced, doubling the number of categories in which a supercomputer could be rated. So now there were potentially six possible winners in the annual supercomputer roundup fashion parade. It’s kind of like an elementary school soccer match—everyone gets a ribbon.
This year, yet another winner category has been added— the fastest AI system in the world dedicated to AI. Presumably, that is AI training, but why stop there? When can we categorize and compare supercomputers on their inference speed? Of course, we’ll have to have a few categories such as INT8, INT16, FP8, and FP16, and maybe, in the future, INT4—why not?
So, next year at the super-duper supercomputer show in the big tent (you’ll be pleased to know they are no longer allowing elephants in the tent), there will be the possibility of having a dozen or more categories of the fastest, most best-est, supercomputers.
But I think there are still categories missing. What about footprint and weight? What does a water-filled super-duper supercomputer weigh? Now that doesn’t affect performance unless one of the pipes gets clogged and the machine gets too hot. Footprint would go into the price-performance evaluation, the cost of real estate.
In any case, we’re looking forward to seeing who the winners are this year, and you can count on us to give it our full and objective attention, fleeting as it may be.