It has the same physical memory as the consumer cards, it just allows in-band ECC which is a software solution. There is no additional manufacturing cost to including in-band ECC. Nvidia even enables the option on the 4090.
You’re right that you’re paying for the certifications though.
The impressive part is that it has IPC on par with Raptor Cove, what’s holding it to Skylake performance is the low max clocks.
It literally has higher IPC than Zen4, which is an impressive showing for Loongson. I’m guessing it’s a very wide design that has to stay at low clocks because it is hampered by the relatively ancient 14nm process it’s manufactured on.
You’ve gotta give credit where credit is due.