Of course, an RDBMS could be designed to do that without a performance penalty, by storing data in a denormalized form and automatically translating queries for the normalized data accordingly.

But SQL doesn't have the features you'd need to control and manage that sort of transparent denormalization. So you'd end up having to extend SQL to support it properly so that the performance penalty in question could be mitigated in all cases.

edit: Rather than "you need to fully normalize everything," I should have said "you need to split all your data across multiple tables to eliminate the need for structured data within records." The performance penalty happens when you need to do this everywhere for sufficiently complex datasets.