This work presents game semantics of Martin-Löf type theory (MLTT) equipped with the One, the Zero, the N, Pi, Sigma and Id types. Game semantics interprets a wide range of logic and computation, even the polymorphic $\lambda$-calculus; however, it has remained a well-known challenge in the past 25 years to achieve game semantics of dependent type theories such as MLTT, and past attempts lack directness or generality. For instance, the approach taken by Abramsky et al. interprets Sigma types indirectly by formal lists, not by games, making an interpretation of universes hopeless, and it is limited to a very specific class of dependent types. The difficulty of this challenge comes from a conflict between the extensionality of dependent types and the intensionality of game semantics. We overcome the challenge by inventing a novel variant of games, while we keep strategies unchanged, in such a way that this variant inherits the strong points of conventional game semantics. Also, our method enables an interpretation of subtyping on dependent types for the first time as game semantics. We finally give a new, game-semantic proof of the independence of Markov’s principle from MLTT. This proof illustrates an advantage of our intensional model over extensional ones such as Hyland’s effective topos.