Linear gyrokinetic simulation is an important method for studying microinstabilities in tokamakplasmas, but complete calculations can become computationally demanding when many eigenvaluesor finely resolved structures are required. In this work, we present an orbit-invariant decompositionmethod for collisionless gyrokinetic eigenvalue problems. By discretizing velocity space along orbitinvariants such as particle energy and magnetic moment, the full eigenvalue matrix separates intoindependent orbit matrices that couple only through the field equation, reducing both matrix dimensionand computational cost. Based on this method, we extend the MGK code [Phys. Plasmas24, 072106 (2017)] with both CPU and GPU implementations, supporting collisionless electrostaticlinear simulations in s–α and Miller geometries with kinetic or adiabatic electrons. For kineticion temperature gradient (ITG) and trapped electron mode (TEM) eigenvalue problems, the solverreduces single-solution times to the 0.01–0.1s range—more than three orders of magnitude fasterthan CGYRO on the same hardware—enabling efficient large-scale parameter scans. The eigenfrequenciesand mode structures are validated against CGYRO benchmarks. The method is applicableto all collisionless gyrokinetic eigenvalue formulations and is extensible to fully electromagneticsimulations.