The design of novel steam cracking reactor geometries has received considerable attention recently. To correctly predict the improved heat transfer and reduced fouling in these 3D geometries it is essential to accurately account for turbulence. Instead of assuming that a universal model exists for all flow scales such as in a classically applied Reynolds-Averaged Navier-Stokes (RANS) approach, in this work the use of eddy-resolving techniques has been investigated. A large eddy simulation (LES) methodology was implemented in OpenFOAM and validated experimentally and with direct numerical simulations (DNS). The dynamic Smagorinsky model formulated by Germano et al. was used to model the subfilter scales. Simulations were performed for internally finned and swirl flow steam cracking reactors at Reynolds numbers ranging from 11,000 to 38,000. The potential of both geometries is confirmed as heat transfer is augmented by 29-37% at the cost of a pressure drop that is only 31-39% higher compared to a bare straight tube. Although general agreement is reasonable, the applied RANS model clearly fails to capture some secondary flow phenomena arising from the anisotropy of the Reynolds stresses and their influence on global heat transfer and friction. The additional level of accuracy provided by eddy-resolving techniques clearly offers substantial advantages for use as a design and optimization tool, and can even provide validation and closure to existing RANS models in cases where experimental data is unavailable.