It is known that Spectre and Meltdown fixes can slow down computers substantially, commonly involving in some cases I/O operations and kernel calls. Last month, MIT Lincoln Laboratory published a (not peer-reviewed) article indicating that pMatlab simulations and Keras with Tensorflow can be slowed down by 21% and 16% respectively for kernels with all mitigations implemented and microcode updates for the processors. With just the microcode updates, the slowdown is still 19% and 15%. However, not that this research was done on unoptimized code.
Spectre and Meltdown are vulnerabilities discovered at the beginning of this year affecting most modern processors. Using these vulnerabilities, it is possible to retrieve encryption keys. The bugs are at the hardware level, so they cannot be directly patched. Mitigation patches have been developed and they can work around these issues, but they come at a cost in performance – they disable certain features in the processor at specific times. These mitigations are still under active development to reduce the performance penalties as much as possible.
Source : https://arxiv.org/abs/1807.08703