It is difficult to generate smidth predictor feedback loop (1-exp(-tds))G(s) exactly unless you have an ideal delay generating device. However, the the delay can be approximated by pade's or any other appropriate approximation ((1-tds/2)/(1+tds/2)). Then you can realize it by analog circuits like opamp, registors, capacitors etc. But if have are a pneumatic controller then probably you can you a long flow chanel which generates the delay
The Smith predictor is a good theoretical tool, but I think it is difficult to implement with an analog controller. It is easy to implement in Simulink, but not easy in real life. The Smith Predictor eliminates the delayed response and replaces it by the response of the system if it was not delayed. For calculating this you need the model G(s) of your system and the exact delay (td).
The best thing would be reducing the delay as much as possible. If this is not possible, you can use a PID with a properly selected D action. It is said that a well tuned PID also has an anticipating effect similar to a Smith Predictor.
Smith predictor involves much computations which can not be implemented using analog controllers. Using digital computers easily it can be implemented.
Yes, the crucial problem is the delay realization and the model accuracy. However, there are many methods how delays can be approximated by a finite-dimensional model, see e.g. works Partington (2004).
(1) The Smith Predictor is a type of predictive controller for systems with pure time delay. In addition, the Smith predictor can regarded as a dead time compensator.
One can see that in the original proposal, the Smith Predictor was implemented in discrete form (digital circuit) instead of continuous form (analog circuit).
https://en.wikipedia.org/wiki/Smith_predictor
Quote Juan Antonio Baeza's comment "it is difficult to implement with an analog controller. It is easy to implement in Simulink, but not easy in real life."
Quote Angeline vijula Dhanral's comment "Smith predictor involves much computations which can not be implemented using analog controllers. Using digital computers easily it can be implemented."
In summary, The Smith Predictor has been widely implemented in discrete form or in form of digital controller (i.e., digital circuit) in the real-world industrial applications.
(2) First-order Pade approximation of a time delay has been widely applied in the real-world control system design, i.e., exp(-s\tau)=(1-s\tau/2)/(1+s\tau/2), as suggested by Prasanta Roy.
However, Pade approximation also places a constraint (i.e., |s\tau|