Yes, HA deployments in public cloud are 2 x 4 vCPU with the 3/1 data vs control plane split. We've been downsizing to 2 vCPU instances though since the throughput requirements are low and connections is only a couple thousand. Knowing this, I will look to upsize them back to 4 vCPU.
Have to say, I am quite puzzled why CheckPoint would correlate the number of TCP/UDP ports to vCPU count. That seems like a legacy architectural decision that should be revisited. Modern CPUs are certainly cable of handling more than 16533 connections per core without issue.
1) Set up a hide behind many with a pool of three NAT addresses, that will give each of your three worker instances ~50k NAT capacity.
I assume this would also 3x the 16533 in the 2 vCPU instances as well and serve as an alternate fix to paying 2x in compute and licensing costs. GCP does support additional IP ranges assigned to an instance, and the CheckPoint dynamic object used to do the NAT hide could be changed to match that range, so this is possible.
The better option for us though is to use the CheckPoint as an L7 proxy, but this requires an upgrade to R81 so that TLS 1.3 and HTTP2 can be supported. We're on that path anyway, I'm just trying to get a firmer understanding of the limits we're hitting under the existing setup.