This machine has been a delusion.
A bit of info, for your own knowledge:
The operating system was installed on a single SATA disk, using ZFS. That leave 2 nvme disk and 1 SATA disk free.I have setup a new pool, made up from the two NVME disks, in RAID 0/stripe.
The NVME disk transfer rate is approx. 1000 MB/s (as reported from diskinfo), the aggregate transfer rate of the striped pool is approx 1500 MB/s. Impressive!
But that doesn't translate to a builder performance improvement, (ok, there is an improvement, but can be measured like +3% or+ 4%). Why so ?
a) ZFS, (and the 64 GB RAM installed), do an astonishing job in caching lot of things, as a result, the disk is really used only to write the built packages, the logs, and eventually the ccache. Additionally the write operations are of course asyncronous, in that synth "request" the write, the kernel/ZFS take in charge the job, and return an ACK to synth, before the files are effectively stored to the disk. The kernel will do the effective write sometimes later, by itself, in the meantime synth will continue with other packages.
The same "write offload" happens with SATA disks too, that's why there is no real difference by using NVME vs SATA. Probably swapping onto an NVME disk would be much faster, but here we don't swap at all.
b) The high speed of the nvme disks come to a price, in term of CPU usage, true you can achieve impressive write speed, but the CPU usage (sys in this case), will be relatively high. By doing that, those CPU cycles are detracted from the build processes. This situation could be eventually improved by using "intelligent/smart/cached/hw controllers, but that would be difficult to manage remotely. (and very likely would come at a cost too).
long story short:
for builders choose the better CPU you can afford.
Now, the good news: we (me and ericbsd) already did it.
SoYouStart renewed their offer, and we decided to grab two machines now that are available:
We have got two of these:
https://www.soyoustart.com/ca/en/offers/173sys20.xml
These two machines (both) come at the same price of the E3-1670, but each of these perform better compared to the E3-1670, and we have two of them.
My personal thanks to the E3-1670, that I stressed a lot,

but we will let it expire.
PS: ericbsd performed also a test using "poudriere" on this machine, the result was 440 pkg/hour.
I'm now doing a last round, just for comparison and will post the result here when completed.