Based on my issues with throttling on the 5500M equipped systems, and after extensive investigation from Apple's engineering team, they agreed to a refund of my 5500M equipped pro.
So, I've ordered another 16" system (2.4GHz, 64GB RAM, 8TB SSD, 5600M) and ran it identically alongside the 5500M equipped model, hoping that the 5600M system would solve my problem.
Testing was performed with the following:
Catalina 10.15.7.
Left side: Apple first gen USB-C to HDMI adapter connecting to a Dell U3017 at 2560x1600, OWC Thunderbolt 3 dock (no display outputs used on this).
Right side: LG Ultrafine 5K Thunderbolt adapter, Apple 96W power adapter.
Lid open, so three active displays.
I plugged in the 96W power adapter first and verified this was being used as the power source (the LG panel can provide 87W).
CPU throttle test performed as follows:
OBS 26, single scene, recording to local disk with one 1080p60 video capture source (via Elgato HD60S+)
Adobe Lightroom with photo export from raw images to jpeg.
Without the CPU intensive resources, I see the GPU sitting at around 10W with all three displays active on the 5600M, but on the 5500M it shoots up to the 25W or so. With a zoom call running, the 5600M sits at around 10W.
With CPU/GPU intensive resources, the 5600M sits at around 20W, whereas the 5500M sits at around 25W.
CPU performance throttle occurs in both situations - it takes slightly longer for the 5600M system to enter this state, but with the GPU pulling 20W+, in both cases, I see the CPU get pegged at 1GHz with 90+ CPU core utilisation, and CPU temp not exceeding 70W (i.e. the throttle cause is not temperature related). The 5600M was able to manage the encoding just fine without the lightroom export, whereas the 5500M system hit 1GHz throttle with just the encode running.
With just a Zoom call running, the GPU sits on around 15W with all three displays active. Running a Lightroom export in the background sends both systems into throttle in less than a minute.
So, the 5600M solves the three-displays-at-idle problem, but as soon as I load up either system, I'm in the same situation.
For high performance workloads when docked, it seems the only solution is to relieve the system of the power load of both the CPU and GPU and use an external Thunderbolt-equipped GPU. Have others had any good experiences with Thunderbolt dGPUs? They seem very expensive given their current technical capabilities in 2021 and the fact that they're not even supported on the new Apple Silicon Macs.