The GPU is a very large fraction of the thermal output these SoCs are engineered to handle, so there really shouldn't be any throttling if you're only loading the CPU.
In the M1 generation there was no benefit from going from Pro to the Max.
Apple made their E-cores a lot faster the past couple generations. Despite being a tiny fraction of the size, they get nearly 50% of the performance (yay diminishing returns) while using around 10x less power.
If your task only scales to 1-4 cores, you need those cores to be fast, but generally speaking, if you can scale more than that, you can probably scale to a LOT of cores. From this perspective, 6-8 cores allow a couple of those lightly-threaded applications at the same time which is usually all the more you'll be running.
After that, putting 4-6 E-cores+cache in the space of just one P-core+cache means you're still using half the power, but getting a lot more performance in those very scalable applications.
AMD is headed this same direction with Zen 4c and 5c chiplets where you'll have one performance chiplet and multiple efficiency chiplets.
The area and power wins are just too big to ignore.
Apple has been shifting away from performance cores with each generation. M1 Pro: 75-80% P-cores; M2 Pro: 60-67% P-cores; M3 Pro: 45-50% P-cores.
This shows up when you look at Geekbench results. A 10-core M2 Pro (6+4) gets 12,100 while a 10-core M1 Pro (8+2) gets 12,202. The 12-core M2 Pro (8+4) gets 14,221. That's a 16.5% increase from having 20% more cores. In some ways, this feels like an odd result. Adding two additional M2 P-cores gets Apple a comparatively small result over the 10-core M2 Pro (less than the average core performance). However, adding two efficiency cores over the M1 Pro gives them the same 16.5% boost over the 8+2 M1 Pro.
If I had to guess, maybe it's thermal throttling when running the benchmark. If the additional P-cores can't truly be P-cores under 100% load, then their impact shows up less. Likewise, if the E-cores can match P-core performance under heavy thermal load, then it could show up as being roughly equivalent in the benchmark.
I wonder if real-world scenarios end up differing from benchmarks in a meaningful way around this. For example, core-pinning can be useful for warm caches and in a real-world scenario you might have a process the OS tries to pin to P-core-1 that has spikes in utilization while another process is pinned to P-core-2 with similar spikes. So both get the performance of a P-core and warm caches while the thermal load isn't that high so the P-cores are still at their peak performance (unlike a benchmark that's trying to use all cores as much as possible at the same time).
Maybe this is a business decision more than something based around how the chip performs. The big selling point of the M1/M2 Max was graphics (maybe the extra RAM as well). You could get the same CPU in the Pro or Max. Now the M3 Pro is a 5+6 or 6+6 CPU while the M3 Max is 10+4. 67-100% more performance cores becomes a selling point for the M3 Max even for those who might not care about graphics as much.