Following the experiences narrated in this thread: Looking for an OLD raytracing benchmark for pure DOS and following a hiatus of a few months, I started using this benchmark on all my retro machines + processor combinations. It's a long process, but I intend to build a database with all the results, as I do for the other benchmarks I run on my retro PCs.

C-Ray is a benchmark created by John Tsiombikas ( https://github.com/jtsiomb/c-ray ). I merely compiled the source under DOS using an old Watcom CPP compiler after implementing some simple QoL tweaks.
The benchmarks runs on any 386 class processor with a FPU , some XMS memory (4Mb should be enough) and a VESA 2.0 compliant video card capable of 32bpp (either natively or though a VBE driver). So far I've been able to run in on old 486 boards with PCI or VLB, and several Pentium motherboards. The benchmark relies heavily on the FPU, even more so than DOS Quake. Cache, memory latency and other aspects of the hardware architecture seems to have a very limited impact on the result. By default it creates a 640x480 image with 1 sample per pixel. Resolution and samples per pixel can be specified. The results scale linearly, so an image of 320x240 with 4 samples per pixel will take as much time to be computed as one of 640x480 with only one sample per pixel.

Attached you'll find the preliminary results for my machine based on a Zida Tomato 4DSP motherboard, 256KB cache, 32MB of 60ns RAM. I tested this configuration with several CPUs from my collection. Benchmark was run with default settings (640x480, 1 sample per pixel)