The MMX has double L1 cache, among other things, which will trump the double L2 (almost) any time.

The Pentium MMX is 8-10% faster per clock at virtually everything. The 512k vs 256k L2 is about +2% more if having the right use.

MMX with 256 KB L2 is faster than non-MMX with 512 KB. You have 15% to 30% gain by MMX (15% due to larger L1 cache and up to 30% if MMX is used) against about 5% by the larger L2 cache. Also, you can use software that requires MMX. Go for the MMX.

Amount of L2 cache is relatively unimportant. Other factors are much, much more relevant.

- L1 cache (in Pentium MMX) as already said
- type of cache. Pentium MMX will probably (but not absolutely always) have PLB cache, Pentium non-MMX could have PLB or asynch. Asynch is significantly slower than PLB, about 10% in RAM benchmarks.
- chipset performance. Clock-for-clock there's a factor 3 performance difference between the slowest and fastest So5/7 chipsets. It's unlikely either CPU will have the very slowest chipsets, but there's still chance of 25% difference between slowest and fastest likely chipset. In general MMX chips get paired with better chipsets than non-MMX, but that's no guarantee.
- presence of integrated graphics - halves memory bandwith, so kills performance. Possible with both CPUs.

So please, give complete specs if you want a good answer. Although in general based on the info you gave the MMX would win, there are more than enough scenarios where the non-MMX would trounce it (think: MMX with SiS 5596 intetegrated VGA chipset vs non-MMX with i430HX and PLB cache).

dionb wrote on 2020-03-16, 00:39:

- presence of integrated graphics - halves memory bandwith, so kills performance. Possible with both CPUs.

Is there a particular benchmark that would demonstrate the improvement when I switch to a discrete video card?

dionb wrote on 2020-03-16, 00:39:

So please, give complete specs if you want a good answer. Although in general based on the info you gave the MMX would win, there are more than enough scenarios where the non-MMX would trounce it (think: MMX with SiS 5596 intetegrated VGA chipset vs non-MMX with i430HX and PLB cache).

Gotcha. I don't know too much about the machine right now.

I appreciate all of your feedback—thank you.

grettke wrote on 2020-03-17, 02:01:

[...]

Is there a particular benchmark that would demonstrate the improvement when I switch to a discrete video card?

Pretty much anything - these early integrated VGA chipsets didn't have the kind of advanced memory management current ones do, regardless of actual usage. Memory benchmarks would show the biggest delta. I once ran RAMspeed INTmem and FLOATmem benchmarks on most So7 chipsets. FLOATmem score for P100 with EDO on 5598 on integrated VGA: 32.27. Same on discrete VGA: 62.69. By comparison, on the fastest S5 or 7 chipset (ALi Aladdin IV+) the same combo scored 69.48 (and 79.49 with SDRAM instead of EDO), on the slowest (UMC UM8891C) 28.9. to 36.89 This tells you exactly what's going on with the integrated VGA: the UMC8891 has a 32b 486 memory controller, whereas Pentium-class chips normally have a 64b memory controller. Using integrated VGA pushes your memory performance down to 486 levels.

Note that I'm consistently referring to integrated VGA, not onboard VGA. Onboard VGA is basically just the same chips as on a discrete card stuck on the motherboard. So long as the VGA chip has its own memory it doesn't impact memory bandwidth and performance, and CPU/memory performance will be the same as on a system with discrete graphics.

One of the strengths that the PMMX has over the older core is an improved method for branch prediction.
But for practical purposes(no matter how much cache) using the older core won't make that much of a difference since you're using "486 era" software. The original P200 is also more rare than the P55C.

The Pentium MMX is also much more versatile when it comes to throttling methods. For speed sensitive titles, this is a very useful feature.