Here are the Speedsys results with my TXP4 with the motherboard cache enabled, and disabled. Memory throughput with the motherboard's cache enabled is 149.23 MB/s, disabled it's 187.26.

Demolition-Man wrote on 2023-10-02, 19:27:

It's not easy with a DOS/Win 98SE PC. With every MHz more, or every little bit more speed (cache), the chance that some DOS game will no longer work properly increases.

Speed-sensitive software is one of the reasons why I have "time machine" computers and it isn't just limited to DOS games.

List of CPU speed sensitive games from VOGONS Wiki

list of speed sensitive windows games (NOT DOS)

This is an interesting topic.
I have the problem with some games ,that I don't know how fast they should run, because I haven't seen them on hardware from that time.^^

No problems with with any of my games on the K6-III 400 or the K6-2+ 500. The K6-III 450 gives me the famous 6003 Error in Monkey Island ,but anything else works.

Games i tried:
Aces of the Pacific on my slower 166 Mhz machines, but only as a free memeory test.
Descent runs very smooth, but too fast? I dont know.
Descent II runs perfect , normal or with Glide (my only 3Dfx Game on DOS). No mouse involved.
Duke Nukem II still works , but I've had sound problems before.
Jazz Jackrabbit (CD) no problem.
Secret of Monkey Island RUntime 6003 error
Theme Park just press R ?!
Tyrian works, but not with my fastest Win 98 PCs.

Thats it.

I got my first in 1996, when I was 12. I only know the games from 120 MHz upwards.

Did my own Speedsys L3 cache test:
On the M5ALA board with he K6-III 400 CPU.
Almost identical Memory throughput: 138.75 MB/s with L3 cache, 144.63 MB/s without L3 cache.
So it doesn't matter at all with this board or chipset?

Demolition-Man wrote on 2023-10-03, 18:56:

Did my own Speedsys L3 cache test: On the M5ALA board with he K6-III 400 CPU. Almost identical Memory throughput: 138.75 MB/s wi […]
Show full quote

Did my own Speedsys L3 cache test:
On the M5ALA board with he K6-III 400 CPU.
Almost identical Memory throughput: 138.75 MB/s with L3 cache, 144.63 MB/s without L3 cache.
So it doesn't matter at all with this board or chipset?

The Ali Aladdin V chipset has a much better L2 cache than the Intel VX. In the now moribund thread for setting the highest 3D Mark 2000 score with a Super 7 motherboard, the Ali equipped motherboards with 1024k L2 have the highest scores. I would also test with other benchmarks, but in this case it will probably be too close to call. Other variables will be the amount of RAM vs. the cacheable area of the chipset.

"One of the most important features the Aladdin V chipset boasts is the internal L2 cache the chipset features, more specifically the M1541 chip has an integrated 16K x 10-bit Tag RAM as well as 16K x 2 L2 cache SRAM, both of which decrease cost and increase performance somewhat. The most important reason for including the Tag RAM on chip is to make using the 100MHz bus speed possible. The 100MHz Bus Frequency puts a heavy strain on the L2 cache, and tests have shown that without fast L2 cache & a capable Tag RAM chip (6ns or faster) the 100MHz bus speed has the potential to become quite erratic. By placing the Tag RAM on the chipset ALi eliminates the possibility of the L2 cache or chipset as being the limiting factor in making use of the 100MHz bus speed."
https://www.anandtech.com/show/72/3

Great tests and interesting results Repo-man11, could you try to test both k6-3 and 3+ with dos benchmarks?
Like quake and doom, they should give more constant results than windows tests.

I've done some quick testing of onboard cache enabled / disabled on my Gigabyte GA-586-T2, using a K6-2+ 450:

enabled
- speedsys memory throughput: 104.52
- PCPbench SVGA: 65.3
- 3dmark99: 2500

disabled
- speedsys memory throughput: 151.31
- PCPbench SVGA: 64.3
- 3dmark99: 2457

The memory throughput increased significantly with external cache disabled, but I doubt that would have much impact on gaming. The other 2 scores are slightly worse with cache disabled.

So I guess I'll leave it enabled.

I ran more tests on my TXP4, and the results are a mixed bag, though slightly in favor of leaving the motherboard's cache disabled.

TXP4 no L3
Chris bench: 533.4
Chris 3d SVGA: 115.5
Quake timedemo 360x480: 39.1
Doom max details: 2134 gametics 1002 realitics
3d bench 1.0c: 265.9
3d bench: 33.3
3d 99: 3201/7543
3d 2000: 2684
3D 2001: 2417

With L3 enabled
Quake timedemo 360x480: 38.4
Doom max details: 2134 gametics 997 realtics
Chris bench 539.4
Chris 3d SVGA: 119.5
3d bench 1.0c: 265.7
3d bench 33.3
3D 99: 3160/7215
3D 2000: 2711
3D 2001: 2383

I did a few more tests with the M5ALA board. Dosbench, Doom & Quake, 3DMark 99. You should leave the cache on. It's not a big difference, but every test with L3 cache is slightly faster.
Dosbench:
With L3 cache:
a) 2134 / 187
b)2134 / 828
c) 969 / 14.8 / 65.6
d) 969 / 36.3 / 26.7
e) 969 / 60.6 / 16
3D Mark 99 (V3 2000 PCI)
Score:2546
CPU: 5635

Without L3 cache:
a) 2134 / 193
b) 2134 / 838
c) 969 / 15.4 / 63
d) 969 / 37 / 26.2
e) 969 / 61 / 15.9
3DMark 99
Score: 2469
CPU:5508

Nemo1985 wrote on 2023-10-03, 19:43:

Great tests and interesting results Repo-man11, could you try to test both k6-3 and 3+ with dos benchmarks?
Like quake and doom, they should give more constant results than windows tests.

Okay, I ran most of the dosbench tests. I cannot do any comparisons with the motherboard's cache enabled/disabled because this motherboard has fake cache chips, and I don't have a COAST module that will work at this speed. The difference is very small, but as before the K6-3 has the edge over the K6-3+.

PCChips M520 with K6-3
3d 99: 2495/5699
3d 2000: 2115
3d bench 1.0c: 240.0
Chris’s 3d benchmark: 469.0
Chris’s 3d 640x480: 96.4
Doom max details: 2134 gametics, 1126 realtics
Quake timedemo 360x480: 30.9
Quake time demo 640x480: 19.5
Topbench: 438

M520 with k6-3+
3d 99: 2430/5648
3d 2000: 2032
3d bench 1.0c: 239.7
Chris’s 3d benchmark: 466.4
Chris’s 3d 640x480: 96.3
Doom max details: 2134/1134
Quake timedemo 360x480: 30.8
Quake time demo 640x480: 19.4
Topbench: 431