Nobody has mentioned write combining yet, which can boost the performance in higher resolutions. If your K6-2 CPU core is CXT, this can be done with K6INIT in DOS or Central Tweaking Unit in Windows. Still, it won't truly outperform a Mendocino Celeron CPU (BTW, it also supports write combining via P6-style MTRR) which has a pipelined FPU that the software renderer of Quake and likely the Fusion emulator does take advantage of.

Gmlb256 wrote on 2022-12-20, 02:37:

Nobody has mentioned write combining yet, which can boost the performance in higher resolutions. If your K6-2 CPU core is CXT, this can be done with K6INIT in DOS or Central Tweaking Unit in Windows. Still, it won't truly outperform a Mendocino Celeron CPU (BTW, it also supports write combining via P6-style MTRR) which has a pipelined FPU that the software renderer of Quake and likely the Fusion emulator does take advantage of.

Hi! thanks, i tested K6init no performance gain at low resolution, in Quake, but a big speed up in 640x480! from 11.2 to 21fps, thanks!

Sadly no performance gain for Fusion, still too slow compared to mendocino, half speed or less than a 333 celeron...

theelf wrote on 2022-12-20, 03:30:

Hi! thanks, i tested K6init but no performance gain, not in quake or Fusion. Im doing something wrong?

No, you didn't do anything wrong. I see that write combing gets successfully enabled for the LFB region in the integrated SiS video chip and that only works with SVGA/VESA video modes. There is also the /vga command line switch (don't use it if B000-B7FF region is used for UMB) that enables write combining for the memory region used in standard VGA video modes.

About my Celeron, even if a lot of benchmarks compare K6-2 to P2, i did nos spected to be even similar... but not half... my mendocino in fusion is 2x faster... or maybe 3x, it feels sooo slow

Yep, I also noticed some slowness with my K6-2+/450 compared to a PII-400 even outside Quake. Benchmarks don't tell the entire story.

Gmlb256 wrote on 2022-12-20, 03:55:

No, you didn't do anything wrong. I see that write combing gets successfully enabled for the LFB region in the integrated SiS vi […]
Show full quote

theelf wrote on 2022-12-20, 03:30:

Hi! thanks, i tested K6init but no performance gain, not in quake or Fusion. Im doing something wrong?

No, you didn't do anything wrong. I see that write combing gets successfully enabled for the LFB region in the integrated SiS video chip and that only works with SVGA/VESA video modes. There is also the /vga command line switch (don't use it if B000-B7FF region is used for UMB) that enables write combining for the memory region used in standard VGA video modes.

About my Celeron, even if a lot of benchmarks compare K6-2 to P2, i did nos spected to be even similar... but not half... my mendocino in fusion is 2x faster... or maybe 3x, it feels sooo slow

Yep, I also noticed some slowness with my K6-2+/450 compared to a PII-400 even outside Quake. Benchmarks don't tell the entire story.

Thanks a lot, ok, maybe conclusion is just K6-2 are slow and end of story. Like you say, benchmarks dont tell the entire story

I was thinking to replace my DX2-66 with this K6-2 build, just to use Fusion, but if fusion is not usable, i will not see any advantage of K6-2 over my actual 486

Sorry for my late tests. I lied, I do have K6-II 400 not 450.... Anyway, here are some test results. I doubt that K6-2 are slow. Problem is your IGP and that's why IMO you are having those issues... Plus You have a frankenstein small PC not a regular one (for example so-dimm ram). Try to disable somehow this bloody integra. Maybe in BIOS you can set ram amount to 0 which will.mean that it is shut down..

Test made without univbe, FSB 100Mhz, 512KB cache, S3Savage4 8MB, mobo Freetech P5F93, 512MB RAM.

Quake 320 1.06, Doom High, all other test for fast PC's (Phils benchmarks)

Sth to add:
https://fops.systems/digipos/digipos.html
This guy mention that it has bad GPU supprt, and it does like TNT2 cards... (???)

K6-2 CPUs are mainly competitive with PII and Celeron CPUs in integer workloads and scenarios where the 3DNow! instructions are used (practically non-existent in DOS). To be honest though, I didn't notice the use of shared memory which can create bottlenecks in memory access. 😜

Spitz wrote on 2022-12-20, 21:56:

Sorry for my late tests. I lied, I do have K6-II 400 not 450.... Anyway, here are some test results. I doubt that K6-2 are slow. […]
Show full quote

Sorry for my late tests. I lied, I do have K6-II 400 not 450.... Anyway, here are some test results. I doubt that K6-2 are slow. Problem is your IGP and that's why IMO you are having those issues... Plus You have a frankenstein small PC not a regular one (for example so-dimm ram). Try to disable somehow this bloody integra. Maybe in BIOS you can set ram amount to 0 which will.mean that it is shut down..

Test made without univbe, FSB 100Mhz, 512KB cache, S3Savage4 8MB, mobo Freetech P5F93, 512MB RAM.

Quake 320 1.06, Doom High, all other test for fast PC's (Phils benchmarks)

Sth to add:
https://fops.systems/digipos/digipos.html
This guy mention that it has bad GPU supprt, and it does like TNT2 cards... (???)

Like i said, the IGP is very fast, any PCI GPU i use is same or slower

At 400 you K6 is slight faster than mine in FPU, and slower than mine in ALU for what i see in your pictures. I see a big difference in Ram speed... 300mb/s vs 100mb/s i have, are you sure is correct?

I think you run quake at 320x200

and Doom shareware 1.9

I think this numbers are more or less the ones mwdmeyer check with a SIS vs Soltek board, that is 1.28 time slower

If i get 55.5 in sis in the soltek will be like 55x1.28 = 71.04

71.04 at 450mhz is 63fps at 400, more or less what do you get

But my Mendocino a 66mhz bus and 333mhz, i get 86fps in quake!!!! to have 86fps in a K6-2 even in a good board i will need something like 550mhz and 650mhz in a sis

This make sense in fusion, when i feel is like half speed my mendocino

Gmlb256 wrote on 2022-12-20, 23:26:

K6-2 CPUs are mainly competitive with PII and Celeron CPUs in integer workloads and scenarios where the 3DNow! instructions are used (practically non-existent in DOS). To be honest though, I didn't notice the use of shared memory which can create bottlenecks in memory access. 😜

I tested dedicated GPU and IGP and no difference in tests

Yes, in integer K6 is good, but FPU look terrible, and i check, Fusion emulator use extensive use of FPU

I will not continue losing time with the K6, i think is pretty clear is slow, and without fusion emulator, i will continue using my 486 PC for DOS

theelf wrote on 2022-12-21, 00:25:

Yes, in integer K6 is good, but FPU look terrible...

The K6 FPU is actually quite fast but it is NOT pipelined. Most common FPU instructions like FADD, FSUB, FMUL have only 2-clock latency on K6 while on the P6 (Celeron) FADD and FSUB have 3-clock and FMUL has 5-clock latency. But because of its non-pipelined nature K6 FPU does not work well with long chains of floating point instructions(like the hand crafted overly Pentium optimized FPU code in Quake).
Quake also uses many FXCH instructions to rearrange FPU stack which is free on Pentium but not free on any other contemporary CPU architectures from AMD and Cyrix.
K6 works much better with applications that interleave integer and FPU code (like spreadsheets etc.)
The MMX and 3dnow! execution units are fully pipelined on K6-2 /3.