Let's benchmark our systems with caches disabled

[lvader]

I did some benchmarks of my socket 7 Pentium 233 MMX build. I have a couple of toggle switches that lets me change the FSB between 50 and 66 FSB and between 3.5 and 2.0 multiplier. I'm also using a few combinations with SETMUL to get a good spread of speeds.

Interesting that FSB has a much greater influence on slowdown compared to multipliers).

233 MHz
Speedsys / 3Dbench2
174.3 / 165.5
144.63 / 138.8 (VPD, BPD)
77.53. / 59.9 (CCD)
35.90. / 38.02 (CCD, DCD)
30.1. / 33.3 (CCD,DCD,VPD,BPD)
12.95 / 15.8 (CCD,DCD,VPD,BPD, L2 disabled)

175MHz (3.5x50)
Speedsys / 3Dbench2
131.11 / 124.3
108.77/ 104.3 (VPD, BPD)
58.30 / 45.0. (CCD)
27.0. / 28.7 (CCD, DCD)
22.56 / 25.0. (CCD,DCD,VPD,BPD)
9.70. / 11.8 (CCD,DCD,VPD,BPD, L2 disabled)

133mhz (2.0x66mhz)
Speedsys / 3Dbench2
99.62 / 128.3
82.64. / 101.2 (VPD, BPD)
67.99. / 54.0 (CCD)
33.3 / 35.8 (CCD, DCD)
25.77 / 29.7. (CCD,DCD,VPD,BPD)
12.62. /. 15.3 (CCD,DCD,VPD,BPD, L2 disabled)

100mhz (2.0x50)
Speedsys / 3Dbench2
77.91. / 96.3
62.15 / 76.0 (VPD, BPD)
51.13 / 40.5. (CCD)
25.04 / 26.8 (CCD, DCD)
19.38./ 22.3 (CCD,DCD,VPD,BPD)
9.46. / 11.4. (CCD,DCD,VPD,BPD, L2 disabled)

[lvader]

Here are the PCBench and Doom benchmarks

233 MHz
PCBench / Doom
59.9 / 81.83
48.3 / 71.37 (VPD, BPD)
16.2 / 24.28 (CCD)
10.2 / 14.34 (CCD, DCD)
9.1 / 12.66 (CCD,DCD,VPD,BPD)
3.4 / 4.51 (CCD,DCD,VPD,BPD, L2 disabled)

175MHz (3.5x50)
PCBench / Doom

45.0 / 71.36
36.3 / 53.65 (VPD, BPD)
12.2 / 18.01 (CCD)
7.60 / 10.76 (CCD, DCD)
6.90 / 9.36 (CCD,DCD,VPD,BPD)
3.0 / 3.87 (CCD,DCD,VPD,BPD, L2 disabled)

133mhz (2.0x66mhz)
PCBench / Doom
41.5 / 65.34
31.2 / 54.0 (VPD, BPD)
14.6 / 22.08 (CCD)
9.6 / 14.21 (CCD, DCD)
8.3 / 11.39 (CCD,DCD,VPD,BPD)
3.8 / 5.22 (CCD,DCD,VPD,BPD, L2 disabled)

100mhz (2.0x50)
PCBench / Doom

31.1 / 48.9
23.4 / 40.65 (VPD, BPD)
11.0 / 16.41 (CCD)
7.2 / 10.50 (CCD, DCD)
6.2 / 8.40 (CCD,DCD,VPD,BPD)
2.9 / 3.80 (CCD,DCD,VPD,BPD, L2 disabled)

[clueless1]

I added your results, lvader, just waiting on the system specs to fill in. I also tweaked the layout of the spreadsheet a little for easier readability. For example, I froze each sheet at column C so when you scroll to view results, you can still see the cache settings.

[Falcosoft]

Hi,
Just out of curiosity:

CPU: AMD Phenom II X4 960T
Clock: 4x230 (920 MHz) - Forced with my Phenom low power util: http://falcosoft.hu/dos_softwares.html#pwr
Chipset: AMD-790X
RAM: DDR2-920 8192 MB
VGA: Radeon HD 7770 PCI-E
No sound under DOS

Results:
All caches ON / All caches OFF (L1+L2+L3)
3DBench2: Too Fast / 36.1
SpeedSys: 1345.58 / 32.35
PCBench: 428.2 / 9.8
Doom RT: 430 / 5212
Doom FPS: 173.69 / 14.33

[infiniteclouds]

CPU: AMD K6-III+
Clock: 5.5x100 (550MHZ)
Chipset: VIA MVP3 AGPset
RAM: PC133 SDRAM 128MB
VGA: Voodoo3 2000 (AGP)

L1 Cache - 64KB
L2 Cache - 256KB
L3 Cache - 512KB (Motherboard)

I don't really have more than a very basic understanding of caches but based on my benchmark results I am assuming that 'location' (L1, L2 or L3) is more important that size (KB) when it comes to a caches affect on speed.

BIOS Settings are 'Turbo Defaults'

Results:

None Disabled

SpeedSys: 631.11
3DBench2: 437.1
PCBench: 173.3
Doom: 530 Realticks /140.94FPS

L2 Disabled

SpeedSys: 630.92
3DBench2: 387.4
PCBench: 155.6
Doom: 628 Realtick/118.93FPS

L1 Disabled (using SETMUL L1D)

SpeedSys: 35.20
3DBench2: 36.3
PCBench: 10.2
Doom: 4775 / 15.61 FPS

L1 Disabled (SETMUL L1DX)

SpeedSys: 35.20
3DBench2: 40.3
PCBench: 10.2
4765 / 15.67 FPS

L1 + L2 Disabled

SpeedSys: 35.20
3DBench2: 36.3
PCBench: 10.2
Doom: 4775 / 15.64 FPS

L3 Disabled

SpeedSys: 631.10
3DBench2: 437.1
PCBench: 160.4
Doom: 539 / 138.57 FPS

L2 + L3 Disabled

SpeedSys: 630.83
3DBench2: 340.9
PCBench: 90.7
Doom: 736 / 101.48 FPS

L1 (SETMUL L1D) and L3 Disabled

SpeedSys: 13.37
3DBench2: 14.0
PCBench: 3.7
Doom: 13660 / 5.47 FPS

L1 (SETMUL L1DX) and L3 Disabled

SpeedSys: 13.42
3DBench2: 16.4
PCBench: 3.7
Doom: 13576 / 5.50 FPS

All Disabled

SpeedSys: 13.37
3DBench2: 14.0
PCBench: 3.7
Doom: 13660 / 5.47 FPS

I did separate benchmarks using SETMUL L1D and SETMUL L1DX because the first one seems to yield the exact same result as disabling both L1 and L2 caches. This leads me to believe that 'SETMUL L1D' automatically disables L2. I suppose L1DX wouldn't be needed, if this wasn't the case.

I'd really like to be able to hit a 486DX2-66 performance. I'm not sure why I can't? With cache disabled SpeedSys shows a higher score than clueless' actual machine but I'm getting almost -9 FPS compared to the real 486/66.

[clueless1]

@infiniteclouds

Thanks for submitting results. You are welcome to put these results in the spreadsheet (link in the first post) if you haven't already. Just put them in at the bottom and when I get around to it, I'll copy them into a dedicated tab and then re-sort.

[infiniteclouds]

No problem -- it's a great reference! I admit I'm disappointed that there's no way to get a K6 to perform inbetween DX-33 and a Pentium 133.

[infiniteclouds]

It is a bit confusing though since I see a lot of people on here check the 'scaling' capabilities on their CPUs based on SpeedSys yet there are clear instances where a test will yield a higher SpeedSys rating than another CPU but lower framerates.

Your own i486DX2-66 test on there has a SpeedSys of 25.14 vs 35.20 of the K6III+ L1D but yields 24.29 FPS in Doom where the K6 was nearly 10 FPS less at 15.67FPS.

[clueless1]

That's why we use multiple benchmarks, because when you disable caches, some benchmarks don't react as you'd expect. It's the game performance that matters in the end.

[infiniteclouds]

It seems like a Pentium MMX233 would let me hit 486DX2-66 and DX4-100 marks. What sucks is that my Aopen X59 can't set a FSB lower than 66. This is a major bummer because the Pentium MMX @ 100mhz (50x2) with CCD seems to be at the perfect speed for the notoriously difficult Ultima VII. From what I've read 'ICD' is immune to Ultima VII's evil cache sorcery -- I wonder if this applies to CCD as well? If so it would make a Pentium the only way you can play the game on a S/S7 system.

I am intrigued by the '27 choices' in SetMul for some of the VIA chips though right now the tested settings for those CPUs seem limited in the spreadsheet.

[gdjacobs]

To the best of my knowledge Ultima VII's unreal mode doesn't touch TR12 registers - just the L1 cache enable/disable bit.

[clueless1]

To the best of my knowledge Ultima VII's unreal mode doesn't touch TR12 registers - just the L1 cache enable/disable bit.

You are correct, sir. I did lots of tests here: http://www.vogons.org/viewtopic.php?f=5&t=51782
While L1D has no effect on Ultima VII, CCD (L1 code cache) and DCD (L1 data cache) each do. And CCD+DCD is roughly equivalent to L1D, so you can use both to effectively accomplish L1D.

[Kamerat]

Quantity isn't always best but perhaps some of you will find this useful. Used clueless1's 486 DX2 66MHz results as a reference.

Notice I also included som results with slowdown utilities like THROTTLE.EXE and SLOWDOWN.COM.

[infiniteclouds]

And CCD+DCD is roughly equivalent to L1D, so you can use both to effectively accomplish L1D.

I wonder though if CCD+DCD is any different than L1D+CCD+DCD?

[clueless1]

And CCD+DCD is roughly equivalent to L1D, so you can use both to effectively accomplish L1D.

I wonder though if CCD+DCD is any different than L1D+CCD+DCD?

L1D+CCD+DCD = L1D
If you check my Pentium Overdrive results on the spreadsheet, you will see some small differences between L1D and CCD+DCD, mostly in Speedsys.
When I benchmark L1D+CCD+DCD, the results are identical to L1D alone.

[infiniteclouds]

I benchmarked every combo anyway... you can filter them out however you like and keep what you want.

I've just added my results for my Pentium MMX233 at 233 and 133 clock speeds with every possible combination of caches and TSRs. We can compare to Ivader's to see differences between 64MB vs 128MB RAM and VIA MVP3 vs Intel430TX chipsets. We're both using the same GPU (V3 2000) with the only difference being he is using PCI and I am using AGP.

It's a great chip and has a lot of advantages over the K6-2/III+ in terms of 386-486 scaling.

I'd like to match a specific 386 or 486 machine for various settings but it's really hard for me to make sense of the real machine benchmarks (also found in Phil's VGA benchmark sheets). As someone who did not own a PC back then it is very confusing... especially the SX/DX situation as I could've sworn from what I've read that SX chips were slower than DX. Yet I see i486SX-25's on there that have higher benchmarks than the i486DX-33...

[clueless1]

I benchmarked every combo anyway... you can filter them out however you like and keep what you want.

I've just added my results for my Pentium MMX233 at 233 and 133 clock speeds with every possible combination of caches and TSRs. We can compare to Ivader's to see differences between 64MB vs 128MB RAM and VIA MVP3 vs Intel430TX chipsets. We're both using the same GPU (V3 2000) with the only difference being he is using PCI and I am using AGP.

It's a great chip and has a lot of advantages over the K6-2/III+ in terms of 386-486 scaling.

I'd like to match a specific 386 or 486 machine for various settings but it's really hard for me to make sense of the real machine benchmarks (also found in Phil's VGA benchmark sheets). As someone who did not own a PC back then it is very confusing... especially the SX/DX situation as I could've sworn from what I've read that SX chips were slower than DX. Yet I see i486SX-25's on there that have higher benchmarks than the i486DX-33...

Holy crap, that's a lot of results! Thanks Until I have time to process them all, I just moved them into their own tab, so others can still check them all out. Again, thank you, that's really awesome!

[j^aws]

I wonder if this applies to CCD as well? If so it would make a Pentium the only way you can play the game on a S/S7 system.

Yes, CCD is also immune.

The Pentium isn't the only way you can play Ultima VII on Socket 7. On a hardware Turbo-switched Socket 7 motherboard, the game speed is immune to cache re-enabling done by this game. This means any CPU compatible with S7, and with enough top-end performance can be scaled down to a suitable speed.

You can also achieve similar results with S7 boards that can force L1 cache policy from Write-Back to Write-Through (slow), and Ultima VII fails to renable L1 because it's still enabled. So, any compatible S7 CPU can work with enough top-end speed, too.

EDIT:

The Centaur-based Winchip and Setmul should also work using ICD on S7, as VIA C3s on Socket 370/ Slot 1 are also immune to the Ultima VII L1 re-enabling (it will re-enable, but you use ICD instead).

[clueless1]

I made a little discussion of Ultima VII and slowing systems down on S7 and 486 and came up with a way to benchmark the game:
http://www.vogons.org/viewtopic.php?f=5&t=51782

[infiniteclouds]

The Pentium isn't the only way you can play Ultima VII on Socket 7. On a hardware Turbo-switched Socket 7 motherboard, the game speed is immune to cache re-enabling done by this game. This means any CPU compatible with S7, and with enough top-end performance can be scaled down to a suitable speed.

I guess I should've said SS7. Even with S7, motherboards with a working turbo switch are quite unusual. Even less that work as a legit turbo like your Rise board instead of just being a FSB switch as many others are.

You can also achieve similar results with S7 boards that can force L1 cache policy from Write-Back to Write-Through (slow), and Ultima VII fails to renable L1 because it's still enabled. So, any compatible S7 CPU can work with enough top-end speed, too.

This sounds really interesting. I'm not familiar - could you please expand on this? Is this through MTRRLFBE? Or some sort of BIOS setting?

Also, as I have an Ezra CPU (866mhz) and a 6BXC on the way I'm curious about the amount of SetMul options for it. It has less then the Pentium I think?

VIA C3 =
L1D
L2D
ICD
BPD

No VPD, CCD or DCD settings for VIA?