WJG6260 wrote on 2022-09-07, 01:42:

Hello pshipkov and Chadti99, […]
Show full quote

Hello pshipkov and Chadti99,

I've been following your guys' work for some time now and it is quite impressive. Talk about taking things to the next level!

@Chadti99
I saw your VLB thread and have to say that I am quite impressed with your work in the area! These benchmarks here are interesting; I have a Paradise Barbados64 S3 Vision864 card and a Miro20SD S3 Vision868 card; having briefly benched both, I can say that the Miro is slower. I find it interesting that your Vision868 is slower than your Trio64; I wonder if memory clocks were more conservative? As far as I understood, the Trio64 is basically an 864 with all components integrated for cost-savings, and the 868 is an 864 without CGA/EGA backwards compatibility and with some basic video scaling capabilities. It seems that this may be something worth looking into, as far as to what differentiates the Trio64 and S3 Vision864/868. Have you messed with memory clocks at all with MCLK to see if that changes things on the Vision868?

Sincerely,
Will

The more the merrier, looking forward to any contributions you can make. I'm a bit obsessed with Socket3, first PC I built, and I'm just piggybacking off of pshipkov's and feipoa's efforts. I think there is still some performance to be had out of this platform. Especially with this new circuit that's been worked on for custom FSB speeds, or the possibility of using faster cache and dram modules.

I wish I knew what the difference was between the performance of some of the similarly equipped VLB cards. Why is the ARK the fastest DOS performer? I'm sure someone could answer, you would think it's mostly wait states and memory speed. I've got a couple more VLB cards on the way, so more tests to come. Is the MCLK utility for S3 chipsets specifically?

pshipkov wrote on 2022-09-10, 18:08:

@chadti99 […]
Show full quote

@chadti99

That N9 BIOS for the S3 Trio64 VLB video card is cool.
On the Asus VLI-486SV2GX4 with POD100 CPU i am seeing these differences:

Diamond V2.01 BIOS:
Wolf3D: 135 fps
PC Player: 28.8 fps
Doom: 60.03 fps (1244 ticks)
Quake 1: 26.6
WinTune 2: 13300 KPixels/second

N9 BIOS:
Wolf3D: 137.3 fps [+]
PC Player: 29 fps [+]
Doom: 59.95 fps (1254 ticks) [-]
Quake 1: 26.6 fps [=]
WinTune 2: 13400 KPixels/second [+]

As you can see - same in Quake 1, tiny bit worse in Doom, better in Wolf3D, PC Player and WinTune.
Also, the Diamond BIOS keeps the card dark for much longer during POST while the N9 one turns on the lights right away.
Will keep the N9 BIOS in the card.

Interesting results, glad it worked out for you!

@Chadti99
Truly, I'm glad to be here!

Socket 3 is pretty versatile, and I agree completely that there's a lot to be had! It's amazing how people are improving upon a 25 year old platform; Atheatos' anyclk system is nothing short of incredible!
Tomorrow I'm going to try out the N9 BIOS you sent. From pshipkov's results, it seems to be an upgrade over the Diamond BIOS.

As far as the ARK, I recall reading that it is something to do with the highly buffered memory subsystem that the card uses. It's clear that the card's memory bandwidth shines, being pretty much unmatched by most/all VLB cards. Interestingly, I believe the ARK1000VL has a 32-bit memory interface. Even the earlier Mach32 has a 64-bit interface. Something has to be architecturally different about it; I've read references to some sort of "acceleration engine" or coprocessor contained within the chip. Anyone have any ideas?

MCLK has some support from non-S3 chipsets, at least as far as I recall. I've attached the version of it that I use.

Mind sharing what other cards you've got on the way?

---------
@pshipkov
Agreed. The NCR is a bit hard to work with, but a beast it certainly is. It has the chops to compete with some later ISA cards in a very favorable way. I would very much like to see how the later 77C32BLT does. It was sold as the Alaris Skyeagle for VLB.

I will definitely be adding some 4x40 results for the 77C22E, once the beast awakens.

I recall reading in PC Mag about a 6MB VLB version of the AGX015. I have never seen one either. I believe some variant of the Hercules Graphite VL was capable.

Speaking of, there are a few weird VLB cards out there that would be interesting to test. The Alliance ProMotion 3210 made its way onto a Miro card, and there's the MediaVision ProGraphics 1024, among the other weird offerings. Maybe someone has numbers for these? I am very curious about the Alliance part.

Thanks for the reupload! I hope that it was just an issue on my end; going to take a look later and see.

--------

EDIT: For a relatively late 486 board, sold in December of 2000 and dated on the chipset at 1999, this board has much to offer. Onboard VLB 2.0 EIDE (for which drivers are unfortunately nonexistent), sturdy build quality, and undoubtedly excellent stability. Uses 95 AMI WinBIOS core, so standard there. Seems to be the only known board implementing this Chips F84041 chipset, a late and odd entry to the market. Lacks support for >256k L2 inexplicably, despite chipset support. BIOS settings are limited and tricky to get right; sometimes board will not boot without timings at auto. Will have to investigate further with -10ns cache.

Here's some first-round numbers for the Chips at 4 x 40, minus LW3D, which I have not yet tested:

Most timings are set to auto, otherwise board won't boot. This is probably related to L2 timings, hinted at below. Will investigate further.
Otherwise, 8-bit and 16-bit memory WS are set to 0, and 16-bit/8-bit I/O recovery are at the lowest permissible (2 and 0 clk, respectively).

ARK1000VL
Wolf3D: 132.2 fps
PC Player: 22.6 fps (6.9 fps @ 640x480)
Doom: 61.5 fps
Quake 1: 16 fps
WinTune 2: 6358 KPixels/second
3D Studio R3: 130 seconds

S3 Vision864 (Bahamas64)
Wolf3D: 122.1 fps
PC Player: 22.4 fps (7.5 fps @ 640x480)
Doom: 60.3 fps
Quake 1: 15.9
WinTune 2: 6438 KPixels/second
3D Studio R3: 130 seconds

Impressions: Something is clearly wrong. At 33MHz FSB, this board shines (numbers to follow). Seems that the BIOS is timed conservatively; potential is here. Some cajoling, faster cache, and time will be necessary. Frustratingly, something happens to VLB timings/WS at 40MHz, likely BIOS-related. Vision864 is leagues ahead of ARK in acceleration, but tables are turned here and the fight is even. For now, will migrate to a faster platform. Tests of S3 864, 868, and Trio32/64 on such will follow.

My understanding about Ark1000 is that it has efficient core and very tight memory timings.
Don't think there is anything radical compared to the rest of the graphics cards in its class.

---

Alaris Skyeagle VLB:
Didn't know about this card until now.
Very interesting.

---

About the Chips Predator benchmark:
First of all - appreciate the time spent on this.
Looks like it is a bit slower than Asus VLI, but still pretty good.
Don't worry about LW3D - the test takes long time and if you feel the board is stable - it won't reveal something extraordinary given the rest of the numbers provided.
This feels like a good all-rounder assembly for sure.
I am going to link your post in the register on the first page, but can you do me a favor - either attach a picture of the motherboard to the above post, or link to the page you provided early with the manual and stuff.
Also, can you clarify in the same above post your BIOS timings - is everything on max, or if there were exceptions - list them out.

---

Curious if you can repeat/confirm the better metrics with the N9 BIOS for S3 Trio64 VLB.

---

Thanks for the mclock util.

@Chadti99

Do you happen to have a N9 BIOS for the Vision868? I have a Miro 20SD that I am tinkering with.
It seems on average a hair slower than the Paradise 864. Conservative memory timings seem to be the issue.
Troubling, to say the least.
Going to try your BIOS later tonight and see if I can replicate pshipkov's results. Seems likely, but curious as to how Diamond v2.02 compares.
I will dump my Diamond BIOS as well.

---
@pshipkov
Re: ARK1000, agreed. Seems to be standard in design, but cost-saving measures are clear compared to later S3/Cirrus/ATI cards.
Non-integrated RAMDAC and clockgen is interesting. Image quality is average, at best.
ARK PCI offerings seem to be equally "cheap," but certainly bring the punch.
Curious as to how ARK1000VL was marketed, budget card or speed demon? Seems like my Bali32 is late 94/early 95.

---
The Skyeagle seems interesting. I think one/two others made such a card w/ this chip.
Not sure if the 77C32BLT is upgraded core, but only notes on it are "Accelerated, supports VLB. Some neat features."
77C22E is not fully accelerated. Seems like a great leap. Curious about "neat features."

---
Re: Predator
My pleasure! I have updated the post. I think I will add some more info over time but am going to wait on faster cache to arrive.
Am also going to see about >256k L2. Frustrating and unreasonable to have 10 DIP-32 sockets and no BIOS support.
Considering calling Corvalent; they are the modern rebrand of APC. Perhaps they still offer support, given this is an industrial product.
BIOS stubbornly reports 64k, disabled when anything greater than 256k is inserted. Perhaps this can be patched?

---
MCLK is pretty interesting for early "overclocks." Reminiscent of my S3 805i w/ hardware-selectable MCLK.
Seems to give a little boost, but architectural limitations are at play in all cases.
On other note, card is a mystery. Why adjustable clock on a "budget-model" 805i?
Maker of card is unknown. Will have to delve further into its history at some point.

---
Might be worth trying XVESA util for mem reads/writes after playing with MCLK.
I use VESA Mode 101, but all modes are listed. Nice little utility, if you haven't messed with it before.
Attached below, for reference.

---
Last note and question: I am considering picking up one of the following VLB cards for testing: S3 Vision964-based Miro 20SV, or Orchid Kelvin64 GD5434.
Have heard that GD5434 is a bit underwhelming on VLB and is better-suited for late ISA shenanigans.
Saw some of your benchmarks, and this seems to be the case. Curious as to thoughts?
Re: Miro 20SV, I want to see if my 20SD can run the Diamond BIOS or N9 BIOS first, and then see if the 964 can run its equivalent Diamond BIOS.
Miro BIOSes seem to favor stability and run conservative clocks on memory. Unhappy with prospective DOS performance otherwise, and may wait for another fish to catch.

That S3 805 is wacky. I think i have seen a picture of it somewhere else. Do you have any perf metrics for it ? It will be interesting if it can go past the standard models.
Found some arbitrary Windows tests here: https://books.google.de/books?id=PITtFPwTaWwC … 7C32BLT&f=false
The NCR77C32BLT is in the mix as well.
That page you provided offers good list of 1994 stuff. Some of the notes are funny. NCR77C32BLT has "neat features". Mmmkay. : )
What other VLB cards you have over there ?

Remember trying long ago mclk on Trio64 VLB/PCi cards - didn't observe change in perf.
Didn't look further into it since then.
Are you able to get anything out of it on a good/fast S3 card ?

About the Predator 747 post:
From time to time i tweak previous posts with corrections and additional details.
Looks like you will be going through 1-2 more edits at least.
The current performance of the motherboard with 256Kb L2 cache is close to Asus VLI with 1Mb L2 cache. So if you succeed to upgrade the buffer size things may get interesting.

Cirrus Logic GD-5434 is a good chipset.
The Orchid Kelvin 64 is a good implementation.
No complains to both. Cannot think of anything negative. Works as advertised.
Pretty much on par with the better VLB cards in interactive DOS graphics.
A bit behind in accelerated Windows GUI.

If you ask me - share more details on S3 Vision VBL cards.
I don't have any because passed multiple times in the past on them in favor of the later S3 Trio64 model.
It will be interesting to confirm or disprove difference in Miro/N9/Diamond BIOSes.
@Chadti99 also has some S3 Vision VLB cards, but don't know if he is that interested in messing around with the BIOS microcodes.

@Chadti99
Thanks for the N9 BIOS! It worked very well.
Results are below. It is certainly faster.

---
@pshipkov

Re: S3 805. Very whacky. I just tested it today on the SuperEISA.
DOS numbers are impressive, Windows above/around average.
I've got a Stealth24 equipped with a "regular" 805-p for comparison to test soon.

Interesting tests you found! Thank you for sharing.
Seems the 77C32BLT is decent. It will be interesting to see how it compares.
I read something today saying the Alliance ProMotion 3210 is also pretty fast; should be in the same class of late/weird VLB cards.
That page is pretty funny, but definitely has some nice reference. Still can't help but wonder where some of the data came from.

I have a few random cards, but mostly the usual suspects: Trio32, Trio64, S3 864/868, Mach32, S3 805/i, WD90C33-ZZ, UMC418F, Trident 9440-1, Chips F64300, ARK1000VL, and a boxed/unopened Mach64 VRAM VLB. I have the upgrade module for the last one as well and was initially planning to unbox it on YouTube. I would like to archive and preserve the manual and everything, and want to see that beast in action. I have a soft spot for the weird Mach cards.

As far as EIDE stuff, I have an EIDE2300Plus, a couple of other IDE adapters that I haven't really looked into, and two VLB ethernet cards. One is a BocaLanCard VLB, NIB, but I'll open and use that soon. I am from Boca so it's almost become a hobby to collect their hardware 😀 The other is stubborn and has an SMC91C92 chip. I need to try it in Win95.

I recall some change with MCLK, but some cards have "factory-optimized" BIOS memory timings, or so it seems.
Needs further investigation, but cards with faster DRAM seem to have tighter timings--except ARK1000VL.
All ARK1000VLs have 70ns RAM. Puzzling.
Planning on testing MCLK on Miro 20SD this week; initial numbers at 4 x 40MHz were not as hoped.
Could be due to 70ns EDO vs 60ns on 864, but unlikely. Numbers within a reasonable delta, but below 864.
Diamond 868 and N9 868 BIOSes from web did not work.

Re: Predator Chips
I plan on definitely optimizing things and adding some edits!
It's an impressive board, but I feel a bit let down compared to "the beast."
SuperEISA takes its lunch money and does so without missing a beat.
Increasing the L2 seems most reasonable. I plan to test at first with 256k 10ns cache to see what that brings.
The LS486E seems totally fast with 256k L2, maybe this is another board in that vein?

Thanks for the suggestions re: VLB cards. I think the 964 may be my next pickup, unless a 968 appears.
Seems that the 964/968 are a bit faster than the Trio64 in Windows at higher res. Interesting.
Curious too, whether ATI Mach64 w/ 4MB VRAM can handle a 964/968.
GD5434 is for another day. While I like Cirrus, it sounds like nothing too out of the ordinary.

I've been doing some BIOS swapping. The N9 BIOS worked on my SpeedStar64 DRAM T very well, but failed in Quake. This will be investigated further.
The Trio32 and Chips F64300 also could not handle Quake at full speed. Seems like they are not as well-buffered.

----
Initial Thoughts: For an early 486 board, with an unsophisticated, non-highly-integrated chipset, dating to 1992, this board is nuts. While the AMI BIOS uses the commonplace HI-Color 6/6/92 core, the board’s performance and implementation is brilliant. All aspects were carefully designed. The PCB is of top-notch quality, and many chips are socketed for ease. This board can take 1024k L2 cache and likes the "10ns" parts, running them at max timings w/out wait states at 40MHz. Can cache all 256MB (seemingly) of 30-pin SIMM RAM and allows for DRAM interleave. Provides WB L2 cache, but no L1 WB, sadly. No VRM, so VRM interposer is used for clockspeed and voltage adjustment. Board runs happily at 50MHz FSB+ and some users have reported 66MHz FSB success. Speed is easily adjusted via oscillator, and clockgens can be used instead. Only provides one VLB slot, but EISA makes up for the rest. Wide, fast bus with lots to offer, and easy configuration.

All timings were set to their minimum, all waitstates were disabled, and FSB was reported at 40.0MHz in ChkCPU.

Impressions: Overall, an excellent board. A top performer, for sure. Impressive metrics in synthetics and real-world benchmarks. The SiS406/411 chipset is potent, and offers much to the end user in terms of stability and reliability. This board can take anything thrown at it; 1024k L2? Check. VLB? Check. EISA? Check. Unfortunately, it is limited to 1 VLB slot. But the double-banked 30-pin RAM configuration enables 256MB of fully cacheable RAM, and no modifications are needed for any of this. This board can take clockgens, oscillators, and has few jumpers relative to most anything its age.

Easy, reliable, fast. Cream of the crop.
----
EDIT: pshipkov's notes:
"Couple of conclusions from the captured metrics:
Slightly behind the average at VGA graphics - Wolf3D and Doom. Predator 747 Chips does better here.
Top notch at CPU/FPU intensive tasks - PC Player, Quake 1 and 3D Studio.
Top notch at accelerated Windows GUI - WinTune 2."

Agreed on all fronts. VLB performance is not 100% w/r/t DOS VGA.
Curious as to whether design, implementation, BIOS, or something else.
Will investigate and report.
----
3DStudioR3: 110 seconds

EDIT :Added more cards, including S3 805-p, Trident 9440-1, ALG2228.A, and UM85C418F-GP.
EDIT #2 :Added more cards, including GD5426 and WD90C33-ZZ. Got Miro20SD stable in Win3.1 for GUI testing. Strange results.
EDIT #3 :Added Tseng ET4000/W32P. Re-ran tests with I/O Recovery increased from lowest at 3 BCLK. Helped some boards clear Quake. Scores annotated below.
EDIT #4 :Added Diamond SpeedStar Pro SE GD5430. Not impressed. Acceptable/Trio32 level in Windows, below-average in DOS.
EDIT #5 :Added charts. Finalized WinTune2 score for ARK1000VL.
EDIT #6 :Added S3 Vision964 numbers (Miro 20SV). Need to update charts...
EDIT #7 :Added S3 new S3 Vision868 numbers (Diamond Stealth Video 2000 Series). Mach32 DRAM tested. (Graphics Ultra Plus). Mach64 Graphics Pro Turbo w/ 4MB VRAM tested. New chart with all data appended.

Summary Charts (Continuously Updated):
Individual Charts are at Bottom(will be updated in a few days)

ARK1000VL (1ws enabled)
Wolf3D: 123.6 fps
PC Player: 26.8 (10.3 fps @ 640x480)
Doom: 57.7 fps
Quake 1: 17.9 fps
WinTune 2: 7634 KPixels/second

Gainward Cardex Challenger (Tseng ET4000/W32P)
Wolf3D: 121.2 fps
PC Player: 26.8 fps (10.5 fps @ 640x480)
Doom: 57.5 fps
Quake 1: 17.9 fps
WinTune 2: 10197 KPixels/second

MiroCrystal 20SV (S3 Vision964 w/ Diamond Stealth64 VRAM H BIOS)
Wolf3D: 114.5 fps
PC Player: 26.2 fps (10.1 fps @ 640x480)
Doom: 56.0 fps
Quake 1: 17.7 fps
WinTune 2: 13525 KPixels/second

S3 Trio64 (Stealth64 v2.02)
Wolf3D: 120.3 fps
PC Player: 26.7 fps (9.7 fps @ 640x480)
Doom: 57.0 fps
Quake 1: 17.9 fps
WinTune 2: 12495 KPixels/second

Diamond Stealth 64 DRAM (S3 Vision868) (Diamdond Stealth Video v1.01 BIOS)
Wolf3D: 122.1 fps
PC Player: 26.8 fps (9.8 fps @ 640x480)
Doom: 57.1 fps
Quake 1: 17.9 fps
WinTune 2: 13347 KPixels/second

S3 Trio64 (Stealth64 N9 FX530 BIOS)
Wolf3D: 121.8 fps [+1.5]
PC Player: 26.8 fps (10.2 fps @ 640x480) [+0.1, 0.5]
Doom: 56.9 fps [-0.1]
Quake 1: DNF [Unstable at even 11BCLK I/O Recovery]
WinTune 2: 11699 KPixels/second [-]

MiroCrystal 20SD (S3 Vision868) -- NEED TO REVISIT
Wolf3D: 114.5 fps
PC Player: 26.1 fps (9.6 fps @ 640x480)
Doom: 57.0 fps
Quake 1: 17.6 fps
WinTune 2: 10868 KPixels/second
**Are there 868-specific Win3.1 drivers? The 868 ones I found seem to be the same as the 864 drivers.**
**EDIT: Found original Miro drivers. Quirky, meant for CRTs, but work well. Offer a "max performance mode" that seems unstable.**

Paradise Bahamas 64 (S3 Vision864)
Wolf3D: 121.8fps
PC Player: 26.8 fps (9.7 fps @ 640x480)
Doom: 57.0 fps
Quake 1: 17.9 fps
WinTune 2: 11955 KPixels/second

Unknown LA805i (S3 805i-p, MCLK at 55MHz)
Wolf3D: 119.2 fps
PC Player: 26.5 fps (10.5 fps @ 640x480)
Doom: 57.2 fps
Quake 1: 17.8 fps
WinTune 2: 7893 KPixels/second

Diamond Stealth24 (S3 805-p) (Spea 7 Mirage BIOS)
Wolf3D: 104.4 fps
PC Player: 24.6 fps (8.4 fps @ 640x480) [VESA TSR required for SVGA]
Doom: 50.6 fps
Quake 1: 16.9 fps* (I/O Recovery 11 BCLK)
WinTune 2: DRIVER ISSUES
**Card will NOT operate at 40MHz with default Diamond Stealth24 BIOS v2.00**

Boca Research Voyager (Chips F64300)
Wolf3D: 112.7 fps
PC Player: 26.5 fps (10.3 fps @ 640x480)
Doom: 53.7 fps
Quake 1: DNF
WinTune 2: 8537 KPixels/second

Diamond SpeedStar SE (Trio32 v1.01), Diamond SpeedStar SE (Trio32 v1.02)
Wolf3D: 121.5 fps / 121.8 fps
PC Player: 26.8 fps (9.9 fps @ 640x480) / SAME
Doom: 56.9 fps / SAME
[Change in Doom min details of 0.4 fps from 171.4 to 171.5]
Quake 1: 17.8 fps / 17.8 fps* (I/O Recovery 11 BCLK)
WinTune 2: 6096KPixels/second

ATI Mach32 Graphics Ultra Pro (2MB VRAM)
Wolf3D: 101.7 fps
PC Player: 26.5 fps (10.4 fps @ 640x480) [VESA TSR required for SVGA]
Doom: 49.1 fps
Quake 1: 17.7 fps
WinTune 2: 9458 KPixels/second

ATI Mach32 Graphics Ultra+ (2MB 60ns DRAM)
Wolf3D: 101.7fps
PC Player: 26.4 fps (10.3 fps @ 640x480) [VESA TSR required for SVGA]
Doom: 49.2 fps [+0.1 fps]
Quake 1: 17.7 fps
WinTune 2: 7090 KPixels/second

ATI Mach64 Graphics Pro Turbo (4MB VRAM)
Wolf3D: 95.6 fps
PC Player: 26.3 fps (10.1 fps @ 640x480)
Doom: 46.2 fps
Quake 1: 17.6 fps
WinTune 2: 12795 KPixels/second
NOTE: This card seems to have compatibility issues on this board, yielding lower performance.
These issues are related specifically to the memory aperture feature, as well as EISA CFG issues w/r/t this particular card.
Will revisit later.

Unknown "Union" Trident 9440-1 (2MB DRAM)
Wolf3D: 122.7 fps
PC Player: 25.6 fps (10.5 fps @ 640x480)
Doom: 57.5 fps
Quake 1: 17.4 fps
WinTune 2: DNF -- Need 9440-1 specific drivers? AGI ones do not work.

Avance Logic ALG2228.A (1MB 45ns DRAM)
Wolf3D: 107.8 fps
PC Player: 26.4 fps (10.3 fps @ 640x480)
Doom: 51.9 fps
Quake 1: 17.6 fps* (I/O Recovery 11 BCLK)
WinTune 2: 5588 KPixels/second

Diamond SpeedStar Pro SE (Cirrus Logic GD5430)
Wolf3D: 82.9 fps
PC Player: 24.7 fps (9.0 fps @ 640x480)
Doom: 39.9 fps
Quake 1: 16.9 fps
WinTune 2: 6196 KPixels/second

UMC VL-1AV UM85C418F-GP (1MB 80ns DRAM)
Wolf3D: 114.8 fps
PC Player: 26.7 fps (N/A @ 640x480) [VESA TSR required for SVGA (Does anyone have?)]
Doom: 53.7 fps
Quake 1: DNF
WinTune 2: 1526 KPixels/second (Does this card even have GUI acceleration features?)
**HWINFO reports 16-bit external bus for this card. Would be interesting to see if true.**

Boca Research Super-X (Cirrus Logic GD5426)
Wolf3D: 102.3 fps
PC Player: 24.6 fps (9.0 fps @ 640x480)
Doom: 50.7 fps
Quake 1: DNF
WinTune 2: 3427 KPixels/second

DFI WG-6000VL (WD90C33-ZZ)
Wolf3D: 122.1 fps
PC Player: 25.9 fps (10.5 fps @ 640x480)
Doom: 57.5 fps
Quake 1: 17.5
WinTune 2: 5434 KPixels/second

I'm really glad you had a chance to do this. You own a bunch of DRAM based S3 cards that I don't have, and now I don't have to waste my time trying to get them. The results using the #9 BIOS are interesting. I'll have to test that on my cards. I have a Hercules Terminator64, and a Paradise Bahamas 64 with that chipset.
My tests with ARK1000VL on the SuperEISA were interesting. I have the common version with 2theMax branding. I think it's identical to your card. That one only works on the 1WS setting. I also have the Octek version, which *sometimes* will boot on the 0WS setting, but there is no performance difference whatsoever. As far as I can tell, there is also a jumper on the SuperEISA which controllers VLB waitstates. I believe mine was set to 0WS, and it takes priority over whatever the jumper on the card is set to. Changing it to 1WS actually gave me lower scores. So, the point I am trying to make is that even though you have 1WS set on your card, it's actually running 0WS as set on the board.
As far as 66MHz goes, I didn't really have much luck there. I can get the system to boot and run some benchmarks, but it's not stable. I believe I had better luck with 60MHz, and was actually able to boot into Windows NT4. The ARK1000VL and the ET4000W32P (45ns DRAM) both worked to some degree at 66MHz 0WS. I don't think I tried 1WS at 66MHz (use the motherboard jumper), so you should experiment with that.

BTW, my Octek ARK1000 has fully socketted memory, and I was able to change it all to 45ns. I didn't recall it making much 0f a difference in terms of overclocking.

Great stuff !
Will link it to the front page. This reminds me about two posts from @Chadti99 for MSI-4144 and Biostar UUD @200MHz that i wanted to add there as well.

---

Let's see where this hardware overclockable S3 805 goes compared a regular one.

---

Always wondered what is the purpose of VLB LAN cards.
Any hints from trying them ?

---

That board certainly looks ahead of its time.
Remember seeing @annoymous coward's post, but he didn't provide much details about it other than "works well". ; )
Also, remember seeing non-VLB motherboards based on that chipset, but discarded them because of my self-imposed search criteria - 72-pin RAM only, socket 3 only, mid-to-late chipset versions.
Perhaps i got too tight of a pattern.

What are the empty sockets next to the L2 cache chips for ?
Cannot make sense of them.

Crystal oscillator >> clock generator. So much easier if you want to mess around with overclocking.
+1

Tightest BIOS timings + completely stable.
+1

Nice fat memory bandwidth number (speedsys).

Couple of conclusions from the captured metrics:
Slightly behind the average at VGA graphics - Wolf3D and Doom. Predator 747 Chips does better here.
Top notch at CPU/FPU intensive tasks - PC Player, Quake 1 and 3D Studio.
Top notch at accelerated Windows GUI - WinTune 2.

What is going on MiroCrystal 20SD (S3 Vision868) ?

Looks like S3 Trio64 is better at everything than Vision 864 and 868.

I don't have a soft-spot for ATI cards. Looking at the numbers - that stance only hardens. : D

So, what happens if this board gets pushed to 3x60, 3x66 or 4x50 ?
I see good things there.

pshipkov wrote on 2022-09-13, 18:23:

Nice fat memory bandwidth number (speedsys).

Does anyone know how this figure is derived, or the supposed significance of it?

It never seems to correlate with actual throughput performance in any way I can observe

You are right - it means exactly nothing.
Just noticed the big number. That's all.
If i remember correctly - it is related to bigger L2 cache usually.

I remember there was a thread about those Boca VLB LAN cards. All I recall was that the transfer rates were worse than ISA.

When you guys are testing other graphic card BIOSes, did you also ensure soft resets via ctrl-alt-del were functional? In the past, I've found sometimes the BIOS swap leads to the inability to soft-reset.

The N9 BIOS on Diamond S3 Trio64 VLB is flawless as far as i can tell.
I remember you saying that you had problems with the soft-resets.

Maybe these VLB LAN cards were used for some industrial purpose - i don't know.
Need to search online to see what others are saying.

Btw, never knew that Boca Systems actually is named after a town in Florida.
Even their present web site looks retro. I like it.

pshipkov wrote on 2022-09-14, 21:09:

The N9 BIOS on Diamond S3 Trio64 VLB is flawless as far as i can tell. I remember you saying that you had problems with the soft […]
Show full quote

The N9 BIOS on Diamond S3 Trio64 VLB is flawless as far as i can tell.
I remember you saying that you had problems with the soft-resets.

Maybe these VLB LAN cards were used for some industrial purpose - i don't know.
Need to search online to see what others are saying.

Btw, never knew that Boca Systems actually is named after a town in Florida.
Even their present web site looks retro. I like it.

Could you remind me - does your Diamond have the 86C764-P or the 86c764X chip? I'm sure you recall this thread, Stealth64 DRAM T VLB Trio64 86c764-P vs. 86c764X

And to re-confirm, you had specifically tried a ctrl-alt-del soft reset with the Number9 BIOS on your Diamond and found no issue?

it is 86C764-P. pictire on the first page of this thread.

yes. c-a-d works with it.
can tripple check later today juuuust in case. but it is fine.

If you are going to check the soft reset, can you ensure that it boots up to the DOS prompt? I think in some cases, the soft reset appeared to work at first, but would not bring me to the DOS prompt.

@pshipkov
Thanks! I hope this info has been helpful. Testing this stuff is pretty fun!
Just following the lead of all the great work you guys have been putting in!

Compared to the regular S3 805, the S3 805i is infinitely superior.
Updated the previous post and added a few more cards; 805i is clearly a better implementation.
Found something funny: default Diamond BIOS hates 40MHz, Spea V7 Mirage BIOS works great (and is faster) w/o corruption.
S3 805 is not bad, but 805i is clearly superior. Memory interleaving has its benefits, even in DOS (seemingly)?

-----
Re VLB LAN:
I think these cards are just plain interesting because VLB doesn't add much to the table.
Most are half-duplex, and 10mbps. Not sure if 10/100 models exist, but it's possible.
They are oddities, but potentially useful in not bogging down the ISA bus.
All advantages are null vs EISA, since EISA supports 10/100, busmastering, plug and play, etc.
Not sure about stability at 33Mhz+, would be interesting to test.

My SMC91C92-based card is a nightmare. The only packet driver is for the SMsC 91c9x, and works and then freezes. Card seems okay and lights up; can retrieve IP from my network. Not sure what the story is there. Probably need another packet driver. Hence Win95 adventures. Seems that chipset was used on Apple Mac cards and in PCMCIA cards. Curious to see it on VLB.

The BocaLan-VL uses the typical AMD PCnet32. Good chip, but only supports half-duplex operation when not in loopback mode. Good ISA cards are better at hitting the 10mbps barrier.

-----
Re: SuperEISA
It is a good board all around. I think the chipset is very "ahead of its time," especially for something dating to 1992 and lacking WB L1.
Probably not a bad setup even on EISA-only boards. Certainly potent and stable at 50MHz.

It had a few competitors, including some VIA setup that I think was once a Symphony design (I am not sure about this, @AnonymousCoward is the knowledgeable sage here), and OPTi's offerings. There are probably others that I have yet to learn about, but I would love to find one of these VIA/Symphony boards; only FIC produced one as the VE-V.

OPTi had the 691/696 "Hunter," but that was featured on almost nothing besides the Acer J3. Their other main offering was the 681/22/6/7. I have heard decent things about this chipset, but it only supports 512k L2. It is also rated for 50MHz on some boards like the TMC PET48PN.

I am not sure about the empty sockets either, but think they're for 400 mil cache chips. @AnonymousCoward, any ideas?

Thanks for the observations!
Adding them to the above post. I agree completely.
Curious as to if VLB is setup in some sort of bridge config w/r/t EISA on this board; mr_Slug's chipset database suggests it is possible.
Could be overthinking things here.

-----
Re: MiroCrystal 20SD and VLB boards
I tried MCLK and tightened some timings.
Nothing improved; scores worsened. This is uncharacteristic. S3 864 has responded well.
Curious if bad implementation?
Ram is 70ns only; this could be nothing though. Card is fast and responsive, but seems unhappy on this board.
Will retest on Predator 747 once things are in order there.

Vision864 seems w/in reasonable range of Trio64. It is pretty even.
Windows is marginally faster, but DOS tests are neck and neck. 864 wins a few even.
Trio64 is definitely the "evolution" though. Integrated RAMDAC and clockgen.

I've always felt ATI cards are wonderful in 2D.
Sharp, crisp image and neat "crystal" font in DOS.
Mach32 is great for ISA and for ~1992/3 VLB. Certainly interested to see how Mach64 kicks things up.
Definitely not screamers, but reliable and different.

Will have to try 4x50MHz first, given the spec. Seems that this board does well there.
Not sure if any of my Am5x86s can do it. Will have to see.
My interposer only gives ~3.4V, and would need a trimmer mod.

Re: Boca Research
Here's a fun one: the IBM PC was designed at IBM in Boca Raton, FL. I have no clue why IBM chose such a random place, but Don Estridge's team was there. One building where some of the R&D was done is now a middle school, and some of the "BRIC" IBM campus has been sold off. I believe a little bit of IBM remains here. From what I have read, Boca Research started as a group of ex-IBMers selling components for the IBM PC. They did quite well and had a plant in Delray Beach. All of their products were made locally, and they built quality products. I believe modems were their top seller. Any product with BRI in the FCCID is a Boca product. I unboxed a NOS BocaLAN ISA card with an AMD chip the other day; it had slight corrosion from humidity or water during storage, but I touched up the solder joints and it worked w/o problems. Anecdotal, but pretty wild.

-----
@AnonymousCoward
I appreciate the kind words! The #9 BIOS is worth a go. I am curious why it was not stable in DOOM. It seems that some cards on this board are not quite 100% stable at 40MHz w/o a wait state. Speaking of, do you know which undocumented jumper adds one? Is it JP9, near the SiS85C405 chip?

I think the 2theMax board is the same as well. I have only seen a few unique designs and think pshipkov's is the most different I've seen from the "stock" configuration. That is interesting regarding the difference between the 0/1 WS Jumper on the card. I wonder what it actually does, any ideas?

Good to know re: 66MHz. I want to try it at some point, maybe with an IBM 5x86C at 66x2, but 60Mhz seems a bit more reasonable. I am imrpessed at the ARK and ET4000W32p taking the speed so well. Seems logical, though, given each card's known excellent performance and reliability at higher-speeds in the VLB spectrum. I recently read somewhere of a graphics board advertised and sold as "VLB 2.0 compliant" with only respect to stability at 50Mhz. It may have been the Orchid Kelvin64.

EDIT: I wanted to ask but almost forgot--I am running this board with 20ns 64k x 4 TAGs.
Any thoughts on if an upgrade is worth it/necessary?
It seems difficult to find these chips in speeds any faster than that these days.
-----
@maxtherabbit
I am not sure what the SpeedSys bandwidth means. I know that in older versions there was some arbitrary memory system subscore thing, but I am not sure if this is even a truly reliable metric of any kind. I have noticed that on faster boards, it's generally on the higher side, but even the Predator 747 is only at 90-something on it, and that's a fast board. It'd be neat to find out what it actually means.

-----
@feipoa
I tested this a few times with my card. I have a 86C764-P chip on my board, with BIOS v2.02. The date code is 9506. I am using 60ns RAM on the board. The two soldered chips are marked USA 4P16257DJ-06, and the two socketed chips are SIEMENS HYB51417BJ-60 chips. Otherwise, my board is identical to pshipkov's. I found no issue with soft resets; in fact, I use them quite often in testing. I found the board totally stable with the #9 BIOS, and never had an issue getting to DOS or Windows. I have not tried Diamond BIOS v2.01 or v2.09 on this board. Do you think either is worth a shot, to see if I can replicate this issue? I am curious as to what might cause it.

With regard to the BocaLAN card, I think I stumbled upon that thread the other day. Performance is disappointing, for sure. I wonder if it's due to the half-duplex 10mbps limitation of the AMD PCNet32 chip? Truth be told, I picked it up solely as a curiosity and out of interest in Boca Research. It's a shame they're no longer around. It would've been neat to see that Delray Beach plant.

WJG6260: yes, I would be very interested to see if you using the Diamond v2.09 BIOS with the -P chipset allows you to soft reset.

Next time I have my IBM BL3 system out, I'll have to try the Number9 BIOS on it to see if there's any improvement. That system is only at 25 MHz FSB, so wait states at 0 should be a problem and any performance enhancement is welcome.

If you have something common like a 3Com Etherlink III ISA card, it would be of interest to compare that with the BocaLAN VLB card in Windows 95.

@feipoa
I will give this a shot and let you know.
I have my TL866-II and a W27E257-12 at the ready.

Is your BL3 system using the Alaris Cougar, or a drop-in upgrade on some other VLB 386 board like the ALi Panda?

I really like your idea re: testing the BocaLAN. I have a 10/100 ISA Intel card, and a number of random 10mbps cards.
Unfortunately, none of mine are 3Coms, but I am pretty sure that the HP card I've got offers full duplex.
Any recommendations on how to go about testing? Large file downloads off of a NAS or something like that?
I am not too experienced testing network cards, and would like to learn more and see what the deal with this VLB model is.

-----
Here's a random funny I had to share:
Thinking about EISA and VLB ethernet cards, I ran across this monstrosity.
It seems fitting for a VLB ethernet card 🤣
Very PCChips, but interesting. Curious about stability and, at any rate, use case.
"SUPER VESA SERVER"? That makes no sense. Why not use EISA?
I guess the answer is two lines above. PCChips.

Yes, soft reset with N9 BIOS works fine.
Best S3 Trio64 VLB so far.

Nice Boca Research history info. Didnt know that. I can see state government incentives at play for IBM to get in the area.

Feipoa is with Ali Panda.

That PC-Chips EISA board - W.T.F.
Are these really VLB slots ?
They must be running at 2.5MHz. : )
So, who has this assembly ?

Btw, i was always "afraid" of EISA boards. Perceived them as rigid and as a transitional tech that didnt stay for long, but you are changing my stance.