mpe wrote:

This is from 2.0 from 1993. I know, what you want to say. But 1.0 spec revision is nowhere to be found and there is a plenty of evidence the 386DX/SX bus support has been always there and is implemented by targets as documented by various VGA controllers datasheets.

Can you link the document then? I already asked you.
Because as I say, it doesn't seem possible to do 16-bit transfers at all, given that the address bus doesn't have the two lowest bits.
I'd like to read for myself what the spec says, and see if they modified the pinout on the bus to add an extra address line required for 16-bit.
I'm reasonably sure you're not reading the spec correctly, and it does in fact not support 16-bit transfers (unless you mean on the ISA-part of the bus, which is not clocked at the localbus speed, so that is moot, it's just an ISA transfer. Even on a VLB card, the ISA portion is still used for interrupts and port i/o. Only data transfers are done via the localbus part. So any VLB-compatible chip will still have a 16-bit interface for the legacy ISA part).

VLB v2.0 Specs

Note the supported hosts.

my 386SX33 beats the SX16 for sure... and it has no Cache.

Have a NCR System with SX16 CPU and its reallly slooow.

VL-Bus Specifications

It looks like they are encoding missing A1, BLE2#, BHE2# signals into BE0#, BE1# and BE2#.

matze79 wrote:

my 386SX33 beats the SX16 for sure... and it has no Cache.

Have a NCR System with SX16 CPU and its reallly slooow.

Pigs alright. My generic 286/16 was a better setup.

mpe wrote:

It looks like they are encoding missing A1, BLE2#, BHE2# signals into BE0#, BE1# and BE2#.

Glad we opened that can of worms 😀
Anyway, after browsing the document, the conclusion is simple:
VLB is built directly upon the 486 interface, as I said.
In theory it is possible to implement it on a 386DX, but it requires various workarounds due to bus differences (one interesting example they name is that a 386DX duplicates the high-word to the low-word on 16-bit transfers, while the 486 only uses the high-word. The VLB bus expects only the high-word on the bus, so a 386DX would need the chipset to 'filter out' the duplicate low-word). So it is actually more difficult and therefore more costly to implement VLB on a 386DX (making VLB a 'too little, too late' option for 386 machines, which explains why so few systems support it).

And if you REALLY want to, you could even implement VLB on a 386SX, but this again relies on the chipset to do even more translations to make the transfers compatible (it would appear that such chipsets/boards actually do exist, but are mainly used for the IBM 486SLC CPU, which uses the 386SX pinout).

So much for the theory. Conclusion remains: with anything other than a 486, you're retrofitting your CPU to the 486 bus, to be VLB-compatible. The 386 bus may be the closest to the 486, so it may be the easiest, but it is by no means straightforward. There's some fundamental differences in data transfer that have to be fixed by the chipset.
The fact that these additional translations are documented in the spec do not convince me that VLB was really trying to be a CPU-agnostic standard.
Compare it to the PCI standard and see what a proper CPU/bus-agnostic standard looks like (or MCA if you want).

A big part of that standard also talks about bus mastering... But as we all know, that never worked in practice. The specification casually says: "Oh yea, and you'll need some bus arbitration", but that obviously never happened. So theory and practice are not necessarily the same.

Yeah i have Pentium 90/100 VLB too, if you choose Fast Mode it will run VLB 60/66Mhz..
so its only useable with a P75 @ 50 FSB or with Divider to 30/33.

While I hear what you say about implementation quirks and practice vs reality:

- I don't currently own one, but I've definitely seen a working 386 VL-Bus system (typically late boards that shipped with soldered 386 or SLC/DLC that also had 486 upgrade socket). Visit the 'System Specs" section on this forum to see some success sorties.
- I also happen to own a working bus-master capable VL-Bus card (Adaptec AHA-2842A) which also has VL-Bus 2.0 feature (write-back cache invalidation) in all write-back 486 system. I would eventually like to get another one to see if I can have two masters in one system 😀
- And finally I own a working post-486 VL-Bus implementation (NxVL board with Nx586 CPU) for which I am yet to find a VLB card that wouldn't work

So while I agree with you these features are rather non-mainstream, implementation can be screwed or problematic it is not that uncommon. Or you just need to be lucky 😀

This is interesting in a hair-splitting-contest kind of way, but my memory of that whole 386 / 486 / early Pentium era is, vendors promised a lot of stuff.

Like, sure you can upgrade the CPU so go ahead and buy that 486 today and run your Pentium CPU in it tomorrow. But in reality, it was almost never worth the cost or trouble. Performance and/or stability were questionable, you were saddled with legacy components (ISA video and controller cards when OEMs and new systems started shipping VLB), or the cost was high enough that it wasn't really saving that much over a proper upgrade path.

That said, it didn't seem like ANYone was onboard with VLB hanging around. Not even the industry advocates. It was a stop-gap that served its purpose and was jettisoned as soon as PCI was ready. They didn't even wait for a complete platform refresh, as evidenced by the availability of 486 PCI boards -- which I'm convinced were really only marketable because, by then, the writing was on the wall, and if you were still buying VLB cards, you were investing in a ship non-nonchalantly floating away from dock. So having an option to buy an affordable 486 system and start migrating to PCI seemed like a reasonable alternative.

The existence of 386 and Pentium VLB systems (neither of which I've ever seen in the wild) doesn't mean much. With enough clever glue logic, you could boot an XT from a USB BD-ROM, and drive a 1080p monitor from a PCIe graphics card. There's just no sense in doing so.

A lot of OEMs just included VLB video, with 512k vram allong with upgrade sockets to go up to 1meg in some cases and Hdd i/o on the mobo with out even needing a VLB expansion card. Only one out of six I have from various major OEMs from 93 to early-mid 95 had one external VLB slot. All shipped with 386DX/486DLC/486SX class cpus originally but the OPs upgraded to 486DX2/66 class cpus. These upgrades happened late '95 around the release of Windows 95RTM. Some of the systems I have came with Win95 upgrade CDs along with original OEM Dos/win3.x restore media.

mpe wrote:

- I don't currently own one, but I've definitely seen a working 386 VL-Bus system (typically late boards that shipped with soldered 386 or SLC/DLC that also had 486 upgrade socket). Visit the 'System Specs" section on this forum to see some success sorties.

I never said it didn't work at all. Just that chipsets/motherboards that even support it are very rare, and even when they support it, the performance gains and compatibility may be very marginal.

mpe wrote:

- I also happen to own a working bus-master capable VL-Bus card (Adaptec AHA-2842A) which also has VL-Bus 2.0 feature (write-back cache invalidation) in all write-back 486 system. I would eventually like to get another one to see if I can have two masters in one system 😀

Yes, even on 486es, VLB was often flawed. Most people wouldn't notice because they only used one or two cards, and 'got lucky', because the cards didn't use bus mastering.
Because there is no bus arbitration implemented, having two bus-master cards in one system is going to fail.
When you're only using one bus-master card, well, it's always the bus-master, so there's nothing to arbitrate.
Adaptec is a big name, so they probably designed their part right, and tested it properly. But the motherboard could still let it down.

I guess it also doesn't help that localbus was a thing before it was standardized in VLB. So there were probably a number of chips originally developed for some custom localbus solution, and later hacked around to be somewhat VLB-compatible.
I wouldn't be surprised if that 16-bit mode for 386SX doesn't actually work on quite a few devices, because either the designers never bothered to implement it at all (they targeted only 486), or because they never actually had any 386SX-VLB systems to test them on (because they simply didn't exist), so the implementation went untested.

mpe wrote:

So while I agree with you these features are rather non-mainstream, implementation can be screwed or problematic it is not that uncommon. Or you just need to be lucky 😀

Yes, you need to be lucky, even on a real 486.
That's not entirely exclusive to VLB either. In the early PII/Athlon era, you had the same thing with PCI, AGP, USB and such. Certain budget chipsets, notoriously VIA, didn't have an official license, and as such didn't have access to the official spec. So they just reverse-engineered everything, and built their own hacked-up implementation. The systems would kinda sorta work, but you had to be lucky. The same chipset worked better on one mainboard than on another because of different settings hardcoded into the BIOS. And some cards would work fine despite the flaws, other cards would not. And some stuff would work, just not very well.

One example I can think of is my Athlon 1400 with VIA KT133A chipset. I wanted to use it as a DAW, and I had a USB audio interface with ASIO support. I had lots of trouble getting the latency down below 150 ms (which is quite a lot). I would get random dropouts and whatnot.
When I tried it on a much slower system, a Celeron 1000 with Intel chipset, the same USB interface ran just fine, and I could get it down to 10 ms latency easily. Obviously the VIA chipset had issues.
I then wanted to see what the GeForce2 GTS in my Athlon 1400 would do in the Celeron 1000. Once again I found that the AGP bus ran much faster on that machine. I could actually get better performance out of the GPU as long as I didn't run into a CPU bottleneck.

I've seen similar issues with various VLB systems, because again they generally use cheap chipsets with poor implementations. I got lucky with my own 486DX2-66. It has a very fast VLB implementation, and can run zero-waitstate. I now run it overclocked at 80 MHz, and it still runs zero-waitstate just fine.
Most of my friends' VLB systems at the time couldn't run zero-waitstate, even at 66 MHz, even if we used my video card and IDE controller. And the performance difference was quite significant because of that.

Scali wrote:

One example I can think of is my Athlon 1400 with VIA KT133A chipset. I wanted to use it as a DAW, and I had a USB audio interface with ASIO support. I had lots of trouble getting the latency down below 150 ms (which is quite a lot). I would get random dropouts and whatnot.
When I tried it on a much slower system, a Celeron 1000 with Intel chipset, the same USB interface ran just fine, and I could get it down to 10 ms latency easily. Obviously the VIA chipset had issues.
I then wanted to see what the GeForce2 GTS in my Athlon 1400 would do in the Celeron 1000. Once again I found that the AGP bus ran much faster on that machine. I could actually get better performance out of the GPU as long as I didn't run into a CPU bottleneck.

That is very interesting. I have an Athlon XP 2100+ here running on a QDI KinetiZ 7E with KT-133A chipset running FSB at 133Mhz with 512MB SDRAM. Using it as a Yamaha Synth PC and generally old musics machine. I have done several benchmarks and the memory benchmarks indicate the system running at top speeds when compared to similar systems or even some with DDR memory. I was aware of the KT-133A being a hit and miss, giving vastly different performance depending on what board one is talking about. Is there any benchmark on whose base of numbers you would be able to immediately tell whether the system performs as it should?

derSammler wrote on 2019-10-27, 13:25:

AlessandroB wrote:

386 and 486 are not far clock to clock, probably 386DX40 is much faster.

No, they are quite far from each other if you compare the CPU only. A 486 is much faster. A 386 can only come close to a 486 if external cache is present. This, however, is no longer a fair comparison, as the external cache is provided by the mainboard with a 386, not by the CPU (unlike with the 486).

Bull comparing a cache less 486SX to a 386DX is fully valid

As “back in the day” no one buying a 486sx bought one with L2 cache EVER
Everyone buying a 33mhz+ 386DX always got l2 and almost always had svga as it was considered a “higher end offering” and SX was stuck as an expensive foot in the door value 486 with minimal specs on everything but the 486 chip.

I literally sold pallet loads of used PCs in a former life, not a single SX of any flavor have I ever seen with L2 cache
Nor did they even have anything but ISA expansion or VGA.

The only exception were very late sx2’s and 33mhz sx’s but by the time those released no one was buying Sx systems as evidenced by the fact I only have seen one of each in the many pallet loads I went through.

In the real world when you would compared the Packard Bell sx25 every up and coming bought with its outdated glacial hard drive, nearly defective Oak vga and not much else to the typical 386Dx33+ being sold at the same time it was no comparison
The higher end 386Dx
Always had
A tolerable hard drive
Cache
SVGA
and we’re almost always bought by professionals and hobbiest looking for the best bang for the buck

Using a real configuration of 386DX40 you typically found in the wild would almost always be a more pleasant experience than the gimped Packard Bell sx25 sitting in every moms home office in 1991 and it would run better because the other components weren’t obsolete and usually you had a bigger HD and more memory for less $$$

There are quite a few motherboards out there with 16 and 20MHz options (hell, I even have two So5 Pentium boards with those bus speeds...), but it's not that common. You might have better luck with old ISA-only 486 boards replacing the oscillator with a slower one.

dionb wrote on 2021-02-05, 00:46:

There are quite a few motherboards out there with 16 and 20MHz options (hell, I even have two So5 Pentium boards with those bus speeds...), but it's not that common. You might have better luck with old ISA-only 486 boards replacing the oscillator with a slower one.

changing the oscillator was the main way to change clock speed in the LIF socket 486 days

rmay635703 wrote on 2021-02-04, 23:41:

Using a real configuration of 386DX40 you typically found in the wild would almost always be a more pleasant experience than the gimped Packard Bell sx25 sitting in every moms home office in 1991 and it would run better because the other components weren’t obsolete and usually you had a bigger HD and more memory for less $$$

Hey, I've got one of those 'gimped Packard Bell 25sx' systems! 😜 Of course I fitted mine with L2 cache (including a dirty TAG SRAM), a much faster processor (Evergreen 5x86 upgrade) and an actually usable hard drive, and also bypassed the craptacular onboard Oak VGA with something less limited. Despite the whole thing being ISA based it doesn't feel too bad at all under Windows 95. It's just too bad Packard Bell couldn't be arsed to put L2 cache on it from the factory, and even if they would have, I sincerely doubt they'd fit the 64Kx1 SRAM needed to turn on write back.

In my defense, this was by far not my first 486. My first '486 class' system back in the day wasn't even a 486, but an Alaris Leopard-based hodgepodge cobbled together from the remains of several old computers and secondhand junk. For what it was, it sure felt a fair bit snappier than the 386DX-33 we upgraded it from.

mpe wrote on 2019-10-30, 14:29:

While I hear what you say about implementation quirks and practice vs reality:

- I don't currently own one, but I've definitely seen a working 386 VL-Bus system (typically late boards that shipped with soldered 386 or SLC/DLC that also had 486 upgrade socket). Visit the 'System Specs" section on this forum to see some success sorties.

I've got two 386 VL-bus boards now, both Alaris made: a Leopard with VLB and L2 cache just like the one I had back in the nineties, and a Cougar, its 386DX-bussed cousin. Both boards have some variant of IBM's interesting 486SLC/2 or 486BL/3 chips, and are fairly tolerant of VLB video cards, though the Leopard didn't take too kindly to my Chips F64300 VLB card. (The Cougar wasn't so picky fortunately.) It's my understanding that these were later 386 boards though, so maybe that had something to do with their relative feature sets and compatibility?

Just as a 2 cent comment, in Estonia, i've seen pretty common 486SX2-50 machines with 128kB of cache and VLB video and io cards.
So in some circumstances, these things happen. 😜

Hope you don't mind me reviving this thread.

I finally got myself to comparing the 486SX-16 against 386DX-40 (where everything else is equal). Something I wanted to settle for a very long time (at least since 1992 when I started wondering about the same 😀.

The results are not particularly surprising. It turns out the486SX-20 is roughly equivalent to 386DX 40 and the 486SX-16 is more or less equal to 386DX-33.

A short write up and more details are here.

Any comments are welcome.

Very Nice!!! It's cool that the 486SX16 lines up so close to the 386-33 and the 486SX20 lines up so close to the 386-40. That lends itself to an easy to remember 2:1 ratio for comparing 386's and 486's clock speeds.

What video card and storage did you use?
Can you do a comparison of ISA vs VLB from 386 to faster 486s? I was always curious to see evidence of a clear MHz inflection point where VLB really starts to become important rather than the 486-66dx2 rule of thumb.
Anyway you can quantify the % performance change for both the 386 & 486 chips from the OPTi 496XLC chipset compared against better performing chipsets?

mpe wrote on 2021-03-08, 14:03:

...and the 486SX-16 is more or less equal to 486SX-16.

Any comments are welcome.

Just imagine a world where it isn't ... 😉