Probably because the board has EISA slots. They probably had to add an extra chip for EISA support.
This thread is getting pretty cool. I'm glad to finally see so much interest in Socket4. I've always been fascinated by it because the technology was expensive when produced, and therefore less of it exists...especially if you're looking for something other than the standard Mercury based OEM stuff. I really want to see benchmarks on the more obscure chipsets like UMC, Forex and Contaq.

I have a new socket 4 puppy. PS/2 mouse, No Ide/ serial on board and removable cache chips.

With 512Kb cache it scores marginally higher that the DFI around 10 realtiks

I recently finished a socket 4 build here: Re: HELP - Voodoo on a socket 4 Pentium motherboard I'm using the Acer V12P motherboard. It scores 28.6 fps in GLQuake at 640x480 using the Pentium Overdrive 133.

A really nice board!

Talking about removable cache chips and extra cache. I upgraded L2 cache on my Socket 4 board listed earlier on this page from 256kB to 1024kB.

Before:

After:

I believe the new ISST branded chips from eBay (marked -10ns) are fake. However, at least they do have advertised capacity. I reused one of the original 32k Winbond chips as a tag SRAM as I needed to go from 8kx8 to 32kx8 to support 1M cache and there is only a short DIP socket.

It gives the motherboard a heathy boost in some benchmarks (Doom, PC Player). About 16%. It is enough for a P66 to go over my P75 in some tests...

I wonder if I can get even better performance by using faster tag cache chip...

I thought direct-mapped cached needed a 64kx8 TAG to use 1024K cache?

feipoa wrote:

I thought direct-mapped cached needed a 64kx8 TAG to use 1024K cache?

Isn't this for a 486? Pentium's cache lines are 32B instead of 16B and thus less tag RAM space is needed...

Anyway, this system only has DIP28 socket for tag ram and yet supports 1024k of cache.

You might be right. I checked some of my socket 5 board manuals and with 1024K and they call for one or two 32kx8 TAGs. The second TAG for dirty.

mpe wrote:

A really nice board! […]
Show full quote

A really nice board!

Talking about removable cache chips and extra cache. I upgraded L2 cache on my Socket 4 board listed earlier on this page from 256kB to 1024kB.

Before:

DSC_3222.jpeg

After:

DSC_4139 2.jpeg

I believe the new ISST branded chips from eBay (marked -10ns) are fake. However, at least they do have advertised capacity. I reused one of the original 32k Winbond chips as a tag SRAM as I needed to go from 8kx8 to 32kx8 to support 1M cache and there is only a short DIP socket.

It gives the motherboard a heathy boost in some benchmarks (Doom, PC Player). About 16%. It is enough for a P66 to go over my P75 in some tests...

I wonder if I can get even better performance by using faster tag cache chip...

Hey MPE,

nice! any chance you could post some benchmarks? keen to know doom realtiks etc 😀

Speaking of tag Sram the board i posted doesn't appear to have one at all 😒 any idea how that affects performance?

feipoa wrote:

I recently finished a socket 4 build here: Re: HELP - Voodoo on a socket 4 Pentium motherboard I'm using the Acer V12P motherboard. It scores 28.6 fps in GLQuake at 640x480 using the Pentium Overdrive 133.

I might have to dig my overdrive 133 out i'm curious now

Doom framerates with Spea S3 868 card.

P60 256k L2 - 35.82
P66 256k L2 - 38.6
P66 1024k L2 - 39.6

For context Pentium 75 on Intel Endeavour MB (430fx) gets 42.15 (with 256 PB COAST).

feipoa wrote:

You might be right. I checked some of my socket 5 board manuals and with 1024K and they call for one or two 32kx8 TAGs. The second TAG for dirty.

Yes. L2 cache coherency is quite a bit different between 486 and Pentium (MESI instead of as simple dirty bit, line size, etc.)

Not sure I've seen a dedicated dirty tag SRAM in a Pentium system, but there could be. But if I am correct some chipsets from this era used internal 32Kx1 MESI for dirty bit and combined internal/external SRAM to get required cacheable RAM limit. But I can't find any information about this particular ALi chipset.

Some Socket4 boards can support 2MB of L2 cache, but the only one I'm aware of has VLB not PCI. It was an OPTi board. Not sure what ALi can support.
I have a dual Socket4 board with 512kb of L2, but it can support 1MB too. I should really upgrade it to see if it makes it faster.

I finally got the Batman's Revenge board which is a kind of reference board for Socket 4. So I was able to calibrate my results against s.mouse's results (despite a slightly different VGA which is S3 868 in my case.

It looks like the MSI/ALi board at default settings is slower than 430LX, but given its expandable L2 and a plenty of tweaking options can be tuned to surpass the Batman board (and I still haven't exceeded potential as I can't run it stable without L2 wait states).

Being made by Intel, the Batman's Revenge has no tunable options. So its defaults is the best it can do.

Thx… similar benches looked for, i plan since a long time an early socket4 rig.

There is one issue with the fast 93 pentium machine theory, the early pci mainboards came out in the very end of 93 or very early 94. The pcmag tested the branded machines just in jan of 94, the pci cards what they got (prerelease cards!), came with acient ati mach32 or s3 928 chipsets. So i wouldnt expect so good dos fps performance (especially not the ark - in the bench - from 95).

A couple of months later intel released the socket 5 with p90 (p100 were not really available in 94) and the decent 64 bit pci cards (vision864, tseng w32p, etc).

So for 93 probably still the dx2-66 with a fast vlb card looked a good choice.

arncht wrote on 2022-12-23, 21:40:

Thx… similar benches looked for, i plan since a long time an early socket4 rig. […]
Show full quote

Thx… similar benches looked for, i plan since a long time an early socket4 rig.

There is one issue with the fast 93 pentium machine theory, the early pci mainboards came out in the very end of 93 or very early 94. The pcmag tested the branded machines just in jan of 94, the pci cards what they got (prerelease cards!), came with acient ati mach32 or s3 928 chipsets. So i wouldnt expect so good dos fps performance (especially not the ark - in the bench - from 95).

A couple of months later intel released the socket 5 with p90 (p100 were not really available in 94) and the decent 64 bit pci cards (vision864, tseng w32p, etc).

So for 93 probably still the dx2-66 with a fast vlb card looked a good choice.

I agree with the lack of options (especially PCI video) for a 1993 Pentium machine, DX2-66 + VLB was the best choice for DOS games until 1994.

However, I have seen quite a few P100 CPUs from 1994. I have an FDIV bug-free SX962 with 1994 datecode, so I assume many of the previous steppings with bugged S-specs (SX886, 910 and 960) must be even older, from before mid 1994. The P90 was much more common, but there were definitely a few P100s around.

TheMobRules wrote on 2022-12-24, 06:02:

arncht wrote on 2022-12-23, 21:40:

Thx… similar benches looked for, i plan since a long time an early socket4 rig. […]
Show full quote

Thx… similar benches looked for, i plan since a long time an early socket4 rig.

There is one issue with the fast 93 pentium machine theory, the early pci mainboards came out in the very end of 93 or very early 94. The pcmag tested the branded machines just in jan of 94, the pci cards what they got (prerelease cards!), came with acient ati mach32 or s3 928 chipsets. So i wouldnt expect so good dos fps performance (especially not the ark - in the bench - from 95).

A couple of months later intel released the socket 5 with p90 (p100 were not really available in 94) and the decent 64 bit pci cards (vision864, tseng w32p, etc).

So for 93 probably still the dx2-66 with a fast vlb card looked a good choice.

I agree with the lack of options (especially PCI video) for a 1993 Pentium machine, DX2-66 + VLB was the best choice for DOS games until 1994.

However, I have seen quite a few P100 CPUs from 1994. I have an FDIV bug-free SX962 with 1994 datecode, so I assume many of the previous steppings with bugged S-specs (SX886, 910 and 960) must be even older, from before mid 1994. The P90 was much more common, but there were definitely a few P100s around.

Sure… the p100 released, just the take out was very poor, so practically the p90 was the available top on the market (sept of 94). It same like at the ghz p3. They demoed in march of 2000, but the mass production started at the end of the year - and the situation was very different (the p4, faster athlons released).

To build a period correct pc, if we are talking about months, years, is not too easy 😀

Another issue with the socket4 builds, most of the boards sold in a 95 config - the socket7, cheap socket 3 cpus released, it was not a good offer anymore. So I would target the first part of 94 to build a decent socket4 for dos gaming, the 93 builds should be much more poor opti based mainboards with vlb/isa/eisa vga for server purposes.

arncht wrote on 2022-12-24, 07:26:

To build a period correct pc, if we are talking about months, years, is not too easy 😀

Another issue with the socket4 builds, most of the boards sold in a 95 config - the socket7, cheap socket 3 cpus released, it was not a good offer anymore. So I would target the first part of 94 to build a decent socket4 for dos gaming, the 93 builds should be much more poor opti based mainboards with vlb/isa/eisa vga for server purposes.

Yeah, I know what you mean... a Socket 4 build is difficult to fit if you're trying to stay somewhat period correct, I really like my Batman's Revenge but never end up using it in any builds. A shame, because it's very reliable and trouble-free... and Socket 4 in general is quite unique.