Great comparison mpe!

Regarding the pipeline issue on the POD - I can't imagine that both pipelines can work on a 4 byte bus, and if they do they would "interleave"(aka no idea how it would work).
Maybe it's possible to find out via the instruction complete signals of the cpu - each pipeline delivers one maybe this could solve the mystery(V-Pipe completion signal on pin B01 according to text).

Yes, Doom is not really "scaling up" nicely on newer processor but the demo of E4M2 in the Ultimate Doom is a VERY strong indicator of firepower for lower clocked Ps or high clocked 486s.

Do you know any method of improving the timings on these old Mercury boards maybe via a "hacked" BIOS?

amadeus777999 wrote on 2020-06-23, 15:30:

Regarding the pipeline issue on the POD - I can't imagine that both pipelines can work on a 4 byte bus, and if they do they would "interleave"(aka no idea how it would work).

Why should that matter? The Pentium has instruction cache. Every cache line fill happens in bursts of 4 lines (4x16 bytes) at a time. And x86 famously has a variable instruction size so the 64bit data bus doesn't give you any guarantees for number of instruction you can execute in parallel.

If you look into Pentium Overdrive Datasheet and all marketing materials you can find multiple references to super-scalar architecture and the V-pipe is present on all schemes, 2 integer or 1 fp instruction per clock etc. Then PODP compares favourably with the "normal" Pentium 90, performance counters update as expected, etc. I'd consider dual pipelines in PODP as pretty much guaranteed.

amadeus777999 wrote on 2020-06-23, 15:30:

Maybe it's possible to find out via the instruction complete signals of the cpu - each pipeline delivers one maybe this could solve the mystery(V-Pipe completion signal on pin B01 according to text).

The PODP has roughly 486 pinout so no IV pin .

mpe wrote on 2020-06-23, 16:11:

Why should that matter? The Pentium has instruction cache. Every cache line fill happens in bursts of 4 lines (4x16 bytes) at a […]
Show full quote

amadeus777999 wrote on 2020-06-23, 15:30:

Regarding the pipeline issue on the POD - I can't imagine that both pipelines can work on a 4 byte bus, and if they do they would "interleave"(aka no idea how it would work).

Why should that matter? The Pentium has instruction cache. Every cache line fill happens in bursts of 4 lines (4x16 bytes) at a time. And x86 famously has a variable instruction size so the 64bit data bus doesn't give you any guarantees for number of instruction you can execute in parallel.

If you look into Pentium Overdrive Datasheet and all marketing materials you can find multiple references to super-scalar architecture and the V-pipe is present on all schemes, 2 integer or 1 fp instruction per clock etc. Then PODP compares favourably with the "normal" Pentium 90, performance counters update as expected, etc. I'd consider dual pipelines in PODP as pretty much guaranteed.

amadeus777999 wrote on 2020-06-23, 15:30:

Maybe it's possible to find out via the instruction complete signals of the cpu - each pipeline delivers one maybe this could solve the mystery(V-Pipe completion signal on pin B01 according to text).

The PODP has roughly 486 pinout so no IV pin .

Sounds reasonable to me - totally ignoring the interface to the cache was pretty foolish of me. But the "narrow bus" causing a secret cut to the architecture was a bit too alluring.

FWIW I have been experimenting with processor upgrades for a not-the-optimal-platform 486DX2/66 (Compaq Deskpro/I) and comparing it to a stock Pentium 60 (Compaq Deskpro XE). In the below, I'm using the same ISA video card which is generally the fastest ISA DOS card I own.

What I've seen is that,
a) in all cases the PODP makes a marked improvement over the 486, of course
b) however, with the exception of Quake, that improvement isn't necessarily greater than what I'd get with the AMD 5x86-P75 upgrade
c) regardless, for an old platform like this that doesn't support WB and has limited RAM bandwidth, a 'stock' Pentium 60 is consistently faster.

amadeus777999 wrote on 2020-06-23, 19:20:

Sounds reasonable to me - totally ignoring the interface to the cache was pretty foolish of me. But the "narrow bus" causing a secret cut to the architecture was a bit too alluring.

Yes. The PODP83 not having 60/66 MHz and 64bit bus of a conventional Pentium might sound like a big cut. However, this is adequately compensated

1. The PODP83 has twice as big all important L1 cache which is one of the best assets a CPU can have improving hit rate and reducing pipeline stalls and bus traffic.
2. Compared to Pentium 60/66/75, the 83 MHz has noticeably faster internal frequency
3. The latency of refilling cache from L2 cache is just as fast on a late S3 motherboard (2-1-1-1) as on a typical early S4/S5 Pentium board (typically 4-2-2-2 or more). Yes. the Pentium fetches twice 128 vs 64 bytes in the same time, but there is no guarantee that the extra data is actually needed considering the same software.

Swiego wrote on 2020-06-23, 19:25:

FWIW I have been experimenting with processor upgrades for a not-the-optimal-platform 486DX2/66 (Compaq Deskpro/I) and comparing it to a stock Pentium 60 (Compaq Deskpro XE). In the below, I'm using the same ISA video card which is generally the fastest ISA DOS card I own.

Thanks for sharing your results. Looking at absolute numbers I think your numbers are greatly affected by using ISA VGA for mostly graphics oriented benchmarks and testing OEM systems. Yes, it affects the 486 as well as Pentium , but clearly it shifts bottlenecks out of the CPU and the PODP83 looks to be particularly penalised.

I know why. I own a Compaq XE5 60 too. So I know there is no PCI slot and most likely also no cache and no tuning potential. Likely the same or even worse for the 486. OEM systems are tricky.

If you given a better foundation to the PODP83, like a modern 1994/95 PCI board, it should have no problem beating a well tuned P60 or even P66 in (almost) any benchmark. See - this.

On the other hand you could argue that OEM systems was more realistic back in 1995 as people were upgrading systems like that...