It's a significant difference in overall performance, and the P75 is already low end for the Pentium era. It might be good for VGA DOS games.

I guess someone was describing hard disk I/O speed where Windows loading times came up. Nothing regarding CPU and memory performance is limited that strictly to loading times.

Depends on the SB16 variant. The AWE64 will always lack real OPL FM synth and the daughterboard header.

To be honest I'm probably not going to use it much once I have my DX2-66, I just want it right now because I /don't/, and I can use it for some multiplayer stuff.
The system I will keep for that era is a P2 233. Is a DX2-66 faster than a P1-75 with no cache? (66 has 128k)

I've got an IBM with a pentium 100 that have no cache (there are sockets to add it though, but I tried and I did not managed to make it work, the system acts oddly, crashes or doesn't POST and make beeps to indicate a RAM issue. The chips I'm using are 15ns chips) and I can still play early 3D games at good speed. I can run duke 3D flawlessly. I don't think that a DX2 outperforms that (duke is really laggy on a DX2), even with 25 less MHz on the pentium side

No L2 cache costs about 10-15% performance on a Pentium system. To be honest, I don't notice much of a difference.

No way a DX2-66 is faster than a P75, if both are configured correctly. The P75 should be about twice as fast.

On a Socket 7 Intel 430TX motherboard and Pentium MMX 233 when I only disable external/motherboard cache (L2) which is 512K on my system, there is an extremely tiny performance hit, insignificant.

1Doom benchmark:
2L2 On: 795.
3L2 Off: 844.
4
5Quake benchmark:
6L2 On: 53.3 fps.
7L2 Off: 48.9 fps.

Here is a nice explanation of how it works: http://www.karbosguide.com/hardware/module3b2.htm
When I only disable L1 (CPU cache) the perform hit is massive, turning the PMMX233 to a 486DX2 (+-), leaving the CPU access only to the slower L2 cache which is still faster than pure RAM.
Turning both L1 and L2 and the system is down to 386 speeds, where there is no cache between the CPU and RAM.

Apparently the larger and slower L2 motherboard cache on a Pentium machine (didn't exist before) is rarely accessed by DOS games probably because L1 is large enough to complete the commands without skipping or missing any data, therefor L2 which used as "backup" to L1 is rarely needed.
L2 becomes the next fastest reachable cache by the CPU when L1 is disabled.

The MMX Pentium's L1 is twice the capacity and much improved, and the bus is a bit faster. L2's effect is a little bit more pronounced with the old ones.

I wouldn't say so.
L2 is just much less significant in DOS gaming, so L1 alone whether 16K (non-MMX) or 32K (MMX) does the job very well without "skipping a beat" and needing L2 to "fix" that.

Pentium non-MMX 100Mhz:

1Doom benchmark:
2L2 On: 1230.
3L2 Off: 1297.
4
5Quake benchmark:
6L2 On: 29.8 fps.
7L2 Off: 27.5 fps.

Thought this might be relevant here:

The main differences in L1 cache, L2 cache and L3 cache are size and speed. L1 cache is the smallest and fastest, while L3 cach […]
Show full quote

The main differences in L1 cache, L2 cache and L3 cache are size and speed.
L1 cache is the smallest and fastest, while L3 cache is the biggest and slowest. L2 cache sits between them.

All three are types of caches used by a computer's CPU. They all exist to speed up the performance of the CPU by anticipating what data it needs.
When that data is stored in an L1 cache, L2 cache or L3 cache, the CPU can access and process it much quicker than it can if it pulls it from standard RAM.

The CPU looks for the data in order, which means it goes to the L1 cache first. This is where the data that the CPU is highly likely to need is stored.
It is very quick, and it is located very close to the CPU. As a result it is also small, so the amount of data that it can hold is limited.

If the CPU does not find the information that it needs in L1 cache, it checks the L2 cache. This is larger than L1 cache so it is slower, but its increased size means it is able to hold more data.
Finally, if the CPU cannot find what it is looking for in L2 cache, it moves to L3 cache, which is even larger .
Often these caches feed up the chain, so if a piece of data is needed regularly it will move out of L3 cache and up to L2 cache, or out of L2 cache and up to L1 cache.

My Pentium time machine is a Packard Bell with no L2. It currently has a Pentium Overdrive 200MMX in it, but I have benched it with a P75. It's on row 52 of this chart:
https://docs.google.com/spreadsheets/d/1uKhCI … #gid=1720967797

performance hit depends on the chipset.
with ali5 chipsets, performance hit is very small, usually within 5%. this is because ali5 has very fast sdram performance alone, which is only very slightly slower than onboard cache.
with mvp3 chipsets, performance hit is huge, sometimes over 30% slower. on the opposite of the ali5, mvp3 has slow sdram performance, which means it must be accessing onboard cache for all time.
i don't have intel chipset socket7 boards but i think they behave similarly to the ali5.

Hmm, this is interesting.
I was more inclined to believe that L1 is sufficient enough to not need L2 cache help, but I did not consider the other side of the coin where the RAM performance is fast enough to minimize the benefit of L2, good to know.
That means when disabling L1 and L2 the resulting performance very much depends on the RAM speed, much like it was on a 386.

Thanks for clarifying things out noshutdown.

My Cyrix MediaGX 266 system does not have L2 cache and is it ever sluggish.

I see, it's the DRAM that makes the difference. The platforms did go from FPM to EDO to SDRAM.

I assume that with faster CPUs L2 is much more significant, but for a Pentium with SDRAM is right on the edge of L2 speeds.
This is only an assumption as I have no EDO ram to test.

EDIT:
Alright, I have 2x32 dual-sided and 2x16 single-sided EDOs on the way.

It partially comes down to the CPU pipeline design, as well. Compare the performance penalty of disabling cache on the P2 compared to the PMMX. There's a reason why those first gen Celerons were so reviled.

James-F wrote:

I assume that with faster CPUs L2 is much more significant, but for a Pentium with SDRAM is right on the edge of L2 speeds.
This is only an assumption as I have no EDO ram to test.

On my 430TX board, SDRAM did not offer any speed benefit to EDO. In fact, I was able to set the RAM timings faster using EDO compared to SDRAM when maxing out the board with 256 MB of RAM and an FSB of 83 MHz. For 256 MB of SDRAM (PC100) to be stable, I had to use an FSB of 66 MHz. Even with the same FSB, the SDRAM had no speed benefit to EDO, that is, the benchmarks did not pick it up.

Deksor wrote:

I've got an IBM with a Pentium 100 that have no cache (there are sockets to add it though, but I tried and I did not managed to make it work, the system acts oddly, crashes or doesn't POST and make beeps to indicate a RAM issue. The chips I'm using are 15ns chips) and I can still play early 3D games at a good speed. I can run duke 3D flawlessly. I don't think that a DX2 outperforms that (duke is really laggy on a DX2), even with 25 less MHz on the Pentium side

I also recently got an IBM Personal Computer 340 (6560-77t) Pentium 1 (133Mhz) with no L2 cache but with sockets for dip chips and places for soldering SMD chips. I added 8 15ns SRAM chip + 1 tag SRAM chip of the same type but the system still doesn't detect the L2 cache.
The System works fine with or without the chips under Win98SE and WinNT 4.0 but I would really like to have my 256KB of L2 cache.
Any advice?

Thank you and best regards,
Null

I don't have that computer anymore. Unfortunately I was never able to get L2 cache working in that computer 🙁

Maybe the "easiest" way to get L2 cache working is actually to solder SMD cache chips on the spots where they're supposed to be x)

feipoa wrote:

James-F wrote:

I assume that with faster CPUs L2 is much more significant, but for a Pentium with SDRAM is right on the edge of L2 speeds.
This is only an assumption as I have no EDO ram to test.

On my 430TX board, SDRAM did not offer any speed benefit to EDO. In fact, I was able to set the RAM timings faster using EDO compared to SDRAM when maxing out the board with 256 MB of RAM and an FSB of 83 MHz. For 256 MB of SDRAM (PC100) to be stable, I had to use an FSB of 66 MHz. Even with the same FSB, the SDRAM had no speed benefit to EDO, that is, the benchmarks did not pick it up.

Sounds like a dodgy board. When I was benchmarking RAM performance of So7 chipsets with Ramspeed INTmem and FLOATmem, I did two i430TX boards (A-Trend ATC-5040 and Epox P55-TX2) and I was able to get between 10 and 15% higher performance with SDRAM than with EDO - and that was with L2 cache on (using a P54C though, so smaller L1 cache). In raw memory performance terms, the hit of disabling L2 cache was around 33% - but that won't necessarily translate to a similar hit in a game benchmark as the game will probably also be bottlenecking on CPU and GPU.

TS's board has an i430FX. In caching terms that's an interesting one as it straddles the divide between asynch and PLB cache. The latter is much faster, so the hit of disabling it (or not having it...) is bigger. On a Biostar MB-8500TEC with asynch, the hit of disabling L2 was about 6%, but on an Asus P55TP4XE with PLB it was about 12%, so the difference between async and PLB is bigger than the difference between no cache and asynch, at least with the i430FX, which had a very fast memory controller. On the Via Apollo Master (I compared FIC PA-2000 with async and PA-2002 with PLB) the memory controller was much slower, so the impact of no cache was bigger (and the difference between async and PLB relatively smaller).

TLDR: yep, it's slower with no cache, but the hit is very limited on the i430FX chipset, at least if you tune up the RAM nicely.

In cases whereby one motherboard shows a marked improvement when using SDRAM over EDO, it would be very interesting to compare such a board with another of an identical chipset which did not show improvement when using SDRAM. It is possible that the board which you suspect is dodgy may just be able to run its EDO as fast as the board which shows a marked improvement when using SDRAM. It could also be a dodgy board. I don't know, but there isn't enough testing to make such determinations.

How did I come up with this theory? I noticed that on some 486 boards which properly support EDO RAM and do show improvement when using EDO over FPM, those boards did not outperform other 486 boards with FPM RAM.