Depending on motherboard chipsets, you may have to install matched pairs of memory sticks as they may require interleaving. I noticed you mentioned just using a single memory stick?

Yeah.. use two sticks and turn interleaving on. That should help quite a bit.

Yes, I am using a single stick of 72 PIN FPM. I'll try a pair as suggested and post the difference, thanks.

Using a paired set of ram sticks made no change on the bandwidth results. I tried disabling the cache and did another speedsys and the bandwidth went up to 110 MB/s.

I have also seen the situation where memory benchmarks are higher when cache is disabled. Some chipsets suffer extra latency when there is a cache miss. But sometimes this only happens when the L2 is set to Write-Back, but not on Write-Through.

100MB/s is pretty high for socket 3.

What happens when you enable "DRAM Slow Refresh"? Contrary to how it sounds, it should actually increase the throughput and lower the latency. The more time between the forced refreshes, the less wait their will be between reads and writes. It just depends on if your RAM will handle the longer time between forced refreshes.

What about "CPU Burst Write"?

What about trying "Write Through" on the CPU Internal Cache and /or the L2 Cache Policy?

I don't actually see anything about interleaving on that BIOS screen.

Just to chip in, i have a Soyo SY-25P2 with AMD486@100MHz, 2x32MB 60ns RAM.

When the cache is set to Write-Back, i get memory throughput under 40MB/s. When cache is set to Write-Through, i get ~100MB/s.
If anything, Write-Back should be faster, but...

On some random dos benchmarks, the speed was mostly not affected. Some tests were maybe 5% faster, some 5% slower between those 2 settings. So i'm also kinda confused and will keep an eye on this topic. 😀

Update on the testing. Setting the CPU to WT had no effect. CPU burst mode on the mobo would hang every time even with different settings in the bios so that stays off.

Setting the cache on the mobo to WT......
Bandwidth now at - 110.06!
L1 - 161.51
L2 - 44.48
Throughput - 37.64

I put the CPU back in WB, that had no effect on the speed. Setting the Dram to slow refresh had a small increase.
Bandwith 112.08
L1 - 45.32
L2 - 38.20

I tried both a pair of sticks and one stick that had no change on the speed. Also before the FPS in 3Dbench went from 64.1 to 71.2 now with the cache on the mobo to WT.

I thought WB was supposed to be faster? would that be effect be chipset specific? One thing to think about, these 4 motherboards that I was testing are all SIS, I wonder if that is why?

^^ Have you tried changing L2 cache tag bits from 8bits to other options and see if that affects the bandwidth?

I did try changing the 8 to 7 bits that made no difference.

I just tried my 486-PIO3 that has poor bandwidth as well. 27.41 but there is no WT - WB options in the bios. I turned the cache off and it went up to 82.37 bandwidth.

I found this thread here having the same poor bandwidth. FIC 486-PIO3 mobo - why's it so slow?

Is there a way to have the bios modded to add WT - WB options?

3D Bench has 71.72 FPS with the cache on and 56.60 with the cache off.

FIC 486-PIO3 mobo

I just found this thread ^ when searching from the same OP, he has his MS-4144 set to WB and he is getting about the same performance as I just got with it in WT. But I don't understand why his board is getting full bandwidth in WB and I am not.

^^ It's quite conceivable that the BIOS writers mislabelled the WT/ WB options and these are actually swapped. If you can't find these BIOS options, try looking for a jumper on the board that is equivalent.

What happens if the board physically doesn't have the L2 cache chips installed? Does it act the same way as if it is disabled in the BIOS?

Now what I'm really questioning is the usefulness of this "memory bandwidth" which basically only seems right in the opening post and wrong by a long shot every other time (including my own benchmark results from many systems), or completely made up in the case of 486s.

An example could be ?my? (I'm still hesitating whether I should keep it or not) PCChips M919. Let's see the numbers from the .txt file it saved:

1                ┌──────────────┬──────────────┬──────────────┬──────────────┐
2                │     Read     │     Write    │     Move     │    Average   │
3┌───────────────┼──────────────┼──────────────┼──────────────┼──────────────┤
4│ Cache Level 1 │   151.69 MB/s│    75.38 MB/s│   195.95 MB/s│   141.00 MB/s│ 
5│ Cache Level 2 │    68.17 MB/s│    75.20 MB/s│    34.07 MB/s│    59.15 MB/s│ 
6│ Memory        │    43.33 MB/s│    75.43 MB/s│    16.56 MB/s│    45.11 MB/s│ 
7└───────────────┴──────────────┴──────────────┴──────────────┴──────────────┘

It looks like this on screen:

So the fastest way for the CPU to access the memory is sustained write, at 75.43 MB/s. Read is slower and if you try to combine them in the hope of squeezing out a bit more of the memory subsystem (like it apparently works for L1 cache) you get a way slower 16.56MB/s result. Where on Earth that 120.53MB/s figure comes then? Apparently there's no way for the CPU reach this bandwidth to/from memory. Is this a measurement at all?

(And slightly off topic, the reason I'm hesitant to keep this board is an x-file. I have a Shuttle HOT-433 with the same chipset, the same memory benchmarks and the same quirk (have to POST at 33MHz then switch to 40MHz if I don't want the hideous 20MHz PCI clock seriously impair the performance). It even works with the Shuttle BIOS. But the VGA speed and many even loosely related (like 3dbench or game) benchmarks are all slower (except for speedsys which shows a very healthy video speed). So should I just use it with a copied Shuttle BIOS? Well, no, because it's just as slow with the swapped BIOS. HOT-433 landmark VGA: 25MB/s. M919 landmark VGA: 17MB/s. HOT-433 3dbench: ~98fps. M919 3dbench: ~93. Same VGA, same BIOS, same settings. The hell is happening here? 😲 )

alvaro84, I was wondering if it matters from what you said, I did more testing and found that it does. This can be seen in cachechk to confirm and doom.

Another Asus mobo, using all the same peripherals with the L2 cache on WB in the bios

Memory bandwidth 37.55 MB/s
L1 - 155.53 MB/s
L2 - 38.00 MB/s
Memory Throughput - 30.97 MB/s
Memory speed in cachechk 16.8 MB/s
Doom FPS 38.0

L2 Cache set to WT in the bios

Memory bandwidth 112.16 MB/s
L1 - 162.54 MB/s
L2 - 45.01 MB/s
Memory Throughput 38.00 MB/s
Memory speed in cachechk 36.6 MB/s
Doom FPS 40.6

From what I am experiencing L2 cache on certain mobos being set to WT is faster then WB. I'm not sure what else could be tested to see the impact on performance.

I think the performance between WB and WT is also going to depend on the program being run.

See this thread for a bit more explanation of WB vs WT:
Identify 486 L1 WT or WB Cache?

It can be a missing „dirty tag“. Without dirty tag the cache in WB can be slower than in WT.