cache size difference has slight impact on 386 and 486 performance. the most significant importance of cache size is the range of cacheable ram: 128kb can only cache 32mb of ram, while 1mb cache can cache 256mb of ram. cacheable ram is halved if you set cache to write-back mode.
for cache speed, there are only two grades of cache timing to choose from: 2-1-1-1 and 3-2-2-2. 3-2-2-2 is much slower than 2-1-1-1.
20ns allows 2-1-1-1 up to 25fsb, anything higher than that would have to live with 3-2-2-2.
15ns allows 2-1-1-1 up to 33fsb, which is default for most 486 machines.
12ns cache is rare but shall allow 2-1-1-1 up to 40fsb.
I have one board with only 128KB of ram and the speed difference is negligible - the mode is set to write through though, to yield 32 instead of 16 MB cacheable area(8K cache per 1MB ram in WB mode - 1:128).
The front side bus speed you can run at tight timings does depend not only on the chips latency(e.g. 15ns) but also strongly on the batch it came from. I have 15ns UMC srams(UM61M256K-15 / 9534) which go higher than 12ns ones clock wise. Then there's a bunch of Alliance chips which do not go beyond 33mhz and are marked at 15ns too.
The difference regarding latency values is that they should be guaranteed to run at frequency blablabla(+a certain setting) no matter what/without error. From experience this is not true for all chips(see above). Also the values can be misleading as a 15ns IC should be able to theoretically run at up to ~66mhz(1000/15). This is not likely possible due to the demand on the sram itself which I guess somebody more knowledgeable will be able to explain.
Best thing is to find a seller with a bigger number of ICs and buy a few with different date codes and test them in a board as a tag ram paired with a set known to work flawlessly at higher speeds.
When I have bought srams I use the LuckyStar board at 60mhz fsb(+tightest timings) as a test system. Insert new sram as tag, boot and when it makes it into DOS I add it to the "good stash" which later runs at 50mhz through memtest and Winbech a few hours.
There must be a timings option unless it's an OEM board with a locked BIOS. Go to "Advanced Chipset Options" usually in the 3rd row, 1st colum on the BIOS screen. There should be settings regarding the cache wait-, ram wait-states and other options(maybe vga, dma, controller related).
More modern boards with a SiS496/497 or UMC8881/8886 chipset may support frequencies up to 66mhz front side bus(some crystals go even higher BUT this is more a theoretical than practical option).
Be aware that the most kinky setting in terms of speed is usually 40mhz fsb where the PCI bus and the (cache)ram when run full steam(ratio 1:1 and lowest wait states) usually deliver pretty good performance.
superchad wrote:I am curious about something, did some research but so far can't find concrete evidence […] Show full quote
I am curious about something, did some research but so far can't find concrete evidence
What difference does the cache size make with a 486
128KB vs 256KB
256KB vs 512KB
512KB vs 1MB
and what difference does cache speed make
20NS vs 15NS
15NS vs 12NS
do they make any performance difference for DOS or Windows gaming? even if very slight differences
I just upgraded the cache in my Gateway 486dx 33 computer with Micronics JX30G mobo.
It originally had only 128kb of cache so I had some cache laying around from 1995 and put it in.
The original cache was at -15 but the NEW cache I added was at -20
Does not seem to impact performance. The computer works fine. I just had to add the cache and reset some mobo jumpers and adjust the bios settings. Works fine with mixed speed cache. I now have 512kb of cache. It is suppose to make the computer perform better by about a factor of about 10%
I got a socket 3 motherboard that accepts up to 1MB of L2 cache.
* S3Trio VLB video card (Diamond Stealth 64 Dram)
* Fastest CF on a VLB controller
* 64MB Ram (4x 16MB simms)
* Hidden memory refresh, fast timings, lowest wait states
I did three bench marks: speedsys - CPU bound test with a small memory footprint Doom - Mostly CPU bound test with a medium memory footprint Win98SE boot time - Large memory footprint with a significant IO component. Seems like an important test, because I do it so many times. I tested by putting an echo statement in the autoexec that copied the time to a file, and then a batch file in the startup that copied the time to the same file. Boot times were pretty consistent, with almost all samples coming in within 3% of the average. I tested 4 boots for each config and threw out the rare outlier.
Below are the results for a 486DX2-66 CPU. In addition to showing the improvement from adding the cache, I also compared the results to the 486DX-33 times, to see how close it came to the theoretical 100% increase that a 66Mhz should have over a 33Mhz CPU in a perfect situation.
1Speedsys 2None 23.70 vs +93.6% 3256KB 24.76 +4.5% vs +99.7% 41024KB 24.76 +0.0% vs +99.7% 5 6Doom Realtics 7None 4242 vs +46.7% 8256KB 3007 +41.1% vs +70.2% 91024KB 2670 +12.6% vs +81.8% 10 11Win98 Boot Time in Seconds 120KB 74.3 vs +9.7% 13256KB 36.5 +103.6% vs +20.5% 141024KB 32.7 +11.6% vs +22.3%
I really wanted to do 64KB and a CPU faster than 66Mhz, but unfortunately the board is a strange one, and it won't start dos or windows with a CPU faster than 66Mhz or support a 64KB cache config.