You can check benchmarks here and decide for yourself:
https://docs.google.com/spreadsheets/d/1lvF9n … wpNU/edit#gid=0

Column V shows L2 cache (if any), but keep in mind that there are other often more significant factors that impact performance.

Main problem in that spreadsheet is that there isn't comparison of same system with cache and no cache/cache disabled.
At least I can't figure out from it if the cache makes any difference.

L2 should more or less double the performance, the effect should be immediately noticable.

I guess you mean "external cache", as there is no such thing as L2 cache on a 386.

But yes, impact is rather large. Without cache, speed is not much higher than a similar-clocked 286, but with cache, you can get close to a 486 with half the clock rate. A 386DX-40 with cache e.g. is almost as fast as a 486-25.

Tiido wrote on 2020-02-01, 08:16:

L2 should more or less double the performance, the effect should be immediately noticable.

Really? I didn't see any indication of that big of a difference in that spreadsheet and I thought many 386 motherboards didn't have any cache on them.

Edit: Perhaps they were limited by the video card and that is why I didn't see that much of a difference. In any case I guess a 486 33Mhz with L1 cache disabled is slower than 33Mhz 386 that doesn't have L1 cache. I wonder why that is.

Baoran wrote on 2020-02-01, 09:09:

Tiido wrote on 2020-02-01, 08:16:

L2 should more or less double the performance, the effect should be immediately noticable.

Really? I didn't see any indication of that big of a difference in that spreadsheet and I thought many 386 motherboards didn't have any cache on them.

Edit: Perhaps they were limited by the video card and that is why I didn't see that much of a difference. In any case I guess a 486 33Mhz with L1 cache disabled is slower than 33Mhz 386 that doesn't have L1 cache. I wonder why that is.

Yes, what the guys above said is true. 386s without cache are painfully slow, even a small amount makes a huge difference (it's not proportional to the cache size though). Also it's L1 cache, since it's the first cache available in the system, we're used to call motherboard cache L2 for some reason but 386s don't have internal cache (yeah apart from that exotic one) so the one on the motherboard is the first

Since there is such big difference, it brings up another question. If you look at system requirements for old games and they mention 386, do they mean 386 with or without motherboard cache?

I used to have a very compact 386 board that had a couple kB of cache integrated into the chipset. Made the board suprisingly fast. I kind of regret selling this board because it was such an interesting oddball. I wonder if that chipset would support additional cache? If only I could remember the name of that board...

Doornkaat wrote on 2020-02-01, 18:03:

I used to have a very compact 386 board that had a couple kB of cache integrated into the chipset. Made the board suprisingly fast. I kind of regret selling this board because it was such an interesting oddball. I wonder if that chipset would support additional cache? If only I could remember the name of that board...

I had several was a pcchips something and another

Just my 2 cents but I've run a number of performance test on 386 CPUs - SX, DX, with and without cache. What I found is most (all?) caching chipsets add memory latency, at about 2-3 extra cycles. So while cached access is faster, non-cached is really slow. On the other hand a cache-less 386SX mobo with 60ns RAM sticks is no slouch - not as fast as full-fat DX but the difference isn't as big as you'd think. For example a 40MHz SX system was just about matching a 33MHz DX system where the cache had to be set to 3 cycles of latency due to slow TAG chip.

Doornkaat wrote on 2020-02-01, 18:03:

I used to have a very compact 386 board that had a couple kB of cache integrated into the chipset. Made the board suprisingly fast. I kind of regret selling this board because it was such an interesting oddball. I wonder if that chipset would support additional cache? If only I could remember the name of that board...

That's the Macronix MX83C306 chipset I think. I've got a 386 DX-40 board with that chipset, apparently it's the MACRONIX CW-DXI3-M40 / INFORMTECH IT-AM33/40-DLC.
Initially I thought it had no cache but was pleasantly surprised to find it's got 8KB of cache in the chipset, which according to the manual is equivalent to 64KB of real cache. Speedsys reports the cache speed is about the same as a regular 386 though so I'm not so sure about that part.

I even found out what it's called thanks to the BIOS string and got the manual here

Thermalwrong wrote on 2020-02-01, 22:37:

Initially I thought it had no cache but was pleasantly surprised to find it's got 8KB of cache in the chipset, which according to the manual is equivalent to 64KB of real cache. Speedsys reports the cache speed is about the same as a regular 386 though so I'm not so sure about that part.

I’d like to see that logic. I can see how you say 8kb of on chip cache might be equal 64kb of chipset cache.

I have a 386sx-20 with 16kb of chipset cache, it is a little snappier than I expected.

BinaryDemon wrote on 2020-02-01, 23:43:

I have a 386sx-20 with 16kb of chipset cache, it is a little snappier than I expected.

Which chipset would that be?

Anonymous Coward wrote on 2020-02-02, 03:54:

BinaryDemon wrote on 2020-02-01, 23:43:

I have a 386sx-20 with 16kb of chipset cache, it is a little snappier than I expected.

Which chipset would that be?

I have no idea, its a very custom board found in a Tektronix RIC386 Tektronix RIC386 (Radisys)

It uses one of the TACT chipsets from Texas Instruments. However, the cache is not integrated into the chipset. It's inside that Toshiba branded chip with the -25ns marking.

Anonymous Coward wrote on 2020-02-02, 04:56:

It uses one of the TACT chipsets from Texas Instruments. However, the cache is not integrated into the chipset. It's inside that Toshiba branded chip with the -25ns marking.

Interesting , is that what the statement above about "8KB of cache in the chipset, which according to the manual is equivalent to 64KB of real cache". Whats the difference between integrated into chipset and on the Toshiba chip? Is my cache attached via the ISA bus or something?

How much cache do you need and what speed cache chips to get most out of a 33Mhz 386 cpu performance?

I think I remembered the name of the board with chipset integrated cache: It's an Octek Jaguar V and it has the MX83C06 chipset as Thermalwrong suggested. 😀
I can imagine chipset-integrated cache has performance benefits because the manufacturer can optimise the chipset for that exact amount of cache and/or use the integration for better/faster caching methods and lower latencies. How that exactly works: No idea.
The pcb date was 1992 though so it wasn't fast for its time and probably aimed at the budget market where a 386 was still passable and integrated cache was a welcome way to reduce cost. I think it also came with a Cx486DLC but I never tested it with that chip.

Baoran wrote on 2020-02-02, 06:35:

How much cache do you need and what speed cache chips to get most out of a 33Mhz 386 cpu performance?

128k is plenty on a 386. I don't think more than 256k are reasonable - if you need higher performance it's time for a faster CPU. I have read about boards having trouble managing 256k as well.
The chips have to be rated for the same speed as the cpu or better, meaning 30ns in a 33MHz 386 is what you need. Faster rated chips don't give you an advantage since they aren't driven faster but I guess you'll have less trouble finding 25ns or 20ns chips than 30ns ones.

My 386 motherboard has 8 cache chip sockets and half of them is used by 20ns chips making it 128KB. Is it worth getting another 4 chips to make it 256KB? Would I also have to switch the tag chip? I understand there is a risk that it won't work, but would there be any change in performance?