It's funny as it specifically mentions Cx486DRx2/Cx486SRx2 which are clock doubling cpus. Not sure how would you run them at the same speed as 386 as 2X multi is always on - not like with Ti486SXL2 which has to be enabled.

It would be interesting to know the differences between Cx486DRx2/Cx486SRx2 and ordinary Cyrix 486DLC. Is it just the clock doubling or anything more.

Ya, really makes you wonder if the "up to two times faster" is simply due to the 2x multiplier.

For the comparison, the idea would be to use an ordinary 486DLC.

I think the main difference between the DRx2 and the DLC is the 2x multiplier and the built-in cache coherency circuit. But perhaps there is more to it as the above quote seems to imply.

There might... but it will be hard to test as DRx2's multi is always on.

Phil performed some tests on the Cyrix CPU's - refer http://www.philscomputerlab.com/cyrix-486dlc.html
Based on my own benchmark tests, I came up with basically the same results (this was on motherboard that directly supported the Cyrix 486DLC and that allowed me to either enable or disable the L1 cache in the BIOS).

Based on these benchmark results, a Cyrix 486DLC-40 still runs slightly faster (with L1 cache disabled) than an AMD 386DX-40.
A Cyrix 486DLC-33 runs more or less the same speed (with L1 cache disabled) than an AMD 386DX-40.

jesolo wrote:

Phil performed some tests on the Cyrix CPU's - refer http://www.philscomputerlab.com/cyrix-486dlc.html Based on my own benchmark […]
Show full quote

Phil performed some tests on the Cyrix CPU's - refer http://www.philscomputerlab.com/cyrix-486dlc.html
Based on my own benchmark tests, I came up with basically the same results (this was on motherboard that directly supported the Cyrix 486DLC and that allowed me to either enable or disable the L1 cache in the BIOS).

Based on these benchmark results, a Cyrix 486DLC-40 still runs slightly faster (with L1 cache disabled) than an AMD 386DX-40.
A Cyrix 486DLC-33 runs more or less the same speed (with L1 cache disabled) than an AMD 386DX-40.

I'm more interested in verifying this statement, "The hardware integer multiplier calculates up to eight times faster than the software multiply functions of 386 and non-Cyrix 486 processors operating at the same frequency."

It would be possible to run some benchmarks that don't use much memory bandwidth and set 20MHz FSB for DRx2 and compare it to 40MHz DLC and 386DX. Run strictly some cpu only intensive benchmarks - maybe some dhrystones or MIPS calculations. I usually use PMIPS.

https://archive.org/details/msdos_festival_PMIPS

An eight times speedup sounds rather surreal - "up to" most likely refers to some rare special cases. I'm all ears for the benchmark results.

I'm actually getting my Cx486DRx2-66 in about a week. I'll do some comparisons between different 386-pin cpus.

Wish I could get one of the Ti486SXL2-66 in hands 😊

Oh nice, where did you score a DRx2-66 from? I bought mine from a collector many years back, a collector who was down-sizing.

The 486SXL2 isn't all that much faster than the DRx2. I uploaded some results somewhere on here. I think it was around 5%-7%.

I'm thinking something like Passmark might show the integer ability of these CPUs.

It's also from a collector. Just got lucky 😀

I have Ti486SXL2-50 but I'm not sure I want to run it at 66MHz... it gets really hot at 50MHz. And cooler isn't easy to mount as it has somekind of interposer installed. Some say it converts voltage from 5 to 3.3V - but I think it's too hot for 3V operation.

I was thinking of doing benchmarks only in pure DOS to minimize main memory usage.

I've read that DRx2-66 is not much faster then SXL2-50. Because it has only 1kB of cache.

Have you checked that PMIPS benchmark I linked before. It test different things... and shows nice graph. It also tests Integer operations separately.

I did not see any way to download the PMIPS benchmark that you linked. Passmark has specific tests for add/sub/mul/div.

Do you have a photo (top and bottom) of your SXL2-50 on interposer? The ceramic SXL2-50 I own can only work up to 55 MHz reliably. I wouldn't even want to run these at 33 MHz without a heatsink.

Download link is on the right at "Download options" and select ZIP 😉

Direct link here:
https://archive.org/download/msdos_festival_PMIPS/PMIPS.ZIP

I only have this photo. It's the one on the left side (yes, I have two but only one with this add-on):

Ahhh, I've seen those on some 386 upgrade website. I forget the website now. I think that circuit is added for increased compatibility. Perhaps it is the NAND-based inverter flush circuit? Do you have the part numbers on those ICs? The CPU looks like the standard 5 V variety.

I found this photo:
http://www.cpu-info.com/html/info.php?ID=639& … E=tblcollection

Excellent. Looks like the chip on the left is a 7-stage ripple counter, and the one on the right is a quad 2-input NAND gate. I know the NAND gate is used to make the flush circuit. I assume the counter some how benefits the flush circuit. I've put together this flush circuit and tested it on my motherboards, but found that none of my half-dozen 386 boards needed the flush circuit.

That chip was almost certainly out of the make it 486 upgrade kit. I used to have one. Mine also came with a heatsink. I got my second hand so not sure if it was part of the package, but I suspect it was.

Did some research to compare the cycle count from official documentation.
For Intel 386 and 486 (just a subset of the different modes, shouldn't matter to much):
http://zsmith.co/intel_i.html#imul
reg16: 9-22 (386) 13-26 (486)
reg32: 9-38 (386) 12-42 (486)

Cx486DLC (official datasheet, hopefully similar enough to the mentioned CPU):
ftp://bitsavers.informatik.uni-stuttgart.de/c … Sheet_May92.pdf
Page 141(6-16)
I'm not sure how to read the values precisely, but my guess is the document says
reg16: 3-7
reg32: 7-13
(worst case already being a cache miss, reg/reg seems to be 3 or 7 respectively).
If those value are correct the CPU is significantly faster than an Intel one, up to two times faster is not an exaggeration.

I also have difficulty comparing those two sources. It would be nice if the results came from the same source.

Hard to say how the values compare, but if I take ADD as another example this looks good:
ADD reg, reg: 2 (386) 1 (486) 1 (Cx486DLC)
ADD mem, reg: 7 (386) 3 (486) 3 (Cx486DLC, 5 for cache miss)
I'd say the 486 values don't allow for cache misses.
The ranges for IMUL in 386 and 486 mean the execution time depends on the operand values (early-out algorithm, this is why the minimum values are the same for different operand sizes).
The different value for IMUL for the Cx486DLC are for different cases (reg/reg, reg/cache hit, reg/cache miss) which means the actual multiplication runs in constant time.
Regarding how much difference this would make in benchmarks or games, that will depend on what you check.
For benchmarks, you could have a look at what Trixter wrote about Norton SI and his analysis of different 3D engines.
I suppose 3D games meant to run on a 386 are optimized in a way that multiplication instructions are kept at a minimum, so this might not make such a big difference in actual games. (IIRC DOS Doom replaced a lot of multiplications with table lookups.) Btw, IDIV doesn't seem to run faster on Cx486DLC.