Performance comparison of CPU: SuperFast-286-25 and AMD-386SX-25-Overclocked

System 1: 286-25mhz

Clocker in 25mhz (100/4=25mhz)
Mobo = Noname with five chips VLSI
CPU = Harris 25mhz.
FPU = Intel 287XL
RAM = only DIP20 Siemens 60ns(super faster) + Aluminium cooler
Cache = No
VGA = CirrusLogic 5420, 512kb, Jump JP1=Closed (for low wait state)
Sound = ESS1868
Multi I/O = Chips "TANS" (china?) for fast speed ISA.
HDD = CompactFlash 1GB, with Ontrack

For more information see here: Kixs's 286 to the Max

286 BIOS default:

System 2: 386SX-25mhz

Clocker in 25mhz (50/2=25mhz)
Mobo = 303n1
CPU = AMD 386SX-40
FPU = IIT 3c87SX-33
RAM = 2x4mb=8mb, 60ns SIMM30
Cache = No
VGA = CirrusLogic 5420, 512kb, Jump JP1=Closed (for low wait state)
Sound = No
Multi I/O = Chip GoldStar Prime 2 (not "2C")
HDD = Quantum Fireball AT 6.4, with Ontrack

386 BIOS is Overclocked:

AutoBios: Memory performance - Fast
to overclocked: Memory performance - Fastest

WARNING: Left only 286-25. Right only 386SX-25 + overclocked.

3D Benchmark:

3Dbench 1.0:

286 = 10.4 fps. Best.
386 = 6.7 fps.

3Dbench 2.0:

286 = 10.4 fps. Best.
386 = Error calculating. Value < 7.0 = overflow.

Formula One Benchmark 1:

286 = 102% load. Best.
386 = 151% load.

Formula One Benchmark 2:

286 = 316% load.
386 = ___% load.

Wolf 3D. PC Speaker (NOT SoundBlaster) + Mouse + FullScreen

286 = 19.3 fps. Best.
386 = 12.2 fps.

3DSpace:

286 = 15%. Best.
386 = 12%.

Morph3D:

286 = 79.0 fps. Best.
386 = 61.5 fps.
.
.
.
.
CPU Benchmark:

Landmark 2.0:

Landmark 6.0:

Checkit 3.0

Norton SysInfo 6.0:

286 = 19.4 XTs. Best.
386 = 12.3 XTs.

PC-Tools SysInfo 9.0:

286 = 15.1 XTs. Best.
386 = 10.3 XTs.

NSSI 60

DoctorHard 3.7:

PC Doctor 1.7:

286 = 6.9 MIPS and Read 16bit = 11mb/s. Best
386 = 5.2 MIPS and Read 16bit = 6mb/s

PC_info 4.04:

286 = 18.1 XTs. Best.
386 = 12.4 XTs.

FPU Benchmark:

Fbench:

287XL =
IIT387 = Best

CABT:

287XL = 6.26 sec.
IIT387 = 3.13 sec. Best

FLOPS:

287XL = 0.04 MFLOPS.
IIT387 = 0.12 MFLOPS. Best

RAM Benchmark:

SST (sorry only 386+)

286: Memory________Failed
386: Memory________R=10 MB/s______W=18.5 MB/s_____Move=12 MB/s

SysTest:

286 = 12mb/s, 0WS.
386 = 5,2mb/s, 1WS

HDD <-> ISA Benchmark:

CheckIt 3.0:

286 = 2774 kb/s. Best.
386 = 1765 kb/s.

OTHER Benchmark:

TopBench:

286 = Score 69. Best.
386 = Score 48

They probably would get the same results since games likely didn't take advantage of an FPU, but have you compared them without the FPUs installed?

FPU presence does not affect the game benchmark 🙁
These games are designed without FPU

For FPU you can do some AutoCAD benchmarks... like using "hide" with stpauls.dwg that is included with AutoCAD 10.

I guess the difference in FPU power is mainly due to lower 287 clock. Internally Intel 287XL is using 387 core as are IIT and Cyrix fpus. Only the first gen Intel 287 and AMD 287 use real 287 core.

Can you do also checkit hdd benchmarks. This would show the ISA bus speed. What ISA divider do you use on a 386?

How do you benchmark F1GP and Wold3D?

How come they get same (or very similar rather) results in 3DBench and Wolf3D given that one is at 40Mhz and the other at 25Mhz? Also one has a 32bit bus the other has a 16bit bus? I am a bit confused.

Regards,
Vlad.

He clocked the 386DX at 25MHz

vladstamate wrote:

Also one has a 32bit bus the other has a 16bit bus? I am a bit confused.

As far as the video card goes, both have a 16-bit bus, since it's ISA. We didn't get 32-bit until 486 localbus.
So for games/graphics-oriented stuff, both systems have about the same bottleneck.

As for the 386 CPU being 32-bit, much like with the FPU, this only matters when the software actually takes advantage of it.
Problem there is that games with 32-bit support generally don't have a 16-bit version to compare against.
So all the tested software here is 16-bit, which doesn't show the full potential of the 386 CPU. What it does show is that in 16-bit code, a 386 isn't really faster than a 286, which is rather interesting.

Ekb, this is brilliant work!

Very easy to compare, well explained and a range of benchmarks 😀

I didn't even know that Wolf 3D has a benchmark built in, how cool is that!

Nice to see that the 286 is so quick, 25 MHz is a great speed for a lot of these older games. Wing Commander should run perfect on it.

Scali wrote:

As for the 386 CPU being 32-bit, much like with the FPU, this only matters when the software actually takes advantage of it.
Problem there is that games with 32-bit support generally don't have a 16-bit version to compare against.
So all the tested software here is 16-bit, which doesn't show the full potential of the 386 CPU. What it does show is that in 16-bit code, a 386 isn't really faster than a 286, which is rather interesting.

Wouldn't the wider bus still give the DX an edge with memory access, as that is not limited to the ISA bus? Given that the 386SX is often regarded as being slightly slower per clock when compared to the 286, could it be that its the bus that is helping the 386DX stay ahead of the 286 at the same clock, even if only slightly?

alexanrs wrote:

Wouldn't the wider bus still give the DX an edge with memory access, as that is not limited to the ISA bus?

No, because memory accesses are done via instructions that are either 8-bit, 16-bit or 32-bit.
As long as you don't specifically use 32-bit memory load/store instructions, it doesn't get faster.

alexanrs wrote:

Given that the 386SX is often regarded as being slightly slower per clock when compared to the 286, could it be that its the bus that is helping the 386DX stay ahead of the 286 at the same clock, even if only slightly?

The reason why the 386SX is slightly slower has to do with how a bus works.
An 8-bit bus is the simplest form, you just generate an address and you receive an 8-bit word (byte). So everything is byte-oriented.
A 16-bit bus has to be backward-compatible with 8-bit.
The 16-bit bus itself can only access 16-bit words that are aligned on 16-bit boundaries, so only even addresses. If you want to access a byte, it will fetch the whole word over the bus, and then load the byte you wanted. So loading a word is just as fast as loading a byte.
The problems start when you want to access an unaligned word. It needs to fetch the two nearest words from aligned addresses, and then extract a byte from each, and put them together into the unaligned word. This makes it slower.

A 32-bit is like the 16-bit bus, except you have 32-bit words, so there are more ways in which you can have unaligned data.
The problem with the 386SX is that the CPU is a 32-bit CPU, so it 'thinks' in 32-bit words, even though the bus only supports 16-bit words. This results in the 386SX sometimes fetching data in a less efficient way than a true 16-bit CPU would.

My guess is that the slight advantage the 386 has in these benchmarks may be because of the cache. It would be interesting to disable the cache and re-run the benchmarks. What is interesting is that the raw memory benchmarks of the 286 are slightly higher than the 386. Might be the 60 ns memory and lower waitstates.

The 386 may also be slightly faster at complex instructions such as mul and div, but games probably don't use enough of these to really make much of a difference in framerate (but synthetic benchmarks do show a boost there, it seems).

It shouldn't be that close. I think something is wrong with your 386. 386DX has a 32-bit memory bus. A 286 is 16-bit. A 386SX should give comparable performance, a DX should destroy it.

Scali wrote:

No, because memory accesses are done via instructions that are either 8-bit, 16-bit or 32-bit.
As long as you don't specifically use 32-bit memory load/store instructions, it doesn't get faster.

Shouldn't the cache "buffer" reads and writes so even 8-bit and 16-bit sequential reads and writes would get a boost due to overall higher throughput capabilities?

alexanrs wrote:

Shouldn't the cache "buffer" reads and writes so even 8-bit and 16-bit sequential reads and writes would get a boost due to overall higher throughput capabilities?

Yes, but that is what you're seeing, I think.
Cache back then wasn't all that spectacular though. At 25 MHz, the memory is not that much slower than the CPU. And the cache is not that fast either, because it is still external to the CPU, so it is relatively high latency.
The cache probably starts to matter more when you clock the 386 to 33 or 40 MHz.

Thank you all ! 😀
I put links BENCH286.RAR in the first post.
and put more pictures benchmarks.

Please test your 386SX ?

To Scali: You are absolutely right in your posts. I agree with you.

sliderider wrote:

It shouldn't be that close. I think something is wrong with your 386. 386DX has a 32-bit memory bus. A 286 is 16-bit. A 386SX should give comparable performance, a DX should destroy it.

No it shouldn't destroy it. In 16bit code the 386 is internally faster than a 286 but not much.

One can go through great pain to find a 286-25 system while there's a crapload of 386DX systems out there. On this case the 386DX is just using default BIOS settings. At 25Mhz he could tune that 386 for 0 WS memory, which would give better memory throughput and probably improve synthetic benchmarks. Running the ISA bus slightly faster wouldn't hurt either and would improve games slightly.

By the way, I think the big difference with the FPU shows the power of the 32-bit bus.
The FPU only supports datatypes of 32-bit, 64-bit and 80-bit.
So even in 16-bit code, the FPU will perform 32-bit loads and stores, and on the 387 this goes twice as fast.

There are two types of 287 for the 286, the original 287, and the 287XL/80C287, which is based on the 387, and is more efficient. So the tested FPU is already the faster option for 286.

Any chance you could use an ET4000 for benchmarking 3Dbench again? I reckon you "might" score a little faster.

Unfortunately, I have no ET4000.
But there is available: Trident8900 (1mb), Trident9000 (512kb), Realtek (512kb), NCR (2mb)
they were slower because of the high WaitState VGA<->ISA.
Only CL5420 - low WS

====

Realtek RTG-3105E, 512kb

286 = 8.4 fps. Best
386 = 8.1 fps.

CompTest 2.6:

286 = 13 WS VideoRAM. Best.
386 = 18 WS VideoRAM

atperf23:

286 = 13.9 WS VideoRAM. Best.
386 = 21.7 WS VideoRAM

Trident 8900C, 1mb

Trident 9000i-1, 512kb

Trident 9000i-1 on 386 is failed. Crashed TextScreen in VolcovCommander and system halted 🙁

NCR 77c22e, 2mb

CirrusLogic CL-GD5420, 512kb

Result

Would be interesting also to see the comparision of 286 on same clock with 386SX. We had a discussion previously about is 286 faster than 386SX.