[FAQ]
[Frequently Asked Questions] [Resources] [Emulators] [Where Is...?] [File Formats] [Technical Information] [Pinouts] [Acknowledgements]

This page last updated on 6 January 2000

[DOS] [Windows] [UNIX/Linux] [OS/2] [Mac] [Archie] [Amiga] [Psion] [Windows CE] [ST/TT] [Java] [Playstation] [QL] [Other Platforms]
[Benchmarks]

In order to gauge the speed of various emulators, the following tests have been carried out:

1. 10 FOR I=16384 TO 23295
   20 POKE I,255
   30 NEXT I

   This is test of video access and the BASIC interpreter: note the remarkable influence of page swapping in 128K modes, especially if there is no EMS (eg Z80 and UKV).

   SimCoupe failed the test due to different video memory, and Sinbas gave 'B Integer out of range' because POKE is not implemented.

2. 10 FOR I=16384 TO 23295
   30 NEXT I

   Removing the writes to video memory from the first test will give the speed of BASIC interpretation. As expected, Sinbas is the fastest here.

   SpectrEm-Dr failed due to a bug (it does not finish the FOR loop), whilst ZX Spectrum Simulator does not respond to the keyboard in BASIC.

3. BEEP 10,0

   This is test of the intensive I/O operations required by BEEP. Nuclear ZX has an option to remap BEEP to PC timing. The sound is clean if the emulator is fast enough and nothing disturbs the tone. The emulators which produced a clean sound were ASp, Nuclear ZX, Roman ZX, SP, Spec32, Spectrum, Warajevo, WinZ80, ZX (Orro), ZX (Rindt & Brukner) and ZX32. ZX32 and ZX (Orro) have a significant delay before the start of the sound.

   Once again, ZX Spectrum Simulator does not respond to the keyboard in BASIC.

4. PAUSE 500

   This is test of interrupt system. If the result is close to 10 seconds, even if other results are not correct, then emulator is using hardware interrupts.

   ZX Spectrum Simulator agains fails because of a lack of keyboard input. Spectrum fails when in 'no screen' mode because it is impossible to determine when the test is over. SpectrEm-Dr quits too fast in overclock mode to be measured.

5. Start the first level of Manic Miner and time how long it takes for Willy to lose all his air.

   This is the only non-BASIC test. Note the influence of fast LDIR on Z80 and Sinclair. There is no difference between 128 and 128/48 modes here.

   Various emulators failed this test as they do not provide sufficient facilities for loading games, whilst Sinbas does not do Z80 emulation, and Spectrum in 'no screen' mode again provides no clue that the test is over.

The various modes used are:

• 48: Spectrum 48K emulation.
• 128: Spectrum 128K emulation, 128K BASIC.
• 128/48: Spectrum 128K emulation, 48K BASIC.
• +3: Spectrum +3 emulation.
• 2068: Timex 2068 emulation.
• Coupé: Sam Coupé emulation.

Many emulators have options that may impact on perfomance, of which some combinations have been tested.

Note: All tests are done with a hand watch, so the results could be shorter by up to 0.3s.

All DOS/Windows emulators were tested on a 33MHz U5 486SX with 16Mb RAM and a Trident TVGA 8900 video card running Windows 95 OSR 2 with Direct X 6.1 and WinG. All DOS and DOS/32 emulators were tested by rebooting into DOS mode without drivers (except if they required EMS) and then started from the command prompt. The only exception to this is Speccyal, which required a more powerful video card. This was tested on a Pentium 100, and the results scaled to the 486 speed. R80 and Spectre both required a coprocessor emulator.

Most of the Amiga emulators were tested on an A1200 with a 50MHz 68030 and 10Mb of RAM. The exceptions are Spectrum128 which was tested on a 40MHz 68030 with 10Mb of RAM, and then scaled (simply by a factor of 0.8), and ASp which was run on a 64MHz 68060 and has not been scaled.

xz80 was tested on a Sparc 2.

No attempt has been made to scale results between the three architectures tested.

The 486 is obsolete, but was chosen for the following reasons:

• It is hard to make benchmarks on too fast a machine. It is easier to tell the difference between 20 and 22 seconds than between 2.0 and 2.2!
• If all emulators can reach Spectrum speed, the testing is pointless. The goal of testing is to investigate the usefuleness of emulators, even on slow platforms.
• The whole emulator scene is a protest against too fast obsolence of hardware and software: any computer older than two years is obsolete. Compare this with your television which can last up to 15 years!
• Hercules graphic card modes can also be tested.

To see influence of the target machine on speed, Z80 v4.00 and Warajevo 2.51 were tested on a variety of machines. These emulators were chosen as they are two of the best emulators for clean DOS and executable on all machines. The results for other emulators can be extrapolated, and you can see if your computer can reach true Spectrum speed. Z80 is tested using the -xt -z8000 options (make it as fast as possible and reduce RAM requirement); Warajevo is tested with SPEC48 /%MAX option. The estimated speed reported by Z80 is given in brackets.

Results for an XT, a 386DX and a 486DX would be appreciated.

• CPU: The slowest machine that can theoretically execute Spectrum machine code (not 100% compatible) is a 286/16.
• Memory: the speed of memory is a bottleneck for the CPU. For example, it is interesting that if you switch off the CPU internal cache then Z80 becomes faster than Warajevo 1.5, although normally the opposite is true: the reason is the longer, although more efficient, processor emulation engine on Z80 that generates cache misses. The type of memory also has an influence: EMS memory can drastically accelerate 128K BASIC.
• Video: The fastest video mode is redefined text mode because it has only attributes. Standard EGA/VGA/CGA/Hercules modes are slower, simply because they require more calculations. For some emulators (eg X128 SAOM) PCI cards can make the emulator much faster than it was with an ISA card.
• Programming language: Emulators written in assembly are generally faster than those written in C, C++, Pascal or Java.
• Level of compatibility: The more compatible emulators are slower. Emulation of R register, T-state counting, rainbow effects, contented RAM and the like all require time.
• Code optimization: The most critical routine is the decoding of Z80 instructions. JPP has the fastest such routine, whilst emulators written in C have a very long such routine. Also, it is important to decide what to keep in registers and what in RAM.
• Interrupt emulation: Z80 interrupts can be generated using either hardware or software; The hardware interrupt method uses the emulating machine's timer, and makes some games useful even on 286/8 or XT/12. However, flickering and interrupt congestion can occur. The software method, which generates the interrupt by counting T-states, has no such problems but it is slower.

Thanks for this page should be sent to Samir Ribic, who tested all the DOS/Windows emulators, all the PCs and wrote most of the text. Credit is also due to Ian Collier, Rob Darke, Joe Mackay and Chris Young.

This FAQ is maintained by Philip Kendall; distribution is permitted only under the conditions specified in the copyright notice.
Primary site for this FAQ: http://www.ast.cam.ac.uk/~pak/cssfaq/index.html.