Dhrystone benchmark QL vs Amiga

[Peter]

Hi,

does anyone know how the Amiga Dhrystone benchmark in their "sysinfo" was compiled?

I had a little look at the Amiga's latest 68060 card, the Warp 1260, and despite a modern FPGA design with 2nd level cache and DDR3 RAM, it scores "only" 75010 Dhrystones/s, even if overclocked to 100 MHz. (http://eab.abime.net/showthread.php?t=98986&page=4)

The Q60 already reached 101010 Dhrystone/s with conservative 80 MHz. That was GCC compiled. I remember that C68 code was somewhat slower, but not explaining that an over 20 year old design with primitive PLDs and EDO RAM could compete these days. Just a C compiler issue?

All the best
Peter

[Derek_Stewart]

Hi Peter,

I always thought SYSINFO was written in assembler, which is confirmed by the Wikipedia entry:

Sysinfo is a shareware program written completely in Assembler for the Motorola 68k equipped Amiga computers to benchmark system performance. Sysinfo shows which version of system software is present in ROM, which hardware is present, and which operating mode the hardware uses.

The link is here:https://en.wikipedia.org/wiki/Sysinfo
Not sure that an Amiga Assembler application is easy to port the QL, as the Amiga uses a C style library system, but I guess not impossible.

[Peter]

Thanks Derek. It would be easier to run the QL Dhrystone benchmark under Amiga QDOS and see how it relates to the figure from sysinfo on the same machine.
Even if that machine is not the Warp 1260, it would provide an indication.

If someone owns an Amiga and wants to do the favour:
Amiga QDOS: http://aminet.net/package/misc/emu/QDOS4amiga1
QL Dhrystone benchmark: http://www.dilwyn.me.uk/benchmrk/dhryst21.zip

[XorA]

I wonder if it runs the test in ChipRam to be fair to all machines?

[Peter]

I don't know the Amiga, but assume "ChipRAM" means RAM on the original motherboard.
In case of the Warp 1260, that is much slower than the DDR3 RAM of the accelerator.

For a fair comparison to the Q60, the best is to use the "normal" main RAM of the accellerator, not something that goes through the bottleneck of the original Amiga bus.

(BTW it seems unlikely that the "ChipRAM" is used for the benchmark, because the Vampire figure is roughly twice as high as the Warp 1260. If the "ChipRAM" was the bottleneck, raw CPU speed would probably not have such an impact.)

[tofro]

Peter,

Chip RAM to the Amiga is basically what Contended Memory is to the QL: Memory that has to be shared between the CPU and the video (in case of the Amiga: also generally, DMA) circuitry. Many Amigas had mixed Chip/Fast RAM motherboards.

Because the Vampire emulates basically all of the Amiga (including the AGA chipset) it does not (or only very partially) need to use the on-board chip ram of its host for I/O, thus can be much faster. A CPU-only replacement like the Warp 1260 obviously has to use the "original" chip RAM on the motherboard to communicate to the chipset.

[Peter]

For a rough impression, it is not important which Amiga and which RAM.
Just the figure of the QL Dhrystone under Amiga QDOS versus the figure of sysinfo on the same machine and same kind of RAM.

My hope was, that an Amiga owner with a little QL background could run both and tell the figures.
My Amiga knowledge is zero, I have no machine, so it would take ages until I can do this myself.

[Peter]

A CPU-only replacement like the Warp 1260 obviously has to use the "original" chip RAM on the motherboard to communicate to the chipset.

Why should that matter for Dhrystone? It does not involve graphics, sound, etc. I can hardly imagine the Warp 1260 spends more than a few percent of it's CPU cycles for motherboard communications during the benchmark run.

[tofro]

A CPU-only replacement like the Warp 1260 obviously has to use the "original" chip RAM on the motherboard to communicate to the chipset.

Why should that matter for Dhrystone? It does not involve graphics, sound, etc. I can hardly imagine the Warp 1260 spends more than a few percent of it's CPU cycles for motherboard communications during the benchmark run.

Just as it matters to the QL (or the Q68): The video circuitry accesses the RAM concurrently with the CPU, thus makes memory accesses slower because the CPU has to wait for the memory accesses of the chipset to be finished.

[Peter]

A CPU-only replacement like the Warp 1260 obviously has to use the "original" chip RAM on the motherboard to communicate to the chipset.

Why should that matter for Dhrystone? It does not involve graphics, sound, etc. I can hardly imagine the Warp 1260 spends more than a few percent of it's CPU cycles for motherboard communications during the benchmark run.

Just as it matters to the QL (or the Q68): The video circuitry accesses the RAM concurrently with the CPU, thus makes memory accesses slower because the CPU has to wait for the memory accesses of the chipset to be finished.

No, because the DDR3 RAM of the Warp 1260 is separate (unlike QL/Q68). Note that slowed down video access by the CPU is not the issue here as graphics are not benchmarked. The DDR3 RAM bandwidth of the Warp is not affected by the video refresh happening the original Amiga mainboard.