Amiga vs PC

[Karlos]

the PC is still using the same design architecture as the 286.

Hmm, let's see. Current x86-64 architecture and supporting components:

16-bit ALU? Nope.
External FPU? Nope.
Segmented memory? Nope.
ISA slots? Nope.
Mono/CGA/EGA video? Nope.

In my previous job, none of the server machines bought in the 4 years I was there had a floppy drive, nor did most of the already installed ones. USB or CD only. Most had their OS installed via a bootimage on a USB or CD and used a network based install procedure.

[Karlos]

I wonder if Commodore 64 forums ever have to put up with this crap.

Maybe from spectrum users? :lol:

[Karlos]

I don't get these posts about, "on the pc, the architecture is old and needs to change". Change? How? Like what, remove super i/o and a bios rewrite? What difference does it make? End result will be the same, a computer that runs windows, *nix, OS X, aros, etc. End user won't even be aware of anything different.

Let's face it, it has changed a lot over the years. Here are just some of them:

Processor:
16-bit -> 32-bit -> 32-bit + SIMD (MMX, 3DNow, SSE etc) -> 64-bit multicore

Peripheral buses:
ISA -> PCI -> PCI + AGP -> PCIe -> PCIe 2.0 (not to mention the other PCI fork, PCI-X)

PCIe couldn't be more different to PCI, using high speed point to point serial "lanes" rather than parallel (which limits the achievable speeds significantly).

HD interfaces:
PC/XT -> IDE (PIO) -> UDMA IDE -> SATA -> SATA II (not to mention SCSI variants)

Other peripheral buses:
Serial/Parallel/PS2 -> USB -> USB2 -> USB3

Video:
Mono -> CGA -> EGA -> VGA -> SVGA -> HD

Audio
Beeper -> 8-bit + FM -> 16-bit + wavetable -> 24-bit HD + wavetable + DSP etc

I'm not really sure what there is that needs to change, other than people's concepts of what a PC actually is. It's not a "PeeCee", wintel box or any of the other decade old names people like to use. It's a modular computer made from standardised parts that is capable of running a myriad of different operating systems.

[Karlos]

So if computing power doubles every 18 months, and the A4000 was circa 1992, do you feel that your 2010 PC gives you the power of 2^12 (ie 4096) A4000's working at the same time?

Well, I built this PC in 2008. Core2 Quad is old news now, all the real hardcore PC enthusiasts are on i7 or better.

That said, depending on the benchmark, it can be a lot more than 4096x faster.

My 040, which is the same speed as those found in the A4000/40 (only without the crippled memory interface) manages about 5-7 MFLOPS in most amiga CPU benchmarks. In real code that I've written (a hand-optimised all-pairs solver for spacing an arbitrary number (read hundreds to thousands) of vertices in a spherical shell such that they are all as far apart as possible), that processor managed the equivalent of about 2MFLOPS, which is a very respectable outcome given the fact that a lot of memory read/write is also required. The unusually stressful nature of the work meant that unless the process was run at a low priority, the machine would freeze until an iteration was complete.

A vanilla, non-threaded version for linux on my PC ran around 400MFLOPS. Of course, that's not really making best use of it, since there are 4 cores. A threaded version (using pthread), compiled for SSE3 reached about 1.2GFLOPS. I suspect it could go higher, with hand-optimised SSE3 vector operations, but I don't really know enough SSE3 assembler to write it.

Again, though, that's not the best utilisation of the machine as a whole. Far and away the fastest version I've built reached ~450GFLOPS via the GPU when the number of vertices to solve for was 30,720 (number of streaming multiprocessors * maximum threads per multiprocessor). That's a speedup of 375x the threaded CPU version and 225,000x compared to the single-thread 040 version.

Suffice to say, the GPU is the coolest coprocessor I've ever toyed with

[Karlos]

Motorola 68040 ~ 27.5 MIPS at 25 MHz
Intel Core i7 Extreme Edition i980EE ~ 147,600 MIPS at 3.3 GHz

That's 5367 times the power of an A4000.

I've never seen a 68040 manage more than 1 instruction per cycle sustained. Of the 4 68040's I have at my disposal, none of them manage more than 0.8MIPS/MHz on the various benchmarking tools I've used. The 27.5MIPS figure looks far more like a 33MHz part to me. Which benchmark was this?

[Karlos]

benchmarks don't reveal true performance.  I wonder if you could render a Lightwave animation in 1/4096th of the time?

I take it you haven't seen some of the recent developments in lightwave 10? It has a realtime "Viewport Preview Render" mode. For some examples, see here.

How long do you think it would take an A4000 based toaster to render this?


Of course, there probably wouldn't be a lightwave 10 without the Amiga trailblazing the earliest versions

[Karlos]

Freescale claims 44 MIPS @ 40 MHz. Who am I to argue?

As with most claims by the manufacturer (intel included), this is probably a best case scenario for a cache hit (no latencies) case and without dependencies on still executing instructions (no stalls), register only arguments (no memory IO). Real code invariably doesn't behave like that.

[Karlos]

Similarly, DOSBOX runs DOS stuff fine but it won't run Windows 3.x stuff.  I.e., it doesn't have the required windows files.

When you say it "won't run", you are giving the false impression that it "can't run". It doesn't do it out of the box since including windows 3.1 with it would get the authors nailed up in Ballmer's trophy room quicker than you could say "developers" four times.

However, you can install windows 3.x on DosBOX, provided you have a copy of it that is still readable after all these years

Why anybody would want to, of course, is another matter.

[Karlos]

Friends don't let friends use windows

I've made the case for open source alternatives to friends that have nothing but pain with Windows day after day. Yet, despite not actually doing anything with their machines that really needs it, they seem unduly resistant to the idea. I guess they have a masochistic streak

Anyway, back on topic, we all know the Amiga was better back when it actually counted; in the days of genuine home computing. That era has gone. People may actually do more stuff with their computers nowadays, but precious little of it is what I'd call computing. Farting about on farcebook and the like seems to be what it has all come down to for many people today.

[Karlos]

I think they should update it to WinBOX so it runs old Windows stuff w/o requiring any installation or mounting crap.  I never even used the mount command before in DOS and now you are required to use it in addition to installing Windows 3.X on top of it.

:roflmao:

Perhaps they should upgrade hard disks so that you never have to install crap. They'd just come pre-filled as if by magic with everything you'd ever want.


Do you ever read any documentation for software?

Firstly, the mount command is used by DOSBox so that you can mount a host directory as a drive from within the emulation.

Secondly, do you have any idea what would happen to any person, organisation or company that went around giving out Windows 3.x for free with their application?

Tell you what, why don't you try that? Let us know how you get on, will you?

[Karlos]

I'll tell you whose REALLY stupid-- the person who just goes and gets a 64-bit OS and makes most of his previous software useless.  You are using ancient technology at 32-bits or 16-bits; why not try the truly over-bloated 64-bit OS that's incompatible  with your software.

No, the person who is really stupid is the one that thinks a 64-bit OS is incapable of running 32-bit applications and offers no benefit over a 32-bit OS.

You obviously know very little about how x86_64 is implemented. My system runs both 64-bit linux and 64-bit windows, both of which have ran every 32-bit application I've tested without complaint (although the only 32-bit applications I  run under linux just now are actually windows ones in WINE). The machine has 4GB of RAM and 896MB of video RAM, which just isn't possible in a 32-bit OS 4GB address space (unless the OS supports PAE). Plenty of the applications (read games) I run in Windows are 32-bit, though drivers and codecs are 64-bit.

Generally, the benefits are that 64-bit optimised code runs faster on the CPU than legacy x86 code does (there are a few rare exceptions, even in some of my own code), since 64-bit code can make use of 16 64-bit general purpose registers for integer code and at least SSE2 for floating point/vector ops.
Furthermore, 32-bit applications in the 64-bit environment can allocate more physical RAM than they could in a 32-bit one, since on 32-bit, only around 2GB was addressable in total (1GB of address space reserved for OS/hardware space, another 1GB used to map in the video memory. Again, PAE can mitigate this slightly). Now in a 64-bit OS, the 1GB address space used for hardware doesn't get in the way and if the process doesn't need direct access to the video memory, it doesn't have to be mapped into it's address space either. You might think that no 32-bit application should ever need 3GB of RAM, but then you probably haven't played Fallout 3 (after patching for large address awareness on 64-bit) with half a dozen resource hungry add-ons and HD texture packs. It certainly helped in this instance.

In short, if you have 64-bit hardware, which let's face it, every new desktop/server PC in the last 5 years (at least) has, using a 32-bit OS is pretty pointless. Even without more than 4GB total memory installed, 64-bit optimised code is usually a better fit for the hardware.

[Karlos]

However I believe that for my main computer use (FEA) a 64-bit system (rather than 32-bit) would actually slow down my solve times, due to the iterative addressing of a matrix of several hundred thousand/millions of simultaneous equations (overhead of translating 64-bit addresses vs 32-bit).

Have you investigated GPGPU for this class of problem? Finite Element Analysis is  one of the areas there are several CUDA precedents for.

[Karlos]

Given that the largest models I've been working with (around 600,000 elements) have tended to be overnight solves (just linear static analysis with a sparse solver) it would be the natural way to go, but as yet I haven't: lack of suitable hardware myself, and lack of forward thinking employers willing to spend a few £ to save a lot of £ (in man-hours) than anything else.

For CUDA, you can test the theory even on an old GTX 8800, provided you don't mind single precision arithmetic only. If you need double precision, the GTX260, though it isn't fully IEEE754 compliant and the double precision performance is only about 1/8th the single.

Not sure what ATI's stream stuff is like, but if you used OpenCL, you are pretty much free from having to worry too much which way to go.

I just need a cash injection of a few thousand though if you'd like me to look into it for you... just a drop in the ocean on those Amiga.org moderator's wages, right? ;-)

You got it. A mod's salary is priceless...

It always surprised me that the likes of Ansys did not supply dedicated hardware on a PCI(/e) card for higher-end customers (dedicated fastest ram available, dedicated co-processor, and optimised code for said processor.

Dedicated hardware can get old fast, I guess. An example would be GRAPE, an n-body solver for gravity simulation. GPU implementations on current hardware are capable of outperforming it comfortably.

[Karlos]

Ha ha ha, you got that that right :lol: It is still very sad that Commodore screwed up what could've still been the best machine.

In some ways, I wonder what might have happened if they'd succeeded. We're quick to assume it would have been a good thing but perhaps we might not have liked the direction it may have taken.

[Karlos]

This is so true.

I guess it's easier to say "we don't need it" than "you're right, it would be great, but we don't have it now, that's too bad". Seems like some people still live in 1989, thinking the Amiga is still cutting edge in a lot of areas... Problem is it's 2010, and it's lagging in pretty much every areas instead.

In the process of becoming more and more obsolescent, I actually find it quite interesting to see just what can still be done with it though.

I still find it much more fun writing code for AmigaOS and m68K than I do for modern kit. The only thing that's really piqued my interest on the latter is GPU stuff.

[Karlos]

Side note: I see classic Amigas as modern machines, new peecees are simply much faster, but that's all they are, faster.

I could add a few things to the list other than "faster". How about

cheaper
better supported
better modularity
more compatible (consumer devices, network protocols etc)
more operating system choice
more application software choice

These are all tangible benefits for most users. Let's face it, the Amiga as a hardware platform is not modern. It was modern in mid 80's, but every hardware advantage it had has been either surpassed by it's rivals or rendered redundant in some other way (planar pixels were great for scrolling but nobody uses them  now). Whatever could not be beaten by elegance has certainly been bested by orders of magnitude increases in performance applied to whatever less elegant solution was used instead.

Try running AROS, amithlon or even just UAE on a current "peecee" and tell me you don't find any advantages at all.