Dawn operating system

[T3sl4co1l]

SUBLEQ is an academically interesting architecture, but it has absolutely no practical application.

Just as a Turing machine, or anything else simple and Turing-complete (like other simple TM-like programming languages, like most any of the *Fuck family; alas that's not vulgarity, that's just their name, like BrainFuck), is academically interesting, but not applied practically.

The trend has been to increase the amount of processing done per cycle, without impairing the overall speed or pipelining of the CPU.  The computation-per-cycle accessible with modern semiconductors is vast, many orders of magnitude greater than that possible with an oversimplified architecture like SUBLEQ.  It might be possible to clock a SUBLEQ at perhaps 20GHz, but it would have to run well over 20THz to even begin to compete with any top commercial CPUs.  It's physically impossible to execute extremely rapid calculations, at least on semiconductors, and with present technology.

The reason, as always, such designs exist, is because there is a market for them; if you are upset that they are expensive to design, and closed source, you are welcome to choose one of the many free and open cores available today.  You'll need to synthesize them into an FPGA, since ASIC design is still financially inaccessible to the casual programmer, but this isn't a big impact on most activities you'd be playing with.

Tim

[Hideki]

Much like solar roadways, this project is guaranteed failure, but you are of course free to use the rest of your life believing that it will become successful.

[Geri]

i dont know much about the actual design of ic-s and logical gates, i cant really form opinion in that, i just want to illustrated the wide scalability of mcu-s in correspondence with subleq, but if you know a better and more accurate way to do it, i will nodify it on the website.

There have been attempts at asynchronous CPU designs. That solves the clock power problem.

interesting and unrelated fact: originally amd designed the FX series to run the fpu at 3 ghz and the alu at 5 ghz like in the old times, but they were unsuccesfull

[Geri]

T3sl4co1l, what is your oppinion about a strong superscalar subleq server cpu? how much pipelines are possible to add, how many instructions per clock is possible? i calculated that for my os 3-10 billion subleq instruction per sec on 6-8 core is enough for multimedia usage, or 3 billion instructions on 4 cores for fluid office usage. of course the os runs far below this too, but becomes more sluggish.

[ataradov]

i calculated

Can you show those calculations?

Because you are basically talking about modern CPU frequencies, but it is quite obvious that subleq architecture requires more instruction to achieve the same same result as any RISC or CISC. So it is basically guaranteed to be slower than modern CPU at the same frequency and the same outside conditions (memory, buses, etc).

[Geri]

ataradov, i multiplied the data from the performance of the emulator.
on 300 million instruction per sec * 4 cores (emulated on my pc) it typically does the things on 0.3-2 fps

[ataradov]

So 4 cores at 300 MHz give you 2 FPS graphics? Comparable ARM cores at the same frequency play video without any problems. So I don't see the advantage here.

[Geri]

no, they dont. they have a built in hardware video decoder to play videos.

also you mistaking emulators with real hardware. in qemu, when emulating arm or x86, those system again just moving on 0.3-2 fps, this is a general speed rate with modern operating systems when they are being used in emulators.

[wraper]

x86 is becoming very much irrelevant aniways, as arm kills it

Tell this to data centers where most of the x86 profit comes from.

[ataradov]

no, they dont. they have a built in hardware video decoder to play videos.

Atmel SAM A5D3 plays 720p video at 536 MHz with purely software encoder on a single core. And that's with Linux running as the OS.

[Geri]

Hideki, no problem, i alreday reached 4000 unique visitors and 300 download, so i am alreday better than most of the kiddos with their very own Black Hat Josiph 2.0 linux distributions

maybe i should email tanenmbaum professor to ask him what grade he would give on this beautiful monolitic kernel design melted together to an unseparated gui controller in 2017

[amspire]

Geri,

Congratulations on how far you have come. I did download and run Dawn, and it said I had disk error on Drive D: at 26844074. It was slower then Windows/286 running on a 80286 CPU with 2M RAM back in the mid '80s, so you are starting a long way back.

I do think that ultimately that the current path of software is not very sustainable. As a mind game, imagine a colony ship was being prepared to leave the solar system. You couldn't use most of the current technology - you would have to limit essential technology to things that are manageable and replaceable on the ship. Definitely no closed source software and no closed source chip architectures.

Trouble is even if you can run on a beautifully simple cpu, the software will grow to be as complex as anything today. If it is successful, thousands of programmers will refine and optimise their speciality area. The OS cannot succeed if the developers are told they have to live in the past.

If there is going to be a breakthrough, I think it will be in how to be able to efficiently navigate a complex system without having to fully understand the complexity. If you think about all the towns and cities in the world, you can find them on Google maps or Openstreet maps, and at  basic level find your way around the town. You need more information and you find a local translator or guide. If you need something done, you find a local expert.

I think ultimately, software needs something like that - some kind of overriding structure for every component of the system to make it fully manageable. In this case, the local translators, guides and experts will be somehow implemented into this structure. Do not ask me how, but just as an example, if something needs recompiling, the system knows exactly how to do this. You do not need to know how to find the compiler, what version, what compilation flags, what libraries, how to install the updates, etc.

Of course if you achieve all this, you now have a hacker's delight. A cpu with no hardware security, and code that is totally easy to navigate and change.

Richard

[Geri]

ataradov - depends on the codec and the drivers. back then a 350 mhz k6 also was able to play 640x480 movies, but again, depends on the format, the codec, and the drivers. generally, anything below 2 ghz is unable to play anything, even my 1,6 ghz atom stutters a bit when playing 480p if unaccelerated nowdays.

[ataradov]

Codec is H.264. You can believe whatever you want. But I don't see the point if you are basing your comparisons on things you remember "back then", and not actual performance of the hardware. There are many things contributing to that performance, and I bet that K6 was running Windows and a ton of antivirus software in the background.

[Geri]

amspire, thankyou very much for spending some times to try my system. the block read error is actually a result of the emulator malfunction on windows, and its not related to the dawn os itself.

only the linux version of the emulator is tested properly, windows is a last minute compile (i ported it within a hour in a rush, becouse i planned to release the system by tonight), since i dont have windows at all, and i was not able to test it. i will fix the problem once i get close with a windows machine, so windows users will be also able to test it with proper and normal disk access.

[amspire]

OK. I will try in on Linux a bit later on.

[cyberfish]

First of all, your website states 50GHz CPU's. Please refer one, because I don't think those exist.

online docs talking about 50 ghz ic-s, like in this doc http://www.ursi.org/proceedings/procGA02/papers/p2121.pdf

50 GHz there refers to things like single latches, and those chips are incredibly expensive. You are definitely not getting a 50 GHz chip for $1USD. You won't be able to make a subleq CPU at anywhere near 50 GHz, for any price.

The subleq instruction is roughly of the same complexity as a conditional branch on ARM, which is one of the most complicated instructions on ARM. That means, more realistically, you are looking at maybe 100MHz chip for $5, or a 1GHz chip for $20.

Now the real reason why subleq is terrible in practice is because all common operations (like adding numbers) that can be done on ARM in one instruction will take hundreds of instructions on subleq.

There's more to an instruction set than simplicity - it needs to be useful, and that means executing common operations quickly. You can have the simplest instruction set possible (like you do here), but it's practically useless if it needs a handful of instructions per "useful instruction", and hundreds or thousands of instructions for more complicated operations like multiplications and divisions. An ARM CPU can do a multiplication in a single cycle.

A modern Intel CPU can do 4 simple instructions per clock. That's equivalent to something like a hundred subleq instructions.

[Geri]

UPDATE:
2017 march 15. 

Hotfix - Fixed the disk access bug in win32/w64 version of emulators. Download links and files are replaced.

[Geri]

(amspire, this fixes your bug under win)

[Geri]

cyberfish - a modern x86 cpu can do as many instructions per clock as many pipelines does it have. if it have 2 pipeline, it can do 2, if it have 4, it can do maximum 4 (in most cases, in theory). the same applies to subleq and arm. modern arm cpus have 3 pipelines. having more pipelines than 1 called superscalar design, and you can put ugly amouts of pipelines into a subleq cpu too if you want, more than into a very complex x86. the question is of course, how much you can effectively add to keep the things easy and effective.

having more pipelines will incrase the alu size, and that decrases the clock. so there should be a sweet spot.

if you say 50 ghz is not possible, you probably have right, maybe i will erase these datas from the website, becouse i dont want to advertise with bullshit

[ataradov]

the same applies to subleq

No, it can't. Branches are killers of pipelines, and all you have is branches.

[Geri]

if C points to the next instruction, its not a real branch, the hardware can optimize it out.

[MK14]

From reading this thread and the OP's website...

So we are talking about a 20,000 Core, 50 GHz, cpu, which is as fast (or faster) than the top Intel cpu at the time of availability.
Which only needs ONE competent Engineer for say 6 months (or so), to design it for the latest chip technology.
Unlike Intel who would need 500 or thousands of engineers, over a ten year period, to design their chips.

Crazy Intel, with their latest, 22 (between 1 and 22+ cores) core desktop/server chips, used in hundreds (or more ?) of millions and millions of computers/servers all over the world, built up over the last 46 years. You don't need the X86's thousands of instructions, silly Intel.

Just one single instruction, 50 GHz, 20,000 cores, and just a few months or so, for the lone engineer with a reasonable amount of competency, to do it all. On the latest generation 7 or 10 nm chips.

So I've got some questions for the OP...

At 50 Ghz, how much power will just a single core use ? (Ignoring some minor/tiny little problems, such as the speed of light and the laws of Physics etc) ?

What is that figure multiplied by x20,000 cores ?

How many dedicated local power stations will each cpu need Will it be practical to cool the cpus at a realistic cost ?

How exactly do you envisage your operating system and the software that runs on it, to usefully/efficiently use the x20,000 cores to do stuff, for most typical/average users ?

Why would people want to use your OS, over existing OS's, such as Linux and Windows ?

[ataradov]

You don't need the X86's thousands of instructions, silly Intel.

Nah, Intel just could not convince people to not use TCP/IP and Flash

[hans]

cyberfish - a modern x86 cpu can do as many instructions per clock as many pipelines does it have. if it have 2 pipeline, it can do 2, if it have 4, it can do maximum 4 (in most cases, in theory). the same applies to subleq and arm. modern arm cpus have 3 pipelines. having more pipelines than 1 called superscalar design, and you can put ugly amouts of pipelines into a subleq cpu too if you want, more than into a very complex x86. the question is of course, how much you can effectively add to keep the things easy and effective.

having more pipelines will incrase the alu size, and that decrases the clock. so there should be a sweet spot.

if you say 50 ghz is not possible, you probably have right, maybe i will erase these datas from the website, becouse i dont want to advertise with bullshit

x86 is CISC. This is the reason it has multiple execution units and deep pipelines to accelerate execution, because the instructions are 'complex'. Microcode in the CPU's chops up instructions into smaller instructions potentially, and also will reorder instructions when it can to keep these units and pipelines filled. Caches and branch predictors are in place to keep the machine running for as much time as possible.

Having multiple ALU's and execution units is helpful, but there probably is a saturation point. Code will eventually branch and branch predictors get it wrong.. There is a balance between all these units.
For example, Intel probably has enough transistors left to give each CPU core double the amount of ALU's. After all; GPU's occupy most of the chip area now. But if the controller will never use those extra ALU's in the datapath because of these limitations, there you have hit a saturation point.

if C points to the next instruction, its not a real branch, the hardware can optimize it out.

C is just a way to express your meanings that is translated to assembler.
SUBLEQ is an architecture that completely functions on being able to branch. This makes it horrific to accelerate.
If you do want to accelerate it, then it probably ends up in the same complexity of x86. Can it be done? Perhaps. Will it be done? Most likely not, because we already have x86 and ARM which are much faster.