IBM's PPC being open sourced .. what does this mean for embedded devices ?

If someone takes on and starts making more power architecture processors the impact could be huge, this is by far the best architecture I have seen in my life - and I have explored most of them (and used power for decades now). Then it might make no impact at all, how many of those making decisions of such magnitude can see further than the next quarter report and can look into something other than what "everybody is doing".

The only maker of Power architecture processors is NXP (since they acquired Freescale). IBM had an embedded core some ages ago, 401 or something like that, which some other company - AMCC I think - took on but I never used any of these.

Thanks for posting, I had missed the news so this is new to me. Good news I suppose, gives some hope that this great architecture will survive after all.

Hopefully I can migrate DPS to 64 bit power sooner rather than later, I have been wanting to do it for quite some time now. Might also be some help.

Dimiter

====================================================== Dimiter Popoff, TGI

ce/)

Maybe I'm not remembering it right, but I thought the PPC fell out of favor because as processors got faster and more complex, the PPC architecture lo st it's advantages. I would seem I am not remembering this correctly.

So why did the PPC architecture stop being developed? Is this still a comp etitive architecture in general? I'm guessing it is still competitive in l ower end processors for sure. How does it stack up to MIPS? Isn't that on e also open source? Or was that open sourced only in a very limited way?

The architecture did not lose its advantages, it is just as silicon got faster messy designs like the x86 could run fast enough for the general population.

I don't know how far did IBM go developing it, may be they still do. I know only the Motorola/Freescale line, some of which I have used and am using. Their latest designs are 64 bits QORIQ (someone got away without being punished for that name) which look quite good really.

MIPS looks like a smaller scale power, I have not used it, just looked at it. No idea if it is open source. Microchip make some 32 bit MCU-s with MIPS, I have not heard people using them complain.

Dimiter

urce/)

ea

rs

avor because as processors got faster and more complex, the PPC architectur e lost it's advantages. I would seem I am not remembering this correctly.

There has to be some reason why they lost the war other than marketing. Ap ple was using the PPC and many others. So there was clearly a market. Per haps I didn't state things correctly. The advantages of the PPC must not h ave been significant enough in larger chips to make continued development o f ever more expensive designs worthwhile. I recall when IBM dropped out of the race. They must have seen the handwriting on the wall.

So how does Freescale keep paying for spins on 64 bit chips? I take it the y still can hold a match to the Intel processors, rather they are now compe ting with the ARMs of the world?

competitive architecture in general? I'm guessing it is still competitive in lower end processors for sure. How does it stack up to MIPS? Isn't tha t one also open source? Or was that open sourced only in a very limited wa y?

Yeah, seems the open source thing is real... They seem to be saying MIPS h as some real advantages over RISC-V but they may not last...

"While RISC-V is gaining grown[sic] the storied history of MIPS means that the MIPS stack is far more complete, and includes things like DSP and SIMD extensions that still don?t exist for the RISC-V platform."

I've read here that most MIPS designs are embedded cores in custom ASICs. So someone already loves them. Seems the current owners of the MIPS archit ecture think open sourcing the design will allow them to sell more IP solut ions.

No, they are not (as far as I understand the situation). The linked article seems a bit mixed up as to what they are giving away (they are mixed up about RISC-V too).

When talking about processor licensing, there are many levels. You can license the ISA - the instructions, register setup, instruction codings, etc. This is the minimum you need in order to make a cpu that will run binaries for that processor. You can license design files (Verlig, VHDL, etc.) for some or all of an implementation. You can license hard macros for some or all of an implementation. You can license the right to use names, trademarks, patents, etc. You can license the right to modify and expand on any of this.

As far as I know, IBM has offered commercial licenses on all of these bits - just as ARM, MIPS, etc. do. Some customers have taken the ISA and made their own implementations, others have taken complete hard macros and put them on ASICs, others go somewhere in between.

From this article, it seems IBM is giving free licensing for the ISA. That means you can make your own processor in an FPGA, ASIC, etc., that runs PowerPC code. It is not clear how open it is - do you need to obtain a zero-cost license? Can you modify the ISA as you want? What about naming - can you call it a PowerPC processor or PPC compatible, even if you add things to it? What about if you remove features? Will groups other than IBM, Apple and NXP be involved in the future of PowerPC and its ISA?

It does not appear that they are giving away a usable /implementation/ of a PowerPC core.

I'm guessing answers to all these points are to be found with a bit of research on the Linux Foundation or IBM's websites, but they need considered before it is clear what IBM have actually done here.

Mostly nothing, I think.

PPC cores are used in a few specific embedded markets - primarily the automotive industry, and other tough industrial uses. They used to be popular in networking systems (along with MIPS), but ARM is taking over that niche. I think you will continue to get PPC-based devices from those that made them before (primarily NXP, with some second-source devices from ST). I don't think anyone is going to design a new implementation for new devices unless they are already invested in the PPC world. PPC simply does not have any advantages over alternatives, and it is a big and heavy ISA compared to RISC-V.

What you might see is a few more second-source parts, perhaps also things like radiation-hard or high-temperature versions from folks like Atmel. And you might see it in use in systems that need a very fast processor and are willing to spend a lot of effort to get it, but want low production or licensing costs - maybe we'll see it again in games consoles, or chips designed specifically for supercomputers.

No, it won't kill RISC-V. It will (in my humble and only somewhat informed opinion) aid RISC-V, by making "open source ISA" more acceptable to conservative and sceptical companies who viewed "open source ISA" as meaning "made by hobby developers" or "made by communists". When IBM says "open source ISAs are good", they are more likely to take note of other open source ISAs.

I'd be surprised if such a thing exists. There are /many/ PPC variations.

There are others that have made PPC based devices in the past. But I think NXP is the only company making microcontrollers and network processors based on PPC in any serious way. ST has (last time I looked, which was a while ago) second-source copies of a number of NXP PPC microcontrollers, but I don't think they had any of their own.

The key change, I think, was the move from x86 cores that ran x86 code, to x86 processors that translated x86 instructions into an internal RISC code that is executed on a RISC core. RISC designs and load-store architectures with lots of registers are much better suited to fast execution, pipelining, OoO execution, and super-scaling than CISC ISAs. Once x86 cpus started using RISC cores internally, much of the advantage of PPC is lost in comparison.

And x86 has several advantages that PPC can't get, beyond the obvious compatibility with x86 binaries. One is the x86 strong memory model - it is a good deal simpler for programmers than weaker memory model of most RISC designs, and a lot of code is written with the assumption of a strong memory model. The PPC exposes the guts of its RISC core to the programmer - and thus cannot easily change it - with x86 cpus, their RISC cores are hidden and designers can change it as they want for each new chip. They can have 37-bit wide instructions, completely re-arrange the encoding, add hundreds of registers, etc., whereas PPC designers have to stick to the ISA.

Of course, PPC implementers could move in this direction too - turning PPC RISC instructions into an expanded internal RISC design for execution. I think the size of the ISA might make that harder than for the x86, and it won't solve the memory model challenge, but it could work. But they haven't had the same motivation - they haven't hit the same performance brick wall as the x86 world faced.

No, there doesn't. IBM have demonstrated in the past that they are capable of technical excellence combined with extraordinary screw-ups by sales, marketing, purchasing and licensing departments that don't understand the products, and don't talk to each other. The whole sad OS/2 story is an example.

That doesn't mean there /aren't/ other reasons, merely that non-technical reasons can be enough.

Certainly keeping in the race was getting more and more expensive. Intel could use their server chip money to pay for it, but IBM did not have high-margin PPC customers (there Power cpu line for big iron is separate) screaming for faster cores.

They don't. No PowerPC devices come close to top Intel/AMD processors. And for the devices NXP make, the core speed is not critical. NXP make two types of PPC devices - automotive / high-reliability devices (up to about 300 MHz core speed), and networking (up to 1.5 GHz core). For the automotive devices, reliability, robustness and peripherals are key, not the core or the core speed. For the networking devices, it is the networking hardware that is key - the core is for control, not for processing. AFAIK all NXP's PPC chips are 32-bit.

There is often a confusion between "Power" and "PowerPC". It doesn't help that NXP talks about their "Power" devices when they mean "PowerPC".

The "Power" architecture is IBM's big iron chip, and these are big, fast, expensive, power-hungry, and full of big iron features like redundancy, hot-plugging (for memory, cores, etc.). They are made by IBM, and they run the many of the top supercomputers as well as every bank in the world.

"PowerPC" was an initiative started by IBM, Apple and Motorola - taking the basic ISA from the Power world, removing all the "big" features, and making something that could be implemented in a single chip that was competitive in size, speed and cost with the x86's and 68k devices of the time, but with a scalable future.

If you haven't heard them complain, you haven't heard me :-) My complaint with Microchip's MIPS microcontrollers is that they used a poor choice of core (MIPS had more appropriate ones at the time), screwed up the chip (it was to be the first microcontroller with USB 2 on the device, but it only worked at USB 1.1 speeds), crippled the compiler (optimisation was disabled unless you paid a large license fee, meaning people got the impression that the core was slow), and called it "PIC32" so that people thought MIPS cores were as brain-dead as a PIC. In my mind, Microchip and the PIC32 are the key reason we don't have a range of manufacturers making MIPS microcontrollers - as there are MIPS cores that are equivalent or better than ARM cores across the range of sizes and powers.

In terms of architecture, MIPS has a different background from PowerPC (or Power before it), but there is definitely a strong similarity.

MIPS is not open source, but easily licensable at different levels and for different cores. I think you can use the ISA in your own designs without a license (there are some Chinese MIPS ISA cores). You can do that with SPARC too, if anyone but Fujitsu cares.

Well emulating a processor on another - like Intel do - is nothing particularly new, I myself have emulated the 6809 on power under dps (had some code written in the mid 80-s which I still can - and do - use). However no matter how efficient your emulation is it is an emulation and has its performance cost. Under equal circumstances power can beat Intel hands down. No such comparison can be made though, if you compare this or that piece of C software you will compare a compiler to another rather than a core to another - so Intel will likely win simply because of its huge popularity, the man hours put into compilers for x86 must outdwarf these put into power compilers. So a performance comparison just cannot be done in a decent way.

At the point where Apple abandoned PPC and went Intel what I remember was Jobs saying Freescale (or wast it still Motorola SPS) or IBM or both could not deliver the silicon they wanted, i.e. it was a production complaint. I am not sure I buy into that, I have my conspiracy theory about what happened but well, it is just a theory of mine, I can't prove it and it might well be wrong on top of that.

Oh no, they have 64 bits for years now. Just check the website, they have up to 8 physical cores and 16 logical ones (or was it 4/8), just look for QORIQ power. Really powerful beasts they are, GHz range, 750 MHz clocked DDR4 etc., that on a SoC.

PowerPC (PPC) was dropped some 10 years ago as a name, it is called "Power Architecure" nowadays. Not to be confused with IBM's Power from the 80s, which is sort of the same but not quite, this is where the architecture originated (the guy who did it has been really good). Of course people use PPC and power interchangeably nowadays so unless you are intimate with what is on offer it can be a source of confusion.

Dimiter

====================================================== Dimiter Popoff, TGI

Hi Rick, all,

First, thanks to all for replying and taking up this thread. This is very interesting.

I wasn't aware that MIPS ISA has been open-sourced. According wikipedia, the descision to open-source was taken in 2018, to be executed in 2019.

Note that they just call it "open" and "open use", not open-source and

-as far as I see it- it is not open source. E.g. the license agreement states that the technology (including the ISA) is still subject to US export-control, which makes the license incompatible with any open-source license. It looks more like 'freeware' then 'open source'.

Note that their website also mentions two cores: the microAptiv UP and UC cores.

This makes them much less open then (say) risc-v or the opensparc licenses.

Kristoff