5 ms·

> when the whole reason Apple shifted to Intel 14 years ago was thermals? Wouldn’t ARM be better? The move to Intel wasn't because the instruction set is, for

> when the whole reason Apple shifted to Intel 14 years ago was thermals? Wouldn’t ARM be better? The move to Intel wasn't because the instruction set is, for some reason, superior (it isn't), it's because IBM couldn't keep up with Intel's fabrication technology. The G5 chip from 2002 was at 130nm, while the Intel design introduced in 2006 (when 4 year old G5 was the best IBM had to offer) was at 65nm. The chip they aim for is built on a 28nm process (which has some impact on voltages used and minimum conversion of power into heat). That's definitely not top-class, but good enough: it's about a factor 3 to the state of the art (compared to factor 2 between G5 and Intel Core in 2006), but in absolute terms it's so much closer. There are enough ARM laptops out there, this provides another venue for more vendor independence (ARM still exerts tight control over its ISA while PowerPC is a close cousin to Power9 which was opened recently). In short, for a project like this, the goal isn't the very best performance on the market, it's "the more the merrier".

gridlockd 7y ago
"The more the merrier" isn't true for proprietary ISAs at all. Think of all the work going into supporting it, of the limitations coming with a very exotic computer. These are things humans have to spend lifetime on. I think it's a huge waste. If this was about pushing RISC-V forward, I wouldn't say that, but only because I hope it takes over, not because I think having yet another ISA in the zoo is somehow good.
- binarycrusader 7y ago
  But POWER isn't a proprietary isa anymore? https://www.servethehome.com/ibm-open-sources-power-isa-and-aligns-openpower-to-linux-foundation/ https://www.servethehome.com/ibm-open-sources-power-isa-and-...
  - gridlockd 7y ago
    Oh yeah, I guess I remember hearing about that now. Well, in that case I don't like it because IBM made it. Also it's old. It's not coming back. RISC-V is new and sexy.
    - dragontamer 7y ago
      > Also it's old. It's not coming back. The #1 Supercomputer in the world is using POWER9: Summit. Granted, its mostly because OpenCAPI allows for POWER9 + NVidia GPUs to interface at extremely high speeds, but a win is a win in my books. https://www.olcf.ornl.gov/summit/ https://www.olcf.ornl.gov/summit/ PowerPC has silly things like decimal-floats support and quad-precision support. I mean, they exist for emulating old COBOL code for the most part, but decimal-floats solve the age-old problem of how to represent financial data with floating-point, and with perfect accuracy at that. There's certainly good and unique ideas in the PowerPC ISA.
      - ajdlinux 7y ago
        Minor correction: Nvidia GPUs on POWER9 are connected using NVLink, which shares its physical layer with OpenCAPI, and like OpenCAPI is cache-coherent, etc, but uses a different data link and transport layer protocol. (I'm one of the OpenCAPI Linux kernel driver maintainers)
    - pgeorgi 7y ago
      > RISC-V is new and sexy. RISC-V is pretty shoddy (M mode, a separate hypervisor mode instead of proper orthogonality, CSRs), but that's exactly what I'd expect from thesis-driven development (remember that RISC-V was an academic project for a long time). Holding their feet to the fire so that they undo some of that damage seems to be a pretty good idea to me, and if it's PowerPC what it takes to do that, that's not the worst choice IMHO.
      - gridlockd 7y ago
        I'm sure your technical arguments are well-reasoned, but I'm not changing my mind. I already made a modest emotional investment into RISC-V.
      - monocasa 7y ago
        M mode isn't a hypervisor mode. It's more like x86 System Management Mode, or ARM's EL3. Hypervisor was originally going to be a separate H mode, but they made it orthongonal with S mode instead as a sort of bank switch. And if CSRs are bad, then POWER's SPRs are bad; they're basically the same thing except POWER makes much less effort to be consistent or introspectable. POWER is way more braindead in it's supervisor state. For instance they finally gave you a sane radix tree page table with POWER9, but now you can't address all the virtual address space you could before, which, yes, is its own problem for some systems.
IOT_Apprentice 7y ago
Ok. but 28nm is nowhere close to where ARM is (7nm) and will be over the next few years (3nm). Smaller, faster, with multiple cores & on-board ML/Neural subsystems. The cost on the ARM side is always going to be lower due to the volume at which they are being dropped into phones. And if one wanted a faster x64 path, then why not AMD Rizen?
- pgeorgi 7y ago
  ARM isn't at 7nm, TSMC is. That's different from the PPC -> Intel transition at Apple 15 years ago: the tech is available to anybody who's willing to pay (while back then, Intel was 1-2 generations ahead of everybody else, and there was no indication that this would change). The primary goal of a project like this is to prove the concept, and if 28nm NXP PPC chips are the sweet spot (enough performance for it to matter, unlike all those RISC-V micros; efficient enough to fit in a laptop without it working double-duty as a grill, unlike Power9 CPUs; cheap enough so that people are willing to risk it, unlike the otherwise fabulous TALOS workstation family) for that, that's not too bad. In case this takes off, there can still be some entrepreneur who cooperates with IBM or NXP or whoever has an efficient Power implementation lying around and shrink it to the smallest process node they can afford, reusing this open hardware design, and send it off to TSMC. That option didn't really exist back in 2004.
- filmgirlcw 7y ago
  Right. Like putting aside everything else, 28nm in 2019 going into 2020? Go ARM or Ryzen.

gridlockd 7y ago

"The more the merrier" isn't true for proprietary ISAs at all. Think of all the work going into supporting it, of the limitations coming with a very exotic computer. These are things humans have to spend lifetime on. I think it's a huge waste. If this was about pushing RISC-V forward, I wouldn't say that, but only because I hope it takes over, not because I think having yet another ISA in the zoo is somehow good.

binarycrusader 7y ago
But POWER isn't a proprietary isa anymore? https://www.servethehome.com/ibm-open-sources-power-isa-and-aligns-openpower-to-linux-foundation/ https://www.servethehome.com/ibm-open-sources-power-isa-and-...
- gridlockd 7y ago
  Oh yeah, I guess I remember hearing about that now. Well, in that case I don't like it because IBM made it. Also it's old. It's not coming back. RISC-V is new and sexy.
  - dragontamer 7y ago
    > Also it's old. It's not coming back. The #1 Supercomputer in the world is using POWER9: Summit. Granted, its mostly because OpenCAPI allows for POWER9 + NVidia GPUs to interface at extremely high speeds, but a win is a win in my books. https://www.olcf.ornl.gov/summit/ https://www.olcf.ornl.gov/summit/ PowerPC has silly things like decimal-floats support and quad-precision support. I mean, they exist for emulating old COBOL code for the most part, but decimal-floats solve the age-old problem of how to represent financial data with floating-point, and with perfect accuracy at that. There's certainly good and unique ideas in the PowerPC ISA.
    - ajdlinux 7y ago
      Minor correction: Nvidia GPUs on POWER9 are connected using NVLink, which shares its physical layer with OpenCAPI, and like OpenCAPI is cache-coherent, etc, but uses a different data link and transport layer protocol. (I'm one of the OpenCAPI Linux kernel driver maintainers)
  - pgeorgi 7y ago
    > RISC-V is new and sexy. RISC-V is pretty shoddy (M mode, a separate hypervisor mode instead of proper orthogonality, CSRs), but that's exactly what I'd expect from thesis-driven development (remember that RISC-V was an academic project for a long time). Holding their feet to the fire so that they undo some of that damage seems to be a pretty good idea to me, and if it's PowerPC what it takes to do that, that's not the worst choice IMHO.
    - gridlockd 7y ago
      I'm sure your technical arguments are well-reasoned, but I'm not changing my mind. I already made a modest emotional investment into RISC-V.
    - monocasa 7y ago
      M mode isn't a hypervisor mode. It's more like x86 System Management Mode, or ARM's EL3. Hypervisor was originally going to be a separate H mode, but they made it orthongonal with S mode instead as a sort of bank switch. And if CSRs are bad, then POWER's SPRs are bad; they're basically the same thing except POWER makes much less effort to be consistent or introspectable. POWER is way more braindead in it's supervisor state. For instance they finally gave you a sane radix tree page table with POWER9, but now you can't address all the virtual address space you could before, which, yes, is its own problem for some systems.

IOT_Apprentice 7y ago

Ok. but 28nm is nowhere close to where ARM is (7nm) and will be over the next few years (3nm). Smaller, faster, with multiple cores & on-board ML/Neural subsystems. The cost on the ARM side is always going to be lower due to the volume at which they are being dropped into phones. And if one wanted a faster x64 path, then why not AMD Rizen?

pgeorgi 7y ago
ARM isn't at 7nm, TSMC is. That's different from the PPC -> Intel transition at Apple 15 years ago: the tech is available to anybody who's willing to pay (while back then, Intel was 1-2 generations ahead of everybody else, and there was no indication that this would change). The primary goal of a project like this is to prove the concept, and if 28nm NXP PPC chips are the sweet spot (enough performance for it to matter, unlike all those RISC-V micros; efficient enough to fit in a laptop without it working double-duty as a grill, unlike Power9 CPUs; cheap enough so that people are willing to risk it, unlike the otherwise fabulous TALOS workstation family) for that, that's not too bad. In case this takes off, there can still be some entrepreneur who cooperates with IBM or NXP or whoever has an efficient Power implementation lying around and shrink it to the smallest process node they can afford, reusing this open hardware design, and send it off to TSMC. That option didn't really exist back in 2004.
filmgirlcw 7y ago
Right. Like putting aside everything else, 28nm in 2019 going into 2020? Go ARM or Ryzen.