discussion.mdwn

   1 # discussion with Roberto from PowerPC Notebook, 09 Apr 2020
   2
   3 basically, you remember EOMA68, i decided to go further down the rabbit
   4 hole after 10 years waiting for fabless semi manufacturers worldwide to
   5 FAIL to produce a modern SoC that has libre GPU *and* VPU drivers.
   6
   7 i therefore began designing a hybrid CPU GPU VPU which unlike a "normal"
   8 GPU which is totally separate and usually  on a PCIe bus, you actually
   9 add GPU and VPU instructions to the *main core's* instruction set.
  10
  11 examples of this include ICubeCorp's IC3128 and the Broadcom Videocore
  12 IV which is actually an ARC Core (ARC was a competitor to ARM that didn't
  13 take off in the same way).
  14
  15 so we started with RISCV on the reasonable assumption that "Open"
  16 actually means "Open", only to find that they are "Fake Open Source".
  17
  18 we therefore had to scramble to find an alternative, which just happened
  19 to be right around when the OpenPOWER Foundation started kicking into
  20 gear.
  21
  22 we're therefore beginning to gear up to POWER9 V3.0B and tentatively
  23 negotiating and opening communications channels with the right (extremely
  24 busy) people, with the kind assistance of Hugh Blemings, Director of
  25 the OpenPOWER Foundation.
  26
  27 this does have to be done carefully because unlike a traditional GPU
  28 (which we do not want to do because the inter-processor communication is
  29 nuts and massively complicates drivers on both sides, CPU *and* GPU) we
  30 are using the same pipelines, the same L1 caches, everything, for *all*
  31 workloads and thus have to be very careful about resource utilisation.
  32
  33  however more than that, because we seek to *extend* (augment) the
  34  instruction set (add ATAN2 to POWER9 for example) this has to be managed
  35  extremely carefully.
  36
  37 normally in a proprietary product there would be absolutely no public
  38 evidence, documentation or publishing of the toolchain for such extensive
  39 custom modifications.  they would definitely not be permitted to be
  40 upstreamed.
  41
  42 however here, with the business justification behind libre (see Intel
  43 Valve Steam collaboration) we have to extend the instruction set *and*
  44 document those augmentations *and* publish the toolchain *and* have it
  45 upstreamed *and* make sure that the entire OpenPOWER Community is ok with
  46 that, and all the implications and ramifications properly thought through.
  47
  48 so that's a potted version of why and what :)
  49
  50 it means that in say 12 to 18 months we will have a quad core SBC-style
  51 processor that you could use for a powerpc netbook.  at last! :)
  52
  53 best,
  54
  55 l.