[rescue] sparc10 cpu - what to do.

Tue Dec 20 15:05:57 CST 2016

On Fri, Dec 16, 2016 at 4:55 PM, Sandwich Maker <adh at an.bradford.ma.us>
wrote:

> the triumph of the marketplace -- best doesn't win; popular wins.  and
> popular affects scale, which drives price, and price is an easy sell.
> quality - however you measure it - is more nebulous.
>

There is an interplay between the very high end (where best performance and
to an extent best design matters) and the general marketplace. With the
export restrictions placed on China regarding Intel CPUs, they have
invested even more in their own (probably variant) architectures:

https://www.nextplatform.com/2016/06/20/look-inside-chinas-chart-topping-new-
supercomputer/

This type of R&D will likely have a trickle-down effect as China will move
toward producing more of it's own hardware across the board, which
ultimately will have a global impact (though we might not see that in the
West immediately).

We are seeing the impact of mobile chips on supercomputers now, very
explicitly in the case of the the European Mont-Blanc Supercomputer
Project, which used the Samsung Exynos 5 (ARM Cortex A15 1.7 GHz dual core
CPU + ARM Mali T 604 GPU, total of 2160 ARM CPU cores and 1080 ARM GPUs):

https://arxiv.org/ftp/arxiv/papers/1508/1508.05075.pdf

This 2012 analysis does some architectual comparisons of power efficiency
in supercomputer designs:

https://www.ida.org/idamedia/Corporate/Files/Publications/IDA_Documents/ITSD/
ida-document-ns-d-4857.ashx

pa-risc is gone.  alpha is gone.  mips is radically retargeted.  if
> sparc goes, power may be the only alternative performance chip left.
> itanium?  i was at hp when the first gen was launching [disastrously]
> and i understand why hp threw in with intel, but i predicted then that
> intel would eat hp's lunch, and it seems to be coming to pass.

I think in the long game, ARM is going to be the one to beat for a general
purpose CPU. Fabless is a distinct advantage. Intel's advantage of having
some of the best fabs and process node technology in the world can only
hold back the tide for so long, and can (and has) resulted in them resting
on their design laurels when it comes to x64. The shifts in the marketplace
to handheld and mobile have repeatedly shown that despite fab and process
node superiority, Intel is not a real competitor when it comes to the best
performance/power ratio in the low power (mobile) marketplace. AMD Zen may
give Intel a jolt and make x64 competitive again, but for x64 to survive in
the long term it needs to be more competitive, and I wonder if the VLIW-ish
core (with the private micro-ops instructions) will always mean there will
be a certain level of complexity to x64 architecture which ARM will never
have and will always mean that ARM can ultimately be more efficient.

The only significant evolution of x64 in the HPC space has been the
Larrabee family. The 'Knight's Corner' generation of Larrabee got branded
as Xeon Phi and made it into several supercomputers, including the Tianhe-2
in 2013 (which still holds the #2 spot in the top 500 and led to the
embargo). The upcoming generation of 'Knights Hill' is going into the
upcoming Argonne National Laboratory bAurorab system:

https://www.nextplatform.com/2016/07/12/knights-landing-will-waterfall-high/

Side note, it will be running Linux (like most supercomputers these days):

http://www.intel.com/newsroom/assets/Intel_Aurora_factsheet.pdf

Meanwhile Fujitsu confirmed this summer that Sparc was dropped from the
design of the next K supercomputer and it will be based around ARM:

https://www.nextplatform.com/2016/06/23/inside-japans-future-exaflops-arm-sup
ercomputer/

Microsoft just announced that they are going to release a desktop version
of Windows 10 for 64-bit ARM that will support realtime emulation of 32-bit
x86 code, which will provide a compatibility bridge that Windows RT never
had:

http://arstechnica.com/information-technology/2016/12/microsoft-pushes-for-a-
more-modern-pc-with-vr-project-evo-and-arm-support/

Meanwhile, there has been ongoing speculation that in the long-term Apple
will transition the Mac desktops to ARM. Apple has seriously neglected much
of the desktop hardware lineup, especially the Mac Pro and Mac Mini, and
it's not clear why. Perhaps it is just Tim Cook being business-logical,
since most of their profit comes from iPhones and iPads, that's where
they're deciding to focus on most of the R&D (and due to their tight
control over the custom design of the CPU and GPU, they can maintain a much
more significant edge in that marketplace, while they are forced to use
largely off-the-shelf designs from Intel on the Mac side). Apple was able
to surprise the world with the release of their 64-bit ARM implementation,
while there is no way they could achieve the same thing with an x64
product. And Apple's recent Touch Bar feature of the MacBook line? That's
an independent ARM system running WatchOS (an iOS variant) that boots from
it's own hidden EFI partition stored on the internal SSD. It's pretty clear
where Apple is focussing it's attention.

Side-note: while I think there will definitely be trickle-down from HPC to
the desktop (we've been seeing that with much of the evolution of things
from NVIDIA and AMD directly the last few years) and trickle-up (power
efficiency that's been so critical to success in the growing mobile market)
that leads to ARM, the role of the traditional CPU/GPU itself may be less
significant over time in the HPC arena. Nervana Systems was recently
acquired by Intel for an undisclosed sum, and indications are they may have
a novel architecture for machine learning, which is becoming increasingly
important:

https://www.nextplatform.com/2016/08/08/deep-learning-chip-upstart-set-take-g
pus-task/

And there is also a Xeon Phi designed for deep learning, 'Knights Mill'.
And Intel has shown off designs putting an FPGA in the same CPU package as
an additional die next to a Xeon:

https://www.nextplatform.com/2016/03/14/intel-marrying-fpga-beefy-broadwell-o
pen-compute-future/

Meanwhile DoE labs has been talking about a "novel architecture" for
supercomputing to arrive in 2021, with a focus on how a new supercomputer
does across a broad set of problems, not just hitting a top 500 score:

https://www.nextplatform.com/2016/12/08/u-s-retracts-exascale-timeline-focuse
s-novel-architectures-2021/

... with China at the top now, and with all the R&D focused on deep
learning in the last year, maybe there's something to the "novel
architecture" (or maybe it's political posturing, or both...) We are coming
off of a year where China went from 108 systems in the top 500 to 171,
pushing the US down to 171:

https://www.top500.org/lists/2016/11/

That's quite a shift in one year.

Side side note: Regarding Android and open source, we may see the entire
platform go closed source in the near future. Speculation on that here:

https://www.androidpit.com/google-and-android-here-comes-the-lockdown

- Nate