Maynard Handley (name99.delete@this.name99.org) on August 15, 2014 4:41 pm wrote:
> David Kanter (dkanter.delete@this.realworldtech.com) on August 15, 2014 3:01 pm wrote:
> > juanrga (nospam.delete@this.juanrga.com) on August 15, 2014 11:39 am wrote:
> > > David Kanter (dkanter.delete@this.realworldtech.com) on August 15, 2014 9:41 am wrote:
> > >
> > > > I also am skeptical that anyone can match Intel in performance while trailing behind by a node. To do that,
> > > > it is necessary to offer a very different product with a
> > > > different system architecture. E.g., target workloads
> > > > where cache does not help and simply slap down more memory controllers and cores (hint: that's a GPU!).
> >
> > > The ISA advantage will be greatly reduced in the top-end
> > > side of the performance spectrum, but will not vanish.
> > > Keller mentioned during Core Day conference that his K12
> > > core will have a "bigger engine" than its x86 sister
> > > thanks to the advantages of ARMv8 over x86-64, which allows to spend more transistors on compute.
> >
> > I happen to know the differences between those two designs. I'm not really sure it's
> > going to translate into a significant performance delta. My guess is maybe 10%.
>
> 10% is two years of Intel's current 5% improvement a year.

And what if final number is more close to 20--30%?

> Maybe that's why there's so much disagreement here? Some of us see "two years"
> and think that's massive, some of us see "10%" and think that's nothing?
>
> ..
>
> I don't want to get involved in the fight over whether various forthcoming ARM64 CPUs are "real" or "servers",
> but I do get the feeling that most of the vitriol is being generated as a result of miscommunication.
>
> In particular, I get the feeling that the x86 partisans have a particular model in mind for what a
> "server" IS, and are vociferous that nothing in the ARM camp meets that definition. (Of course it
> doesn't, if the definition includes "runs x86 code", but what we're seeing is more subtle than that.
> The expectation is that a server is a certain type of box, running certain types of code, with certain
> types of IO, costing a certain amount, and being sold to a certain type of customer. )
>
> On the other hand, the ARM proponent are trying to be a lot more catholic in their tastes. They see
> a "server" as a much more generic term, and see everywhere they look, jobs that could be done perfectly
> well with boxes that are cheap, that don't demand particular RAS features, that can run fairly generic
> software, that can be unbalanced in various ways (whether it's amount of RAM, or RAM bandwidth, or
> IO or whatever). They see a world which needs ever more CDNs, ever more memcached boxes, maybe boxes
> that sit between IPv6 and IPv4 networks and translate between them, maybe boxes that do nothing but
> scan twitter feeds looking for keywords and correlating activity, etc etc.
> Point is, they see a world of computing tasks to be done, and which can be done adequately, and a lot more
> cheaply, on various models of these upcoming ARM "server" SoCs. In many ways this is no different from the
> point in the mid-90's where first Google and then a few other internet companies made a big deal about the
> fact that they did not need (and were not prepared to pay) for "server" quality hardware --- they were quite
> happy to put together boxes that did what they needed and did not do a damn thing more.
> The frustration arises because the pro-x86 people appear to refuse to engage with this possibility


I cannot tell for others, but I have been using the term "server" in a standard way:

Rather than just web serving, these systems are being built to also power data analytics on Hadoop clusters, fetch and put data in NoSQL data stores, streaming media and high-performance computing, sharing processing duties with GPUs, FPGAs or ASICs.

http://www.techrepublic.com/blog/data-center/arms-battle-for-the-datacentre-the-contenders/

During presentation of new ARM-HPC systems Nvidia and Applied Micro claimed that first gen ARM SoCs "can drive full GPU performance" and Applied Micro added that

will be able to show that an X-Gene 1 plus a Tesla K20 coprocessor will be equivalent to an X86 processor plus the same Tesla K20 floating point motor.

http://www.hpcwire.com/2014/06/23/gpus-add-arm-chips-hpc/


> (I'd draw an analogy to gaming, where the types of people who play a particular type of game game on souped
> up $3000 rigs simply will not accept that playing a game on your iPhone or Android tablet counts as gaming
> in any "real" sense --- and thus are mostly blind to how the world around them is changing.
> One can understand their anger/fear --- the sooner the world is content with that sort of
> gaming, the sooner the world will stop manufacturing, at least at prices and in form factors
> that are useful to gamers, the types of GPUs that hardcore gamers utilize. But pretending
> the world is not changing because you don't want it to has never served anyone well.)
>
> ..
>
> Finally, over the past three months or so, something appears to have broken on this board.
> The way I see it, the value (and joy) of this forum is that people can throw out ideas and have
> them critiqued. Some of the ideas we throw out will be flawed --- and with luck the arguments
> presented against them will convince us. Some of the ideas we throw out will be undecidable, at
> least for now, but hopefully we'll all learn something as they are attacked and defended.


It looks as I joined at wrong time then.


> An important point in all this, however, is that just because ideas are presented does not mean that
> they are strongly held, and does not mean that they need to be attacked like a rabid wolverine. Posts
> that consists of nothing but "It seems to me plausible, but I could be wrong, that maybe the right
> thing for Intel to do is..." or "Some people say ARM is like this, but then I've also read it's like
> that" are difficult and irritating to read. That is not the way you talk to your friends in real
> life, and it shouldn't (IMHO) have to be the way you talk on a forum like this.
>
> I'd urge everyone on this board to step back a little and reimagine what's happening here as something like
> a group of EE seniors chatting with each other in the dorm room late at night. We all have interesting FACTS
> that we can use to bolster arguments. We all have interesting OPINIONS about where things will (and should)
> go, opinions based on patterns we perceive, on our interpretation of history, on our understanding of the economics
> and business issues involved. We all have interesting QUESTIONS, some fact based (is this possible? has anyone
> ever done this), some opinion based (how would you handle this engineering/business problem).
>
> The joy of the conversation is in learning stuff along the way, it's not in "winning" by reducing your
> opponent to a pulp. Even if your opponent is a fool, in real life the way you would handle that is to
> explain (once, maybe with one followup) that he is wrong, and then to gently ignore him. Maybe take him
> aside and explain that his enthusiasm is appreciated but he's not really following dorm etiquette.
> EVERYONE needs to understand that the fun, in discussing "does ISA matter; will ARM64 servers take off;
> these features will definitely (not) be in the A8; etc" is in the first few posts where the initial arguments
> on each side are made. And EVERYONE needs to appreciate that once each argument has been made twice, it's
> time to move on. We'll revise the issue in a year when we have a year's more data, but until then nothing
> is achieved by repeating the same point a third time or more, or switching from technical discussion to
> impugning motives, ad hominem attacks, and accusations of stupidity/incompetence/lying.
>
> Let me end by giving just one example of all this. A few weeks ago Apple's Cyclone was being discussed. I said,
> based on my understanding (ie facts, opinions, my personal history), that this seemed like a 4-wide design to
> me. Reading the subsequence discussion, I've been converted to the opinion that it IS actually a six-wide design.
> That's valuable; that's how the system should work. I begin with a legitimate (but likely incorrect) opinion,
> and on seeing various good argument, I change to an opinion that's (likely) closer to correct.
> BUT
> I've also come to the opinion that the way Apple achieved this is not through a "simple" 6-wide
> design, but through a clustered design specifically two 3-wide clusters, presumably one basically
> a full 64-bit version of the Swift execution core, and the other slightly stripped down (missing
> division/sqrt and maybe various special purpose [OS, atomics, etc] instructions). This, I think,
> would have been interesting and pleasant for everyone to discuss, but I was reluctant to raise
> the matter given the unpleasant tone that so much of the discussion had taken.
>
> So that's my attempt to bring things back on track.
> September looks likely to be the most interesting CPU month in a long long time. Many people are
> going to want to say many things about a variety of new devices launching or about to be launched.
> It would be nice if we all treated what we learn/hypothesize about as entertaining and good-natured
> ribbing among friends, rather than a match to the death in which every piece of bad news for my
> tribe has to be crushed, along with those who support it, lest shame befall my god.