Hacker News Reader

Comment by brey | Hacker News Reader

I wonder how long until we have that kind of computing power on our phones and look back with a wry smile ...

or - how long ago would $173M have bought you the kind of power we carry around today in a pocket?

While not quite $173MM, the 1985 Cray 2 cost $32 million in 2010 dollars[1], weighed 2500kg and consumed 150-200 kilowatts in power.

Estimates put the processing power of an iPhone 5 at at around 2.7x times that of the Cray 2[2]

[1] http://www.theregister.co.uk/2012/03/08/supercomputing_vs_ho...

[2] http://pages.experts-exchange.com/processing-power-compared/

Wish there was a good book about the history of crays. It's the kind of thing you hear talked about but are unlikely to ever see.

Mainframes in general too.

I'm currently listening to Soul of a New Machine.

The Supermen: The Story of Seymour Cray and the Technical Wizards Behind the Supercomputer, Charles J. Murray, 1997

Next time you're in the Bay Area, don't miss the Computer History Museum in Mountain View. Lots of Cray action there.

Apparently there is now a working CDC 6500 (designed by Cray) at Paul Allen's computer museum in Seattle.

> Mainframes in general too.

Mainframes are great, already using almost[0] memory safe systems programming language on the 60's with Burroughs, followed by IBM and a few other vendors.

Virtualization and containers with the 360.

Bytecode as universal binary format with JIT/AOT at kernel level, DB based file system, System/38 and AS/400.

Object based OS, AS/400.

[0] - They still have the issue of leaks and double free though, but everything else is safe Algol style with explicit unsafe blocks/modules required.

Talk about a commitment to backwards compatibility – you can run binaries built 30 years ago for god knows what proprietary processor on a modern POWER8 system without recompiling.

I also find it kind of amusing that IBM, a primarily consulting company, developed AS/400, given that part of its sales pitch is that integrated database requires no maintenance and you can forget entirely about your IBM i and just leave it running for a decade.

It's a neat system. I wish I had the opportunity to use one. In many ways it feels like we're still catching up to what System/38 was doing in 1979.

I was responsible for doing backups on a AS/400 during a Summer internship in the early 90's. Sadly did not do much more than exchanging the tapes, logging in and starting the backup.

> "Wish there was a good book about the history of crays. It's the kind of thing you hear talked about but are unlikely to ever see."

That would be a great read.

My favorite Cray tidbit: for fun, Seymour Cray dug tunnels underneath his home, and had a lot of his breakthroughs while doing so.

   he attributed the secret of his success to "visits by elves" while he worked in the tunnel: 
   "While I'm digging in the tunnel, the elves will often come to me with solutions to my problem."

https://en.wikipedia.org/wiki/Seymour_Cray#Personal_life

He'd also famously construct a sailboat, sail around in it, and then torch it at the end of the season so he'd have to construct another one.

My favorite Seymour Cray anecdote is from Alan Kay when he did an AMA here on hn a few months ago:

"Seymour Cray was a man of few words. I was there for three weeks before I realized he was not the janitor."

https://news.ycombinator.com/item?id=11941941

There were 27 years between the release of the iphone 5 and the Cray, and at a $649 release price, it was approximately 50,000 times cheaper.

So, if this was linear, we can all expect 130 petaflop computers around 2043 for around $3,500?

Lots of caveats here though, things aren't usually as linear as all this, and this is very much a back-of-a-napkin calculation.

Nobody needs more than 640TFLOPS.

-Bill Gates?

I think it's speeding up. 1975 Cray-1 had approximately same 160 MFlops as first iphone had. And it was 32 years between them. So, probably we'll have 130 petaflops around 2033-2035? :)

For many services the phone is just a relatively dumb front end, and the actual computing happens on cloudy clusters elsewhere.

Given that Moore's Law is creaking I wouldn't expect pocket petaflops any time soon. I'd expect a serious outbreak of cloudy clusters everywhere, and perhaps a dynamically reconfigurable Internet 2.0 with completely transparent non-localised computation.

This might change if computing finally goes optical and/or quantum. But if we're pushing electrons around wires, current hardware is close to the physical limits. The only way to speed it up is to build a lot more of it and speed up the connections.

Note that these measures are biased by bandwidth. CPUs today have massive peak FLOPS, but limited sustained bandwidth, while the Cray being designed with a very large one to ensure regular performance.

You are probably overestimating the Cray, modern CPU's have a lot of bandwidth. And due to being physically smaller fewer issues with latency, and fairly large on chip cache.

Someone did the math and a Cray was on par with an intel core i3 series (early gen) in sustained perf even though in terms of raw GFLOPS the intel cpu was two leagues above. That's all I'm saying.

I'll link the url when I've scanned my bookmarks.

psedit: look at jojomonkeyboy's comment http://www.techrepublic.com/blog/classics-rock/the-80s-super... he says an i7 2600 (not an i3) has less sustained compute power.

That's not a Crey 2. That is talking about the X-MP it's successor.

https://en.m.wikipedia.org/wiki/Cray_X-MP "The Cray-2, a completely new design, was introduced 1985. A very different compact four-processor design with from 64 MW (megaword) to 512 MW (512 MB to 4 GB) of main memory, it was specified to 500 MFLOPS but was slower than the X-MP on certain calculations due to its high memory latency.

The X-MP-succeeding Cray Y-MP series was announced in 1988; it also had a new design, replacing the 16-gate ECL gate arrays with a more compact VLSI gate array with larger circuit boards. It was a major improvement of the X-MP supporting up to eight processors."

Note latency is huge for these systems meaning for most workloads modern cellphones absolutely crush them.

Err, I mean they where talking about a "Cray Y-MP C90"

While that machine was a beast at the time it cannot compete against a modern iPhone. There's just too many things working against it.

I guess that's theoretically possible, according to Landauer's principle [1]. It's hard to say if it will ever be feasible. It feels like we are running into some serious limitations, since CPUs don't seem to be getting any faster. The CPU will probably have to be totally redesigned from the ground up, using different chemicals, and perhaps light instead of electricity. Or maybe some insane molecular structure computed and created via genetic algorithms and machine learning. Evolution did it to form our brains, so it's definitely possible.

I'm not sure how to do the calculation to find the minimum amount of power required for 130 petaflops, but it seems like it's well within the capacity of today's mobile batteries.

[1] https://en.wikipedia.org/wiki/Landauer's_principle

We don't need that much power to run our Facebook, Camera and Mail apps. Really, we have fast enough hardware, good enough screens and headphones. You can't hear over 16khz and can't see over 400dpi, so they can't improve image and sound any more.

What we need to advance is specialized Deep Learning cores that are compact, fast and low power, so they don't kill the battery. On the other hand, for general apps it's much less necessary to make the CPU faster.

Assuming Moore's law holds and that a $50 is a suitable cost for a smartphone SoC, about thirty-two years.

That's a big assumption. I think few people believe we'll have another 32 years of Moore's law.

That's the best estimate I can provide. Besides it maps well with the past thirty years of computing history.

And ignores what Moore's Law is.

In a (simplified) nutshell: We can fit more transistors in a given unit of area because we make them smaller. That used to just mean we increased clock frequencies (make it faster) but comparatively recently (decade or two) meant we increased parallelism.

Moore's Law is expected to fail because we are now reaching the point where smaller transistors are very heavily impacted by actual limitations imposed by physics.

So while it is possible we'll have a technology shift and see similar performance gains, it won't really be Moore's Law anymore (unless we start using Pym Particles or something).

> So while it is possible we'll have a technology shift and see similar performance gain

What kind of technology shift do you mean? Like a totally different computing paradigm?

There is some work regarding making transistors out of different materials as a way to eke a bit more out. Similarly, there is a lot of work regarding layering circuitry to an even greater degree. And, of course, there is the usual pie in the sky solutions.

But none seem all that promising and my gut is that we'll focus more on interconnects and algorithmic improvements.

But time will tell.

Computing using coupled magnetic spin. Or photonics. Or nanomechanical rod logic. Or nano-electro-mechanical logic. Or ballistic electrons. Or single electrons in nano-structures similar to what a cell uses in the Krebbs cycle.

>Pym Particles

For those reading along, this is a fictional particle named after Hank Pym - AntMan - from Marvel comics. It's not a technology in a lab somewhere.

The economic implications of Moore's Law are quite clear to me. Roughly one observes that every 18 months the expected computing power one can buy per unit currency doubles. Of course this is predicated upon physical possibility. 32 years would imply another 21 process halving or so, and that would take features down from 14 nm to an untenable 0.7 nm. However, if you have a better estimate, I'd be glad to hear it.

Yes. Economically, my bank account gets larger every year. Ergo, I must be saving up all my birthday cash

I don't have an estimate. My point was simply to explain to you why your logic was flawed as we are nearing the limits of what Moore's Law can give us without some pretty massive changes. This isn't a case of "Clock speeds are capped. We are doomed. Oh, wait, we can just put two slower ones on the same die" and is more "So... we are out of physical space..."

Trends are great when you are trying to make sense of data and estimate how to move forward. But they should not be used in a manner that ignores actual data.

Note that 21 halvings doesn't take you from 14nm to 0.7nm, but rather to about 7 femtometers, or about 0.000007nm. For comparison, that's roughly 10x the size of a single proton.

I'd say a better estimate would be to assume density stops increasing around the point when feature size is the size of a silicon atom. I'm sure that'll be way off, but closer than estimating 21 more doublings.

It also ignores that we're likely entering a period of anti-rational and anti-science thought, technology may have a few pull-backs before continuing its rise.

I think this is a relatively mild period of anti-science. Religion is largely on the decline. Fake news is being called out as "fake". College educations are valued even by those without them in general, and the job market certainly values them.

I think it seem like there is a much larger anti-science sentiment then there is because these people have been given a fresh voice with social media and for the first time in a long time they can connect with each other and build echo chambers to shout at eachother in.

Some of this spilled out in the last election and provided a non-trivial number of votes for a candidate who was clearly a demogogue, instead of voting for a different demogogue who appealed less to the uneducated.

It's just too much computing power – there will come a time when many things have enough.

If you look at Apple's recent offerings, it would seem they think that time is more or less here.

I certainly almost never use the full computing power of anything I'm using – the limiting factors aren't hardware any more but the software running on it.

We're still coming up with everyday tasks/usage that require more and more power, storage and internet bandwidth. First everything was text. Then we started using photo/audio, then we started using low res video, then video resolution kept enlarging (though recently this slowed down, we got to 1080p pretty fast, switching to 4k isn't as fast), now we invent VR, probably some 3d/holograms format will follow. I don't know what can come after that, maybe we'll finally will have enough power/bandwidth/storage :)

We're reaching the flat part at the top of the sigmoid innovation curve. We're still using 1960s technology. It's just been through a good few generations of refinement.

VR might just about squeeze through now, but given the hardware limitations - and the fact that people look really dorky using it - I'm not expecting it to drive a new explosion of user interest.

We really need some completely new tech to drive a new wave of innovation. The obvious candidates are optical/quantum and perhaps direct neural interfacing. Both are still science fiction, but that may change by 2030.

More extreme technologies may also be possible, but they're beyond speculative.

For now it may be useful to remember that technology rarely develops linearly, so speculating about future CPUs is like speculating about the future of transatlantic cruise liners, while ignoring the fact that someone somewhere is working on heavier than air flight.

> perhaps direct neural interfacing

I would really like to be excited by the prospect of neural interfacing, but all I can imagine is people catching computer viruses.

Watch the "Ghost in the Shell: Stand Alone Complex" series for an excellent science fiction exploration of this

you might not, but many gamers, streamers, audio/video creators create content today, often 4K (or even higher) which requires a lot of processing power and internet bandwidth, so assuming that demand will not go up is a bit naive. For everyday tasks, sure what we have is fine, but as soon as you move into producing or consuming complex media (which is more and more mainstream these days) i don't think it will stop anytime soon.

Reply on news.ycombinator.com