Slashdot Mirror
Opportunities From the Twilight of Moore's Law
saccade.com writes "Andrew 'bunnie' Huang just posted an excellent essay, Why the Best Days of Open Hardware are Yet to Come. He shows how the gradually slowing pace of semiconductor density actually may create many new opportunities for smaller scale innovators and entrepreneurs. It's based on a talk presented at the 2011 Open Hardware Summit. Are we entering an age of heirloom laptops and artisan engineering?"
As long as this technology is patentable, corporations will not allow it.
the market wont tolerate it, look at the 1980's everyone and their brother was making computers, often times using the same core parts but totally incompatible with each other. when the IBM clones started to hit market all those makers vanished. It was not because the IBM format was better, it was because everyone got even footing on a platform and their confidence in making an investment to hardware was reassured to not be worthless 6 weeks later.
We still see this today, ie ohh windows 8, oh desktop windows8 doesnt run the same shit as arm windows 8, well fuck that!
Was a wood and brass encased laptop with exquisite scrollwork around the keyboard and webcam, inherited by an archeologist who caries it around for data analysis and note taking.
My grandpappy made his own CPU's, you lazy whippersnappers! And if we're going to get back on top, American kids gonna start having to learn how to again. Now git' your lazy ass in that clean room and get to work!
SJW: Someone who has run out of real oppression, and has to fake it.
and i seriously dont think that moores law will end soon. Bumps in both directions, extending over some time are nothing unusual. New technologies will rise and Metal-Oxide-Semiconductor processes wont be dominating forever.
Good for them, but bad for everyone else. Users lose the continual improvements we're used to, manufacturers and retailers have to deal with people making their kit last years longer as they have fewer reasons to upgrade. Probably very good for Google and other companies who "rent" computers.
Some new bit of software will require some widget that can't be supported on an existing platform
and poof, there goes your old platform.
Thats what the development graphic of open stuff is like.
you can expect supercomputers made of open source chips in 10-15 years.
Read radical news here
The number of people predicting the end of Moore's Law doubles every two years.
You should at least open the bet with 'back then, we ate our own shit' -> then add 'young whippersnappers' and whatnot. noone can top 'eating our own shit back at that time'
Read radical news here
It seems computers have been stuck at 3GHz (plus or minus a bit) for a while.
Sure, we've added more cores and the like, but it's interesting to see that plateau at the end of the curve.
I'm sure some things actually are faster, but in terms of what's available to consumers, it hasn't seemed to get all that much faster the last few years.
Lost at C:>. Found at C.
In the beginning, hardware was not "open." Any antique radio collector will tell you that the schematics of 1920s radios were some of the best-kept secrets the manufacturers had (at the time), since the parts used in their products were readily available. Giving the user a schematic was viewed as a license to compete, and there were hundreds, if not thousands, of radio receiver manufacturers -- many of whom got started by reverse-engineering an existing design.
It was only in the 1930s that schematics began to be shipped with products, as the shakeout of the manufacturers meant that competition was based on economies of scale and factors other than just knowing how to build a radio.
By the time Moore's law slows down, we'll also have systems on a chip. Replaceable parts? We've moved the other way from the days you could solder chips and until today. Extension cards are almost gone, more and more of the north/south bridge and motherboard chips is moving into the CPU, now even the graphics card is moving into the CPU for many.
His argument sounds to me to use the same logic as arguing that once computers don't get faster, we'll have to make applications faster so we'll see a return of assembler language and hand optimization. Somehow, I don't think that's very likely. I'd make a fair bet that custom hardware is even more of a niche in 20-30 years than it is now.
Live today, because you never know what tomorrow brings
Do not tempt rule 34 ... I shudder to ponder it.
Lost at C:>. Found at C.
Firstly it's high unlikely that Moore's law will be retiring any time soon. All we are seeing is a slow down in the advancement of shrinking the manufacturing process. That doesn't say anything about performance improvements by other means. We are continually seeing advancement in performance per watt which is enabling CPUs to spread their dies not only "out" but even now we're seeing the prospect of "up" with promising research in layering techniques. Beyond that there are carbon rather than silicon based materials coming online that promise to further improve upon the performance per watt angle. We're even starting to see glimmers of hope in the quantum computing arena which would be game changing.
With respect to small companies being able to enter the market and compete with the "big guys" I would seriously doubt it. The first and obvious reason being the cost factor being a barrier to entry. The equipment isn't cheap and contending in the patent arena is worse. The most you'll ever see here is university level research being sold off to the big guys.
Two of my imaginary friends reproduced once
Are you a semiconductor or a hemiconductor
Moore's law has applied, and will apply - at least by inference - to all past and future computing paradigms.
The exponential growth trends of price/performance started long before CMOS processes were developed. While Moore's law specifically refers to integrated circuits, the facts remain: exponential growth trends were present in relay-based machines, vacuum tube based machines, transistor based machines (pre-IC), and integrated circuits.
In fact, the exponential growth trends are actually accelerating at an accelerating pace, as we are just now approaching the "knee" in the exponential curve.
The simple truth is today's ICs are manufactured at the nano scale ( 100nm), and will continue to shrink for several more generations of chips.
Before transistor size even begins to approach theoretical limits, new paradigms will emerge to replace current metal-oxide-semiconductor technologies.
We can already see this today. As we approach the limit to two dimensional ICs, we see the new emerging trend of three-dimensional circuits. We see the rumblings in research circles of optical systems at nano-scales. We're just now beginning to scratch the surface of quantum computing.
While Moore didn't invent the exponential, the trends he predicted more than four decades ago will be alive and well throughout the 21st century, even if by inference, as we transition away from CMOS to new, as-of-yet undiscovered paradigms.
To those seriously interested in this field, please consider reading Ray Kurzweil's "The Singularity is Near". You may not agree with everything the man has to say, but the man's data on this subject doesn't lie.
And it's far from the only thing that impacts a computer's performance. The advent of SSDs proves there are still major areas of computer performance to be addressed outside of the CPU. While the GPU will be brought onboard before system RAM is, I'm sure that's another area that will find its way onto the die eventually. Gigabit ethernet is good enough for now - I don't know of any broadband connections anywhere that exceed that just yet, but an increase to 10gigabit isn't out of bounds in the nearterm, and will last a very long time.
Once SSDs can affordably hold all the media we have and spinning disks go away completely, we'll pretty much be where we need to be for the foreseeable future. Modern GPUs are capable of moving 4K displays pretty easily, and the Ivy Bridge-era GPU will be capable of that even without a discrete card, so displays will be able to scale up easily. The advent of the 10" 2560x1600 panel are upon us - put three of those together for a nice 30" display, and you're good to go. The more things are offloaded FROM the CPU, the more irrelevant CPU density and speed becomes.
At 3 GHz light travels 10 cm in one clock cycle. Faster speeds would make it hard to send a request of data from one end of the chip to the other and get a response during the execution of a machine instruction. Having to insert no-operation cycles while waiting for data to arrive would negate the usefulness of faster clock rates.
funny... i was just writing up a post to the http://openhardwaresummit.org/ mailing list about a way to accelerate the process by which enthusiasts can work with the latest mass-produced embedded hardware.
the initiative, which has a specification here - http://elinux.org/Embedded_Open_Modular_Architecture/PCMCIA - is based around the fact that, just as mentioned above, the development of processor "speeds" is slowing down. this funnily enough allows so-called "embedded" processors to catch up, and it's these embedded CPUs which are low-power enough to base an entire computer around that is still desirable yet consumes between 0.5 and 3 watts instead of 10 to 500 watts.
if anybody would like to participate in this initiative, please do join the arm-netbook mailing list - http://lists.phcomp.co.uk/mailman/listinfo/arm-netbook
A quad core ARM with 3D video chip, some custom silicon to do specific tasks, like encode and decode ogg vorbis audio and theora video at 10 times realtime rates, a 100,000 neuron neural net chip and some multi trillion dollar software investments and we could have HAL on all our desktops.
FTA:
In the post-Moore’s law future, FPGAs may find themselves performing respectably to their hard-wired CPU kin, for at least two reasons: the flexible yet regular structure of an FPGA may lend it a longer scaling curve, in part due to the FPGA’s ability to reconfigure circuits around small-scale fluctuations in fabrication tolerances,
Uh, no.
FPGAs are built from the same elements as normal chips, transistors. Process improvements give us new transistors which are some combination of smaller, faster, and more power efficient than their predecessors (generally speaking, not all 3 at once). This doesn't favor either FPGAs or ordinary CPUs, as both are built from transistors.
His comment about reconfiguring around process variation is total nonsense. FPGAs are, if anything, more sensitive to variation, simply because the company making the FPGA never knows which part of the FPGA will become a critical timing path for the end user. Every programmable cell and every part of the routing matrix has to pass minimum standards. Otherwise, there's no way for the FPGA mfr. to guarantee that when their timing analysis software claims a customer design has timing closure, it really does.
Fixed function chips, on the other hand, have lots of known paths which have no problem meeting timing at any process corner, and thus can easily tolerate process variation. The key difference is known vs. unknown -- with FPGAs, all chips have to be capable of running all possible designs, so the only thing known when you make the chip is that it all has to be good. With ASICs, you know the characteristics of that fixed design and can design test programs for looser tolerances where it makes sense to do so.
In fact, Xilinx already has a program where they do roughly what he's talking about, but the implications are a bit different than he thinks. If you qualify for the program (you need to have enough sales volume, and you need to be using a relatively large and expensive FPGA, and you need to be willing to pay an upfront fee), you can submit your design to Xilinx and get cheaper prices. Xilinx uses your design to write a custom test program which doesn't test parts of the FPGA which you don't use. This lets you use chips which Xilinx would otherwise have to reject as defective, whether due to process variation or outright defects. It also means you can't change your design! (I think you get one mulligan before they want you to pay another fee to have a new test program written.)
It's useful mainly for certain niche industries like high speed telecom backplanes, where the product is too low volume to justify taping out an ASIC, but sufficiently high volume that losing the ability to update hardware in the field is worth it to shave some dollars off the unit cost. As for the suppositions made by the blogposter, it doesn't improve performance, it just reduces cost.
and because the extra effort to optimize code for hardware acceleration will amortize more favorably as CPU performance scaling increasingly relies upon difficult techniques such as massive parallelism. After all, today’s massively multicore CPU architectures are starting to look a lot like the coarse-grain FPGA architectures proposed in academic circles in the mid to late 90’s.
Er... no, no they aren't. Those coarse-grained FPGA architectures fell far short of having a full-featured CPU in each tile.
An equalization of FPGA to CPU performance should greatly facilitate the penetration of open hardware at a very deep level.
Good luck with that, "Bunnie".
I don't have the patience to dissect the rest of it... suffice it to say that he's clueless, and doesn't know it.
In my day, we just shouted zero's at the window.
The Kruger Dunning explains most post on
It's all just propaganda.
These posts express my own personal views, not those of my employer
I'm not convince of the central postulate here. We may not be able to continually increase CPU performance by slavishly increasing the transistor count any more than we could do it by slavishly increasing the clock frequency, but that doesn't mean the Intels of the world can't maintain dominance in performance. There are all sorts of possibilities that haven't been as yet fully explored because the industry has billions (trillions?) sunk into the silicon-specific end-to-end fabrication process, and they still see incremental improvements using current technology.
It may be true, as the author says, the farthest we can push silicon is about 5nm design rules. But what about cnt and graphene? What about Diamond Age style nanoprocessors? What about fully optical processors and interconnects? That's not the kind of stuff people are going to build in their garages.
We will enter an age where people will write efficient code again.
Probably not - when we design the chips we spec a lifetime that the chip will survive under worst case conditions before 'Electromigration*' kills it. Generally systems are currently well obsolete before it becomes an issue - but run your laptops under worst case conditions for ten years and the semicondutors will stop working. To fix the issue we'd have to make the wires bigger, which would stop Moore's law all by itself. EM isn't even to only lifetime issue for chips, it's just the most significant. *Electrons flowing through the copper interconnect gradually move the metal atoms out of places with higher current density towards places with lower density, causing the wire to fail or even a bridge to another wire. Most significant on power and ground (where current goes one way) but also happens on the signals. S