[Discuss] Article: The Death of Moore's Law Will Spur Innovation
ilebedev at mit.edu
Fri Apr 3 20:02:58 UTC 2015
The way things stand now, a deliberate effort costs so much money, you'd be
hard pressed to have anyone who actually could influence such a decision to
take you seriously. It isn't an isolated problem: ASIC EDA tools are
proprietary expensive and bad, fabs are secretive, expensive, and finicky,
the big customers of these fabs rely heavily on trade secrets and legacy.
Every player in this industry uses money as weaponry to keep their niche.
I've heard a few opinions along the lines of the following:
As we approach ~10nm feature resolution in silicon, there's not a whole lot
of room for improvement (it's ~20 silicon atoms across for crying outloud).
Reliability drops, yield (manufacturing reliability) plummets, static power
becomes ever more an issue, etc. New technologies that will fundamentally
change the game are unlikely to appear. Let's look at Intel. They innovate
in their current technology, and then make money shrinking the design to
newer process nodes. Nahalem was built in 2008 in 45nm, then shrank to
32nm, and so on. Here's a pretty picture detailing just that:
^ Anyway, point is, if Intel gets to 10nm, their incentives for building a
new architecture are slim: doing so is expensive, and their chances of
successfully shrinking it to 7nm are not great, meaning they may not be
able to make money the usual way.
Large systems companies like google/facebook/microsoft surely would love to
pursue vendor independence, and would support open silicon if it is
If we do hit the limits of silicon, and new technologies do not emerge,
foundries may have to compete on cost and reliability instead of feature
size. This will drive prices down, incentivize additional fabrication
companies (The US is now down to a precious few Fabs, sustained in no small
part by government contracts. IBM is selling theirs to Global). One can
dream that if we remain at ~10nm for a long time, startups, students, and
various brave crazy people will create ever more affordable fabrication
techniques. Fewer trade secrets will obfuscate the ASIC supply chain. CAD
tools will improve. Maybe one day we will see a "desktop fab", or FPGAs
cheap and fast enough to be competitive. That'd be a neat.
On Fri, Apr 3, 2015 at 2:39 PM Matt Maier <blueback09 at gmail.com> wrote:
> Do you think that the incentives will naturally result in open silicon
> eventually, or do you think it will require a deliberate effort?
> The typical benefits of open sourcing a thing seem to be social capital,
> faster development, and lower individual costs. It also seems like the more
> capital investment in the industry the slower it is to discuss and/or adopt
> open source. So maybe it's just that producing actual chips requires so
> much up-front capital investment that nobody is willing to entertain the
> idea of disrupting their current process. But, you know, maybe they'll
> gradually come around even if it's on the order of decades.
> Is there any potential for pulling a RepRap? Additive manufacturing
> machines all cost hundreds of thousands or millions of dollars but Dr.
> Bowyer got a useful amount of performance out of something that only cost
> hundreds of dollars. Could we get some useful chips out of a process that
> costs less than a thousand dollars?
> On Fri, Apr 3, 2015 at 9:12 AM, Ilia Lebedev <ilebedev at mit.edu> wrote:
>> This is more or less my field of research, and people have been saying
>> this in every architecture talk since like 2006. There is not a whole lot
>> of innovation happening in computing machinery because computer
>> architecture is currently a fairly dry, conservative field accepting
>> largely incremental work.
>> Open source hardware is great! I hope it will become more mainstream and
>> will receive additional university and industrial support. As things stand
>> now, there are a few low-level ISA projects (RISC V, OpenRisc) attempting
>> to encourage open source processor design, a LOT of open source FPGA things
>> of dubious usability and credibility, and a rich body of excellent open
>> source systems like Beagle that use non-open components. We will likely see
>> more and better open systems built using purchased components. These will
>> be more supported by industry, have better dev tools, better-optimized
>> kernels (Raspberry Pi is a poster child for this, although their Raspbian
>> kernel has a long way to go).
>> Making the entire stack (silicon all the way up to software) open is not
>> quite realistic considering the absolutely staggering cost of producing a
>> custom silicon chip (and the high per-unit cost of FPGAs. Yes, they're not
>> *that* expensive, but they are ~2 orders of magnitude more expensive than
>> ASIC per unit, as per Rose paper). Furthermore, foundries tasked with the
>> manufacture of ASICs are extremely secretive with their process: AMD no
>> doubt has offices full of NDAs and a rigid system in place for the secrecy
>> of foundry-specific expertise. Academic projects have to sign a hefty NDA
>> and are prohibited from publishing low-level details of the foundry process
>> or performance. This is not an atmosphere where open source can thrive.
>> Wow, this turned into a rant, sorry.
>> My personal hope is that industry (Google, probably) will invest
>> significantly into building processors with an open ISA (Risc V, probably)
>> and publish their full spec. They will rely on cryptography instead of
>> industrial secrets (*cough* undocumented register wizardry in x86) to keep
>> their platforms secure. Companies and organizations will build their
>> Raspberry Pi and Beagle Bone clones with these semi-open components. The
>> platforms will run Linux. Everyone will rejoice. The USB, HDMI, Bluetooth,
>> and other integrated components will remain sketchy, buggy, and closed. The
>> processor SoCs too will remain proprietary, though well-documented, and
>> built upon well-audited and open processor design.
>> On Fri, Apr 3, 2015 at 11:47 AM Toni Klopfenstein <
>> toni.klopfenstein at gmail.com> wrote:
>>> Just FYI for anyone interested. Interesting article on open-source
>>> electronics hardware from Bunnie Huang.
>>> discuss mailing list
>>> discuss at lists.oshwa.org
>> discuss mailing list
>> discuss at lists.oshwa.org
> discuss mailing list
> discuss at lists.oshwa.org
-------------- next part --------------
An HTML attachment was scrubbed...
More information about the discuss