Slashdot Mirror
HP Answers The Question: Moore's Law Is Ending. Now What? (hpe.com)
Long-time Slashdot reader Paul Fernhout writes:
R. Stanley Williams, of Hewlett Packard Labs, wrote a report exploring the end of Moore's Law, saying it "could be the best thing that has happened in computing since the beginning of Moore's law. Confronting the end of an epoch should enable a new era of creativity by encouraging computer scientists to invent biologically inspired devices, circuits, and architectures implemented using recently emerging technologies." This idea is also looked at in a broader shorter article by Curt Hopkins also with HP Labs.
Williams argues that "The effort to scale silicon CMOS overwhelmingly dominated the intellectual and financial capital investments of industry, government, and academia, starving investigations across broad segments of computer science and locking in one dominant model for computers, the von Neumann architecture." And Hopkins points to three alternatives already being developed at Hewlett Packard Enterprise -- neuromorphic computing, photonic computing, and Memory-Driven Computing. "All three technologies have been successfully tested in prototype devices, but MDC is at center stage."
Williams argues that "The effort to scale silicon CMOS overwhelmingly dominated the intellectual and financial capital investments of industry, government, and academia, starving investigations across broad segments of computer science and locking in one dominant model for computers, the von Neumann architecture." And Hopkins points to three alternatives already being developed at Hewlett Packard Enterprise -- neuromorphic computing, photonic computing, and Memory-Driven Computing. "All three technologies have been successfully tested in prototype devices, but MDC is at center stage."
Law has a scientific and a colloquial usage. âoeMurphyâ(TM)s Lawâ is equally not a law, and yet weâ(TM)re not swayed by that argument to change its name. Mooreâ(TM)s Law was named by someone and itâ(TM)s in the common vernacular now. Nobody will have any idea what youâ(TM)re talking about if you refer to it as âoeMooreâ(TM)s Observationâ or âoeMooreâ(TM)s Conjectureâ. Itâ(TM)s not that youâ(TM)re being overly pedantic, youâ(TM)re just fighting against how language works.
Anyway, itâ(TM)s fine. The era of Moore will end soon enough, and maybe you can help name the next observed trend correctly.
Yes, I know "Moore's Law" isn't a law but an observation.
When I RTFA, it seems the author is looking at different technologies to continue growth of computing capability for a given unit of space. I also get the impression that Mr. Williams is looking to fund projects that he has an eye on by saying that Si based chips will soon no longer be economically improved and VC/Investment Money should be looking at alternative technologies rather than continued shrinking of Si chip features.
Unfortunately, I don't see a fundamental shift in what Mr. Williams is looking for the resulting devices to do. I would think that if he was really planning on dealing with the end of Moore's Law, he would be looking at new paradigms in how to perform the required tasks, not new ways of doing the same things we do now.
Regardless, the physical end of our ability to grow the number of devices on a slab of Si has been forecasted for more than forty years now - Don't forget that as the devices have gotten smaller in size, the overall wafer and chip size has grown as have yields which mean a continuing drop in cost per Si capacitor/transistor along with an increase in capability per chip. I would be hesitant to invest in technologies that depend on the end of Si chips' trend of becoming increasingly cheaper with increased capabilities over the next few years.
Mimetics Inc. Twitter
Hypothesis is a bit strong - it's more like an observation. An observation that have pushed the silicon industry forward by having that continuous improvement as a goal. But nobody thought it was some natural law.
We still have many ways to continue Moore's law. Going 3D with either monolitic (several logic layers per chip) or die stacking (or of course die stacking of multiple monolitic 3D dies). Increasing the size of chips is another way. Most chips aren't near the reticle limit and one can actually make chips larger than the reticle limit even though it gets harder.
Stopping Moore's law means that millennial programmers addiction to bloatware like react.js and electron will be curbed and we will be forced to write lean mean programs again. Remember when a gigabyte of ram was a high end workstation. Now it is a cheap netbook.
Sounds like HP is about to make an itanic breakthrough
Pain is merely failure leaving the body
It was simply an observation of the then-current progression of increases in transistor count in a commercially available part over a given period of time. While it has been used as a rule of thumb regarding the complexity of parts you should expect from your competitors in the same period of time, it has mostly at this point become a form of religion amongst Silicon Valley types, leading to it being a self-fulfilling prophecy, rather than something that was a necessity for the evolution that happened. Optimizing designs, such as has been happening recently with DRAM, dropping the required components for a single bit of memory from 5 transistors to 1 for instance could have provided similiar levels of performance (given the necessary process technologies to allow it) 20 years ago, meaning there could have been a period with no transistor count growth during which performance/density could have increased up to 5x. But instead they kept shoving more and more transistors into designs (not necessarily the same size silicon, since many of these transistor count hops happened as a result of few defects and thus larger dies offered by a mature process technology rather than the miniaturization allowed by a new smaller feature process...)
Now that the feature size is offering dwindling returns over the old processes (needing more error correction logic for the same size die, etc) we are seeing more of the chip design optimizations taking place, since throwing transistors at it is no longer working. Different logic designs are having to be used either to benefit from the otherwise negative interaction taking place at these small feature sizes (that 1 transistor(gate?) memory design mentioned above takes advantage of quantum scale physics that wouldn't necessarily have worked via the same design on older process technologies) means more work and different testing methodologies leading to uncertainty about the time to RTM for new components, potentially longer design times to figure out all the new interactions caused by the design, compared to older 'tried and true' modules, that only required more transistors to increase the performance or size of logic elements. Essentially chips are moving from a rural to suburban to urban layout with all the design methodology changes necessary to allow that. Urban in this case referring to the new 3d layered designs coming out, suburban to the current standard of 'simplifying/standardizing' infrastructure to allow compacting it together closer.
We start to megazord the chips, kinda like those HBM memory modules, but for several things like cores,GPUs or even lower level components like ALUs and pipelines etc..
Of course, we're talking about a cooling and interconnecting hell here, but probably will be more economically feasible than trying to make sub atomic transistors.
It's not good for anyone else. Fast, simple, cheap improvements that means my computer today is absurdly much faster than the C64 I had 30 years ago. And my dad can tell how much faster they are than vacuum tubes 60 years ago. A friend of mine has two classic cars, ~30 and ~50 years ago. Maybe they're not quite as reliable or safe as modern cars, but they go fast enough to mingle well with other cars. I think it'd be pretty sad if in 2047 base performance is pretty much the same and we just do it "smarter" through more cores and better algorithms.
Live today, because you never know what tomorrow brings
There's nothing good about this at all for a consumer. Maybe for a lazy has been corporation like HP who can wallow in their simplistic and outdated designs that barely need to change. This is a sign that a market has reached stagnation and has nowhere else to go. As a hardware and computer nerd, this is a dark age.
This might be possible. A Pentium-II on current die tech would be so small that pretty much transistor switching speed is your only concern. However you'd need the 82459AD-version L2 cache controller/Tag RAM chip to address 3.5-4GB RAM.
Still waiting on Serviscope_minor to wake up to fucking reality and realize that Jessica Price isn't going to fuck him.
We've all observed that as CPU and RAM have increased in speed/capacity, software has bloated horribly. Maybe now it's time to rethink crap like .net* and layers of virtualization, and go back to efficient code writing.
* (I've seen many major software projects more than quadruple in size, become buggier, and run much slower when they switched to .net)
Moore's Law has allowed too many to justify lazy design decisions and programming in computer architecture. With the end of Moore's Law, progress will come by offloading the main CPUs as much as possible, echoing construction of mainframes way back when. By coprocessing I/O, audio, and even most of the GUI, and OS kernel, the main CPUs can be reserved for actual user processes. Much of the housekeeping of a modern OS does not demand the degree of processing power of the main CPUs, but often hardware has been made cheaper by offloading things onto the main CPUs.
Moore's law might not directly hold true with multi-core x86's, but we now live in a world of differentiated processor power. ARM's specialized for hd streaming, or gaming, or AI, or Autonomous cars, or sensors for a wearable. You can buy an $80 tablet that will stream HD better then a nice 4 year old laptop. The reason is engineers are now focused on low cost processors for specific purposes. See Intel's purchase of Nervana for how Moore's law has forked.
Because higher frequencies mean higher temperatures. Remember when CPUs didn't need active cooling? So, manufacturers use the extra die area they gain from the shrunk transistors in other ways.
And two for being an Anonymous Coward Neckbeard.
I think even P2 could address 64 GB of RAM, but with a maximum of 4 GB per process. It is something a lot of people don't realize because Microsoft disabled that on consumer OS versions.
It's also hitting a wall for 2 reasons:
- As shrinks get closer & closer to atomic scales, they become more difficult, and therefore, more expensive. As a result, despite other trends like larger diameter wafers, process shrinks no longer result in cost savings, which is the only reason (other than capacity) that one would wanna do those in the first place
- Unlike past years, where applications would grow in complexity to quickly overwhelm CPUs at the time, multiprocessing has completely changed the game. Although programming using multithreading & multiprocessing techniques have been around for a while, there ain't too many applications that can overwhelm multiple cores. That is a good part of the reason that Intel & AMD have slowed down in their CPU sales: not too many people have to replace laptops that they've had for years. When that gravy train is drying up, there ain't much of a case to spend billions in process shrinks.
Don't tell them they might have to bin their handholding, inefficient bloated frameworks, or have to trade in their scripting or VM languages for something that compiles to machine code and where they might - horrors! - actually have to have a clue about how memory (de)allocation, threading, multi process, DB normalisation, sockets actually works. Or know how to pick the best sorting algorithm for the data size and complexity they're working with and not just hope the 21 year old hipster who wrote the dUdeFrAmWeRk sorting subsection while kite surfing stoned in Bali actually knew what the fuck he was doing. We don't want to scare them until its absolutely necessary.
What it really means is that "AI" won't happen. Autonomous car driving won't happen. Lots of things that AI and Space Nutters want to happen won't happen. We have been spoiled by Moore's Law and recent technological process has depended on it. If you look at the claims of AI nutters they all say "well computers are X times as fast now as they were in the last decade, image how fast computers will be in the next 10 years!!!". The answer is "not much faster". In fact, the computer you own now isn't significantly faster than the one you had 5 years ago.
Many CPU's still don't need active cooling. The fans in my laptop (XPS 13, i5-6200U) only come on under heavy load, and that only because of the cramped interior of the machine. The "M" cpus in things like the new Macbooks use even less power, and the Atoms less still; I had an old Atom netbook whose cooling fan died, but the machine still ran fine (including under load) without it. (I think its CPU had a TDP of 2.5W.)
Because higher frequencies mean higher temperatures.
That is nonsense. Modern CPUs need more power, hence they produce more heat.
Remember when CPUs didn't need active cooling?
Yes, my 6502, got about 110C hot. But the transistors still worked at that temperature. Modern CPUs have so small transistors that they can not work at that temperature anymore as the electrons would simply jump several transistors far through the substrate.
Cost free eBook I read (by iBook/Kobo/Amazon/ObookO/Gutenberg etc.): "The Green Odyssey" by Philip Jose Farmer.
Ever question why there are constantly so many versions, updates, patches and problems? Because the hardware keeps getting updated, which gives us new potential features and introduce new problems.
Old phones keep slowing down, and losing battery life because of this.
With consistent hardware, we can:
a) Take more time and develop software WITHOUT bugs. Yes, I know this sounds ridiculous, because no one can take years to develop software when their competitors do it weeks. No longer a problem once the hardware becomes consistent.
b) Build electronics to last, rather than with planned obsolescence.
excitingthingstodo.blogspot.com
https://msdn.microsoft.com/en-...
Registry settings that can be modified to improve operating system performance
This section provides a description of recommended values for several registry entries that impact operating system performance. These registry entries can be applied manually or can be applied via the operating system optimization PowerShell script included in Windows PowerShell Scripts.
Increase available worker threads
At system startup, Windows creates several server threads that operate as part of the System process. These are called system worker threads. They exist with the sole purpose of performing work on the behalf of other threads generated by the kernel, system device drivers, the system executive and other components. When one of these components puts a work item in a queue, a thread is assigned to process it.
The number of system worker threads should ideally be high enough to accept work tasks as soon as they become assigned. The trade off, of course, is that worker threads sitting idle consume system resources unnecessarily. Modify and/or create the following REG_DWORD values in the registry and then set to the recommended values listed below.
The AdditionalDelayedWorkerThreads value increases the number of delayed worker threads created for the specified work queue. Delayed worker threads process work items that are not considered time-critical and can have their memory stack paged out while waiting for work items. An insufficient number of threads will reduce the rate at which work items are serviced; a value that is too high will consume system resources unnecessarily.
AdditionalDelayedWorkerThreads
Key: HKLM\SYSTEM\CurrentControlSet\Control\SessionManager\Executive
Value: AdditionalDelayedWorkerThreads
Data Type: REG_DWORD
Range: 0x0 (default) to 0x10 (16)
Recommended value: 0x10 (16)
Value exists by default? Yes
The AdditionalCriticalWorkerThreads value increases the number of critical worker threads created for a specified work queue. Critical worker threads process time-critical work items and have their stack present in physical memory at all times. An insufficient number of threads will reduce the rate at which time-critical work items are serviced; a value that is too high will consume system resources unnecessarily.
AdditionalCriticalWorkerThreads
Key: HKLM\SYSTEM\CurrentControlSet\Control\SessionManager\Executive
Value: AdditionalCriticalWorkerThreads
Data Type: REG_DWORD
Range: 0x0 (default) to 0x10 (16)
Recommended value: 0x10 (16)
Value exists by default? Yes
For starters, I've been reading "the sky is about to fall" articles for at least 20 years about how: "In 2-3 more years Moore's Law is going to slam into a barrier imposed by the laws of physics.". The entire world of computing will come crashing down and burn, the beast will rise from the pit, the keymaster and gatekeeper will find each other, the dead will dig themselves up from their graves, dogs and cats living together, mass hysteria. The doom and gloom crowd have been wrong every time so far. Every time, some clever person at IBM or Intel has figured a way to cross the streams and save the world. So why should I trust the chicken littles that we're living in the end times now?
And even if Moore's Law slows down or pauses, there's plenty of room in the hardware we have today for continued improvement on the software side. Developers will just have to rely less on "pay no attention to the man behind the curtain" languages and frameworks and go back to optimizing their code for performance... like they used to before crap like Java, .Net, and Rails encouraged everyone to be lazy and rely on ever-improving CPUs to make their apps not suck. Why should the hardware guys do all the work, after all? Hell, they can start by writing their code to be properly multi-threaded. My desktop, for example, has a core i7 with 4 physical cores and 8 virtual ones via hyper-threading. I couldn't begin to count the number of times I've watched some program or another run a single core up to 100%, stop there, and ignore the 7 other threads it could be running simultaneously to improve its performance nearly 8-fold, no new or faster hardware needed.
Imagine all the people...
Actually, law doesn't have a scientific meaning. It once did, but that was back when they thought of the universe as being run by the laws of God.
Related scientific terms are things like theory, hypothesis, conjecture, guesstimate, etc. None of them imply perfection. A theory is a hypothesis that has passed several tests. A conjecture is an incompletely formulated hypothesis. And a guesstimate is a conjecture that doesn't have any calculations behind it, but may fit the known data.
When you look at how scientific terms are defined (i.e., there are essentially rough estimates of the probability of correctness, with certainty of either accuracy of falsity being rejected) it's clear whey they don't easily fit into the colloquial.
I think we've pushed this "anyone can grow up to be president" thing too far.
OT: your browser needs new batteries so your comments down look like this - "âoeMooreâ(TM)s Observationâ or âoeMooreâ(TM)s Conjectureâ. Itâ(TM)s not that youâ(TM)re"
damaged by dogma
A slight disagreement, R. Stanley Williams is interested in other solutions as he specific refers to options other than von Neumann architecture computing. Considering he is from HP one might surmise he is looking to DMC as well as their vague (to me) The Machine concept. I have yet to read the other article that is concerned with The Machine.
The issue he offers up for consideration is that further spending of even more $B to move Moore one step to 5nm or beyond would be better spent on looking to other directions for computing itself. He feels clearly that research into other directions have been starved because of the relentless metronome that was Moore's Observation (aka Law). Making a statement of intent is different from his observation that the huge investment in Moore has stunted the very research you are suggesting.
Even your subject was incorrect (ftfy). He is looking at a semiconductor industry that has been investing everything (or nearly so) into extending Moore when the very things that are needed if we are indeed at the end of it are being starved of funding.
As the transistor's get closer to the minimum allowed and with them unable to push past 4.3Ghz (unless extreme cooling is used) the CPU has reached stagnation and it's just clever marketing with little features built into the CPU that differentiates each process(along with the stupid naming conventions now used which mean nothing to many people means we should be looking elsewhere.
We need to dump silicon, or find a hybrid way of using current nm production process with a new material to increase computing power else we shall just end up hugh multi-core CPU like 64 or 128core.
Update, Williams piece does expand/explain The Machine idea. I had thought it was in the other article but that was mostly fluff (no insult intended) and linked to the Williams piece.
Slogging through The Machine. Better than nothing but I tire of not having actual gear to use, same for 3D-Xpoint for that matter.
Um... the recent server offerings are pretty impressive. You should check it out.
Another consultant who stuck it out.
"We are the Priests, of the Temples of Syrinx..."
So if Moore's law is ending, we have Betteridge's law to let us know it is not. Thanks for putting that question mark on the end of the headline! We are saved.
PlanetVulkan.com
My Kaby Lake thermal throttles at like 101C iirc. Not too far off your 110C.
On the Oregon Cost born and raised, On the beach is where I spent most of my days
We still have many ways to continue Moore's law.
We have many theoretical ways of going forward, problem is, there hasn't turned up an actual way. Computing speeds have stalled the past several years. Too few data-points to be certain that Moore's law is dead, but it does not look good.
Fast Federal Court and I.T.C. updates
It's also hitting a wall for 2 reasons:
The purpose of the article wasn't to argue about the reasons it's hitting a wall, but that the fact that it is hitting a wall is a good thing for the industry. It's an optimistic view of the industry, but not terribly convincing. If the industry can't deliver results quickly, investment will dry up and Moore's law will turn from exponential, to linear, to flat.
Fast Federal Court and I.T.C. updates
I always wondered about that (until 64-bit machines became common and made it moot.) But were there any 32-bit processors that brought out more than 32-bits worth of address lines to physical RAM?
" If the industry can't deliver results quickly, investment will dry up and Moore's law will turn from exponential, to linear, to flat."
Turn Moore's Law around a bit and bare with me:
I've begun to wonder if Moore's Law is better thought of primarily economically rather than scientifically... addressing issues with product pipeline management, upgrade demand/appetite etc. Given that Moore was managing a technology that grew in jumps and spurts as discoveries were made and problems overcome; he might want to level that road out for the near future... (which yes, ended up being 50 years)!
By publicly stating "Moore's law" he was setting a pace for the industry that allowed some stability for the entire industry ( and maybe kept them from competing each other out of business).
Moore's Law wasn't predicting the technological advancement per say, but instead stabilizing to better manage it.. and in that way it also enabling the advancements to continue at pace. So technologically it turned out it was a self-fulfilling prophecy.
So with "Moore's Law" coming to an end the industry needs to find a new stable state... and TFA points out some possibilities... and as you say... whether stability is found somewhere off of the pure performance curve is another question.
Just my 2 cents at 2am... Thoughts?
"The main reason for die shrinks is reduced power usage and/or better performance." - NO it is for yield. Halve your feature size and you get 4 x as many die on a wafer.
Time for bed, said Zebedee - boing
Lay off half 50% of the employees every 18 months.
Law has a scientific and a colloquial usage. Murphy&'s Law is equally not a law...
With usual usage, it's not even a law. Typically, a law is something that can be expressed as mathematical form, e.g. the law of gravitation, the laws of thermodynamics. Murphy's Law should actually be Murphy's Theory.
is Apple doomed, or will it be the Year of the Linux Desktop?
I always wondered about that (until 64-bit machines became common and made it moot.) But were there any 32-bit processors that brought out more than 32-bits worth of address lines to physical RAM?
Huh? Many 32 bit processors and especially those which supported some form of physical address extension supported more then 4 gigabytes of address space. That includes all Intel processors starting with the Pentium Pro which supported 64 gigabytes of physical address space. After that, practicality and market segmentation are usually what limited the maximum amount of memory. Microsoft held back greater than 4 gigabyte systems in the consumer space for years.
-
More on topic; I did a quick search to see if anyone shared my view from an economic perspective and thought this was interesting: https://techpinions.com/moores-law-begins-and-ends-with-economics/46575
Not exactly what I was stating but I found it interesting that he also made the observation that "It was an estimation that became a self-fulfilling prophecy".
But even that power reduction has limits. Like 0.7V is the voltage that a diode must have before current flows, so internally, there is going to be clamps on Vdd: one can't just keep reducing it w/ process shrinks the way one did when going from 5V to 3.3V to 1.8V. So while there may have been in-circuit level shifters in the past, one can't keep shrinking and hoping to go from 12 hr battery life to 24.
CPUs generally will be made on the cutting edge processes, and so will higher density semiconductor memory - both RAM and Flash. What I was arguing was that we are getting to the point - if we ain't there already - where it makes less sense to shrink a CPU. Skylake or Kaby lake - I've lost track of which one's most recent - may be the last process shrink for the core series.
True. But yield also gets complicated when factoring in other parameters that may be affected as a result of a shrink. Yeah, you'll get the square of as many die/wafer, but that doesn't necessarily imply that you won't have more defects per wafer.