Another issue: currently the biggest bottleneck , cost wise, is in lithography - the process we use to draw the transistors into chips. Because of this issue, the two latest generations of chips are more expensive(per transistor) than an older version - stopping moore's law.
And moore's law probably won't return to life, until we learn how to solve that problem, which the current work doesn't help with.
This is where 450mm wafers and EUV (extreme ultraviolet lithography) were supposed to come in. EUV relieves the need for double patterning and the tremendous additional costs that entails (and was used to manufacture this 7nm chip).
The CEO of Applied Materials, Gary Dickerson, has stated that the 450mm wafer timeline “has definitely been pushed out from a timing standpoint.” That’s incredibly important, because the economics of 450mm wafers were tied directly to the economics of another struggling technology — EUV. EUV is the follow-up to 193nm lithography that’s used for etching wafers, but it’s a technology that’s spent over a decade mired in technological problems and major ramp-up concerns.
And for comparison, scale fans, let's remember that we're talking about making 7nm features on wafers that are nearly a foot and a half wide, using near-as-dammit x-ray wavelengths.
A few teething problems would be expected.
7nm has been struggling for a while, 5nm is likely to be late, and I don't think anyone really knows what happens after that.
Longer term, industrial manufacture is probably going to have to move to something exotic like nano-assembly of individual atoms, with some extra finagling to work around tunnelling effects. (Easier said than done...)
... and why would we invest the money to do that when there is not enough (software-driven) demand for that performance.
The average person uses PCs and mobile devices to browse the web, write documents, order an Uber, and maybe play games. Nothing much is being done on the software front that challenges current systems. Maybe if VR took off or we got home applications for AI like domestic robotics that would change. I could see a domestic robot capable of folding clothes, cleaning up, etc. needing a low power chip that can do what a dual-12-core Xeon can do on smart phone power and thermal profiles. <5nm might be needed to accomplish that.
I'm not sure server and high-end compute demand is sufficient to pay for the R&D that would be required to far beyond 7nm.
But the good news is that we haven't even scratched the surface of what current systems could theoretically accomplish. Look into the demo scene and prepare to be blown away by what 8-bit CPUs in the 1980s could accomplish with non-crap code running on them. Maybe we need a software Moore's Law to take over for the hardware one -- right now software has more of an erooM's Law.
One thing is clear: if you do software, ball's in your court either way. Either you need to invent killer apps to keep demand high for high performance computers -- things that really need that much power -- or you need to take over for the hardware people and start finding new efficiencies.
Ball's really been in software's court for a while anyway with multi-core... linear performance max'd out (for consumer chips) a while ago.
I agree that most of the demand requires software innovation , but there are other good sources of demand, for example:
1. AI - variety of applications, both for consumer and business markets.
2. Telepresence. If we can get the real feeling of "being there" to telepresence, at a price point that's attractive for the consumer.
3. Simulation. Currently it's a complex process ,mostly done by experts. If it can be a tool for regular engineers , and maybe further down the road - for combining that with some sort of genetic-algorithms , maybe there's potential for a huge demand increase.
Electron-beam lithography is a technique that works, but because it's very slow, it's also very expensive. There are efforts to parallelize e-beam writing, but they face hard challenges. For example, if you want to pattern faster, you just need to shoot more electrons. But if you shoot too many electrons, they repel each other and the pattern blurs. It's a difficult problem to overcome, but people are working on it.
I think we've been wasting far too much processing power in inefficient software for the past few decades, and it's only Moore's Law that let it happen for so long. Now that it's coming to an end, maybe we'll see more emphasis on efficiently optimised software and mindful resource usage.
I think you're incorrect and remembering performance that never existed based on shortcomings you glossed over at the time because you have that all too human bias of believing that things were better when you were younger.
There is so many low hanging fruit in software design that is simply there because of legacy design trade-offs and the cost associated with replacing those - we could gain huge performance gains over night if we for eg. eliminated reliance on hardware memory protection and context switching from kernel space to user space by using languages that can prove memory safety in software.
Then imagine how much performance you could get out of OS level VMs that understand the processes at VM level (ie. can access code in some IR that they can analyze easily, recompile it on the fly, etc.) there is already stuff like this in specialized markets (eg. kernel level GC for JVM) but it's still fairly specific.
Then there's all the shitty legacy abstraction layers in things like filesystems - ZFS is a perfect example of what kind of gains you can get for free if you just rethink the design decisions behind current stack and see what applies and what doesn't.
If the benefit of rewriting these systems ever overcomes the cost - we have huge potential areas for performance gains, modern systems are very far from being performance efficient, they are efficient based on various other factors (development cost, compatibility, etc.)
I wish Linux would just merge a ZFS implementation into the kernel already.
I also wish ZFS would grow an encryption layer (one that isn't based on Sunacle's implementation, since Sunacle doesn't want to share that one thus no one can use it).
I understand what you're saying, but do you really think most of a modern Android (to take the theme to its Javaesque extreme) stack is the most efficient way to accomplish processing?
Compare that to some of the code people ran through 6502-derivatives.
Abstraction may be more efficient in terms of programmer time, and performance efficiency may be high enough so as to be immaterial, but the two shouldn't be conflated.
> do you really think most of a modern Android (to take the theme to its Javaesque extreme) stack is the most efficient way to accomplish processing?
Reminds me of a version of this image[1] which has a discussion superimposed over it that says, "but if he had a big enough pile of ladders he could get over the wall!" and someone responds, "welcome to Android optimization." I think we see something similar with Javascript performance.
The older I get, the more I start seeing over-complexity in stacks as a security risk as well. I feel like there's a fundamental maximum to the number of levels of abstraction one can keep in one's head "enough" to avoid creating layer interaction bugs. Stack overflow, indeed. :)
"Coming to an end"? The sky isn't falling yet. Just because they're having some trouble with one process doesn't mean the whole party is over.
There are other materials to make chips out of besides silicon, gallium arsenide and carbon for instance, each of which has different scaling properties.
There's also ways to make chips more dense by stacking wafers instead of trying to shrink features.
It can sometimes be comforting to know that the universe imposes fundamental limits on how how efficiently computation can be done. Comforting because IIRC we've still got at least 15 orders of magnitude improvement available. But while I'm sure you're right that we're going to be able to switch to different materials (or maybe away from transistors entirely) when progress in silicon runs out we might have to expect an interregnum while other computational substrates are developed to the point where they can provide higher performance.
Stacking is certainly a thing and it's good for memory (see AMD's newest graphics card) but power dissipation provides limits in terms of how much high speed logic you can put under a given area.
And moore's law probably won't return to life, until we learn how to solve that problem, which the current work doesn't help with.