Submitted via IRC for AndyTheAbsurd
Hammered by the finance of physics and the weaponisation of optimisation, Moore's Law has hit the wall, bounced off - and reversed direction. We're driving backwards now: all things IT will become slower, harder and more expensive.
That doesn't mean there won't some rare wins - GPUs and other dedicated hardware have a bit more life left in them. But for the mainstay of IT, general purpose computing, last month may be as good as it ever gets.
Going forward, the game changes from "cheaper and faster" to "sleeker and wiser". Software optimisations - despite their Spectre-like risks - will take the lead over the next decades, as Moore's Law fades into a dimly remembered age when the cornucopia of process engineering gave us everything we ever wanted.
From here on in, we're going to have to work for it.
It's well past the time that we move from improving performance by increasing clock speeds and transistor counts; it's been time to move on to increasing performance wherever possible by writing better parallel processing code.
Source: https://www.theregister.co.uk/2018/01/24/death_notice_for_moores_law/
(Score: 4, Insightful) by takyon on Sunday January 28 2018, @12:14PM (4 children)
Moore's Law: Not Dead? Intel Says its 10nm Chips Will Beat Samsung's [soylentnews.org]
Transistor counts are still increasing. Intel's 10nm has been delayed and will probably be delayed again by the "in-silicon" fixes they have alluded to for Meltdown (Spectre?). And although their transistors per mm2 makes them look better than the competition, it's clear that rather than "10nm" Intel competing against "10nm" Samsung/TSMC/GlobalFoundries, it will be against "7nm" Samsung/TSMC/GlobalFoundries. Intel has slipped enough that the competitors can probably reach similar transistors per mm2.
Where will those transistors go? More cores and graphics (and less to security risk enabling optimizations). The Ryzen/Threadripper, Intel Core i9, and "mainstream" Intel 6-core launches greatly boosted the amount of cores/threads you can cheaply acquire (Intel was previously selling the 10-core i7-6950X for $1700). From 6 cores to 18 cores, it's all much cheaper. Xeon and Epyc are pushing to 32 cores and beyond.
2017 saw performance/$ massively increase, IF you can take advantage of parallelism that is. And the performance "Meltdown" only really affected Intel AFAIK. AMD users potentially got a massive increase in multi-threading capability as well as a ~50% increase in IPC. So their single-threaded performance didn't decline at all.
New nodes are still on the roadmap. [soylentnews.org] "7nm" is assured. "5nm" is likely. "3-4nm" is possible. EUV might not be needed although it would help.
When CMOS scaling does become certifiably dead, it will eventually be replaced by something else. Even if we have to endure a few years in which no significant improvements are made at all (meaning even less than the 3-10% IPC improvements certain Intel generations have made, no increase in core count at the same size, no increase in transistors per mm2, etc). And what we need to see is a new processing element with a huge reduction in power consumption and heat so that it can be stacked. A growing amount of software will be able to take advantage of thousands or millions of cores.
[SIG] 10/28/2017: Soylent Upgrade v14 [soylentnews.org]
(Score: 0) by Anonymous Coward on Sunday January 28 2018, @02:11PM (1 child)
You mean... like the Mill CPU?
https://www.youtube.com/watch?v=LgLNyMAi-0I [youtube.com]
(Score: 2) by maxwell demon on Monday January 29 2018, @05:33AM
For those who prefer not to watch an hour-long video to find out what this is about, here's the Wikipedia link. [wikipedia.org]
The Tao of math: The numbers you can count are not the real numbers.
(Score: 2) by TheRaven on Monday January 29 2018, @11:29AM (1 child)
This is only kind-of true. The thing that's changed in the last 3-4 years is that new process technologies are not bringing down the cost of transistors. The other part of Moore's law is that the dollar investment is constant: the number of transistors on a chip that you can buy for a fixed amount doubles every 18 months. With the last couple of process advances, the cost per transistor has gone up.
It used to be that for large volume runs, you always wanted the newest generation. It might have higher up-front costs and a higher cost more per wafer, but the cost per chip would go down because the amortised cost per transistor was lower. That's no longer the case: even for enormous runs where the fixed costs are negligible when amortised across the entire run, the cost per wafer has gone up by enough that the cost per chip has also gone up. Oh, and those fixed costs have gone up by a lot, so for smaller runs this is even worse.
sudo mod me up
(Score: 2) by takyon on Monday January 29 2018, @12:11PM
Intel disputes that cost per transistor is not going down:
Intel: Our 10-nanometer chips will cost 30% less than the competition’s [venturebeat.com]
Intel’s 10 nm Technology: Delivering the Highest Logic Transistor Density in the Industry Through the Use of Hyper Scaling [intel.com]
Everybody is creating their latest generations of chips without using EUV. Once EUV is working at the right scale, that will drive cost down some more. It could even be useful in the scenario where we completely stall out in terms of improvements with no more shrinking, because it could still lower the cost. As for 450mm wafers [eetimes.com]:
[SIG] 10/28/2017: Soylent Upgrade v14 [soylentnews.org]