Hacker News Board Hacker News, browsed like a classic message board
Live data from Hacker News

Moore’s Law is dead – Long live the chiplet

semiwiki.com

Post reply on HN
Print view View original

111–120 of 123 posts

Re: Moore’s Law is dead – Long live the chiplet

#111

Post by galangalalgol »

pjmlp wrote: post #110
galangalalgol wrote: post #96

Earlier quoted context omitted.

Pytorch isn't just for ML, it can do normal signal processing or physics too. The julia libraries for cuda and roc and oneapi also are general enough for those uses and approachable. Both can fall back to cpu without much modification to the rest of your code. If you aren't doing signal processing, physics or something that would benefit from simd, then the gp is correct, a gpu won't do much for you.

Pytorch belongs to "... libraries written by the GPU druids..." on my comment. And still requires specific skills to use, and is constrained to Python, C++ and Java based languages. GPUs need to be exposed like SIMD, something that the language runtime takes care of, even if not perfect, better than not using them at all.

IME simd very rarely gets used by the compiler or runtime unless you make some slight changes in your data structures or flow, that require specific knowledge of the simd hardware. Asking a compiler to target unknown GPU architecture seems more likely to slow execution than speed it up. Even when writing my own cuda kernels I sometimes realize that something I am doing won't work well for a particular card and it is actually making me slower than the cpu. I'm sure we'll get there, but cards will have to converge a bit.

Re: Moore’s Law is dead – Long live the chiplet

#112

Post by pedrocr »

Andrew_nenakhov wrote: post #26
dragonelite wrote: post #25

Earlier quoted context omitted.

Moore's Law is transisters per square mm or something like that. Just drop a second layer on top with 3d fabbing and im sure moore's law will go on for a decade. inbe4 we have 2,4,8,16,32 etc layer architectures in the future.

You'll need to find ways to remove the heat form inner layers, or the whole pie will melt.

If all you wanted was to deliver on the density you could just keep most transistors dark and still stay within the power budget. That's not very useful in general but maybe there are some aplications for it. AMD has been shipping their 3D VCache setup where they layer an extra cache only die on top of a CPU die. That's been benchmarking really well and is already an effective doubling that can be cooled with a normal PC setup. Maybe there's a few other tricks like that to get a few more vertical layers out of the same processes.

Re: Moore’s Law is dead – Long live the chiplet

#113

Post by ilaksh »

The history of computing is moving from one paradigm to another. We are well past the fast speedups in single thread transistor-based performance phase and into the hyper-parallelization phase.

3d stacking is another innovation that can help.

But I think within a decade or two there will be a move away from silicon-only transistors to something like memristors or some type of optical or optoelectronic system that hasn't even been invented yet. This will provide some iterations with again radical parallel interconnect and quite possibly single thread speedups.

Re: Moore’s Law is dead – Long live the chiplet

#114

Post by pedrocr »

urthor wrote: post #63
BirAdam wrote: post #53

Unless I’m missing something, the article did mention Moore’s law proper with transistor density doubling every 18 months, but then meandered to talk about other things. M1 has 16 billion transistors thanks to TSMC. Each new node has delivered on Moore’s law with AMD and Apple. I don’t doubt that Moore’s law will stop. I can even say that it Moore’s law may have failed from time to time, but the spirit of the law liv…

Moore's law is: Transistor density doubling every 18 months for a similar cost. https://t7m8e9c8.rocketcdn.me/wp-content/uploads/2020/09/pre... https://en.wikipedia.org/wiki/Moore%27s_law#cite_note-Moore_... By those terms, Moore's law is totally extinct. Folk haven't noticed however, because the "leading edge" logic manufacturers have 60% gross margins. The vast majority of their costs are in design, distribution an…

Most people believe Moore's law has already stopped on the density side. Jim Keller claims we have 50 more years of it at least. If it failed on the economic side only instead then the implications are very different. I do wonder if on the economic side of the law we are missing a bunch of innovation because every node step also reduced the number of players in the market. Right now on the bleeding edge there are only 3 players and that keeps trending down. They're also all ASML clients so we're pretty close to having a single monoculture in semiconductor fabrication. Maybe that's an inevitable result of the problem space but I do wonder if at least some of our lost efficiency comes from there.

Re: Moore’s Law is dead – Long live the chiplet

#115

Post by urthor »

pedrocr wrote: post #114
urthor wrote: post #63

Earlier quoted context omitted.

Moore's law is: Transistor density doubling every 18 months for a similar cost. https://t7m8e9c8.rocketcdn.me/wp-content/uploads/2020/09/pre... https://en.wikipedia.org/wiki/Moore%27s_law#cite_note-Moore_... By those terms, Moore's law is totally extinct. Folk haven't noticed however, because the "leading edge" logic manufacturers have 60% gross margins. The vast majority of their costs are in design, distribution an…

Most people believe Moore's law has already stopped on the density side. Jim Keller claims we have 50 more years of it at least. If it failed on the economic side only instead then the implications are very different. I do wonder if on the economic side of the law we are missing a bunch of innovation because every node step also reduced the number of players in the market. Right now on the bleeding edge there are onl…

Costs are rising by ~50% every cycle give or take.

Exponential growth is exponential growth.

They'll exceed the GDP of the United States in a few cycles.

The question becomes, when costs hit the ceiling that Apple is willing to pay, will TSMC still be able to front up with the good stuff and deliver the growth.

Re: Moore’s Law is dead – Long live the chiplet

#116

Post by urthor »

gilbetron wrote: post #80
urthor wrote: post #60

Earlier quoted context omitted.

> single-thread performance and clock frequency have plateaued 10 years ago https://www.cpubenchmark.net/singleThread.html It's amazing how often this is parroted. Anyone with a passing familiarity with the numbers knows this is actually not true at all. Better caching, branch prediction, plus vast amounts of SRAM. There's been a slow & steady increase in the vast variety of single threaded workloads grouped together…

Sorry, doubling in 10 years vs doubling in 18 months effectively is plateauing! Especially since it isn't really a consistent 10% growth per year, but a decelerating growth over that decade. Furthermore, much of the purported single thread performance is taken from a small set of benchmark tests, and so chip makers just optimize them for those tests. Generic single thread performance has undoubtedly not doubled in th…

In any field of any kind except probably silicon, 100% growth in a decade would be marvelous. I don't think anyone could call it a plateau.

> Furthermore, much of the purported single thread performance is taken from a small set of benchmark tests, and so chip makers just optimize them for those tests

"Single thread" is a notoriously difficult benchmark to quantify. Instruction queue depth, floating vs integer, branching vs linear, there are so many variables.

Passmark is fine. Workload simulation is state of the art.

Re: Moore’s Law is dead – Long live the chiplet

#117

Post by pedrocr »

urthor wrote: post #115
pedrocr wrote: post #114

Earlier quoted context omitted.

Most people believe Moore's law has already stopped on the density side. Jim Keller claims we have 50 more years of it at least. If it failed on the economic side only instead then the implications are very different. I do wonder if on the economic side of the law we are missing a bunch of innovation because every node step also reduced the number of players in the market. Right now on the bleeding edge there are onl…

Costs are rising by ~50% every cycle give or take. Exponential growth is exponential growth. They'll exceed the GDP of the United States in a few cycles. The question becomes, when costs hit the ceiling that Apple is willing to pay, will TSMC still be able to front up with the good stuff and deliver the growth.

Obviously exceeding total GDP is not possible and we'll stop well before from just commercial considerarions as you describe. The question is how much of that is inherent to the problem and how much is lack of innovation. And we'll get to find out because if the processes stall from the economic side all the research turns to that whereas today it's an arms race for the next node and the increasing cost is taken as a given because as long as you get there the customers also arrive. Once that's no longer the case it frees up competitors from just chasing the node at equivalent costs to actually innovating on the total cost. If that leads to nothing then maybe we actually stop almost completely.

Re: Moore’s Law is dead – Long live the chiplet

#119

Post by Ma8ee »

lpedrosa wrote: post #40
Ma8ee wrote: post #36

Earlier quoted context omitted.

But the inefficiencies you sometimes see today can’t even be explained with any bad choice of language. You can create more than fast enough programs with interpreted languages with garbage collection. But then of course you need to know at least a little bit about data structures and not doing dozens of REST calls anytime anyone taps the screen.

What if doing dozens of REST calls is effectively the product?

Could you explain? I’ve got a hard time imagining a user interested in the rest calls themselves.

Re: Moore’s Law is dead – Long live the chiplet

#120

Post by urthor »

pedrocr wrote: post #117
urthor wrote: post #115

Earlier quoted context omitted.

Costs are rising by ~50% every cycle give or take. Exponential growth is exponential growth. They'll exceed the GDP of the United States in a few cycles. The question becomes, when costs hit the ceiling that Apple is willing to pay, will TSMC still be able to front up with the good stuff and deliver the growth.

Obviously exceeding total GDP is not possible and we'll stop well before from just commercial considerarions as you describe. The question is how much of that is inherent to the problem and how much is lack of innovation. And we'll get to find out because if the processes stall from the economic side all the research turns to that whereas today it's an arms race for the next node and the increasing cost is taken as a…

Tbh, it's less of a research problem per se, and more of an implementation one, in many ways.

The current pace of development is such that that they are plugging stuff in and hoping it works in LOTS of different ways.

A small army of electrical engineers work night shifts at TSMC as well, something Intel could never pull off given American labour market demand for EEs.

I suspect you're right, when costs have to come down timelines will blow out to something more reasonable.

Post reply on HN